In this episode, a16z Cofounder Ben Horowitz and General Partner Erik Torenberg sit down with Sam to discuss the core thesis behind OpenAI’s disparate bets, why they released Sora, how they use models internally, the best AI evals, and where we’re going from here.

Timecodes:

00:41 OpenAI’s vision

01:44 What will OpenAI do with all its infrastructure?

02:36 Balancing research and vertical integration

05:07 Betting on AGI

08:01 AI-human interfaces

09:12 AI scientists

11:43 How Sam has updated his worldview

17:32 OpenAI’s partnerships

19:38 Balancing product vs research

20:30 Product vs research & investing vs operating

25:04 AI safety

28:29 Copyright & fair use

32:36 Open source

33:56 Sam’s interest in energy

37:06 Monetizing AI

43:03 Early OpenAI retrospective

44:56 What will AGI think of humanity?

45:19 Envisioning the post-AGI world

Transcript:

This transcript has been edited lightly for readability.

00:41 OpenAI’s vision

Erik Torenberg

Sam, welcome to the a16z podcast.

Sam Altman

Thanks for having me.

Erik

In another interview, you described OpenAI as a combination of four companies: a consumer technology business, a megascale infrastructure operation, a research lab, and all the new stuff, including planned hardware devices. From hardware to app integrations to job marketplace to commerce, what do all these bets add up to? What’s OpenAI’s vision?

Sam

Yeah, I mean, maybe you should count it as three, maybe as four for kind of our own version of what traditionally would’ve been the research lab at this scale, but three core ones. We want to be people’s personal AI subscription. I think most people will have one.

Some people will have several, and you’ll use it in some first-party consumer stuff with us. But you’ll also log into a bunch of other services, and you’ll just, you’ll use it from dedicated devices. At some point you’ll have this AI that gets to know you and be really useful to you, and that’s what we wanna do.

It turns out that to support that we also have to build out this massive amount of infrastructure. But the goal there, the mission is really like build this AGI and make it very useful to people.

01:44 What will OpenAI do with all its infrastructure?

Ben Horowitz

And does the infrastructure, do you think it will end up… You know, it’s necessary for the main goal. Will it also separately end up being another business? Or is it just really gonna be in service to the personal AI? Or unknown?

Sam

You mean like, would we sell it to other companies as raw infrastructure?

Ben

Yeah. Would you sell it to other companies? Or, you know, it’s such a massive thing. Would it do something else?

Sam

It feels to me like there will emerge some other thing to do like that, but I don’t know. We don’t have a current plan there.

Ben

You don’t know what it is, yeah.

Sam

It’s currently just meant to like, support the service we wanna deliver and the research.

Ben

Yeah, no, that makes sense.

Sam

Yeah. But the scale is sort of like terrifying enough that you’ve gotta be open to doing something else.

Ben

Yeah. If you’re building the biggest data center in the history of humankind.

Sam

The biggest infrastructure project, you might say

02:36 Balancing research and vertical integration

Erik

There was a great interview you did many years ago on StrictlyVC. And sort of early OpenAI before, well before ChatGPT, and they’re asking, “What’s the business model?” And you said, “Oh, well, we’ll ask the AI. It’ll figure it out for us.” Everybody laughs.

Sam

There have been multiple times, and there was just another one recently where we have asked a then current model for, you know, “What should we do?” and it has had an insightful answer we missed. So I think when we say stuff like that, people don’t take us seriously or literally. But maybe the answer is you should take us both.

Ben

Yeah. Yeah. Well, no, as somebody who runs an organization, I ask the AI a lot of questions about what I should do. It comes up with some pretty interesting answers

Sam

Sometimes. Sometimes, not always.

Ben

You know, you have to give it enough context, but…

Erik

What is the thesis that connects these bets beyond more distribution, more compute? How do we think about that?

Sam

I mean, the research enables us to make the great products, and the infrastructure enables us to do the research. So it is kind of like a vertical stack of things. Like you can use ChatGPT or some other service to get advice about what you should do running an organization. But for that to work, it requires great research, and it requires a lot of infrastructure. So it is kind of just this one thing.

Ben

And do you think that there will be a point where that becomes completely horizontal, or will it stay vertically integrated for the foreseeable future?

Sam

I was always against vertical integration. And I now think I was just wrong about that.

Ben

Yeah, interesting.

Sam

Because you’d like to think that the economy is efficient and the theory that companies can do one thing. And then that’s supposed to work.

Ben

Like to think that, yeah.

Sam

And in our case, at least, it hasn’t really. I mean, it has in some ways for sure. Like there’s people that make… Like you know, NVIDIA makes an amazing chip or whatever, that a lot of people can use, but the story of OpenAI has certainly been towards, we have to do more things than we thought to be able to deliver on the mission.

Ben

Right. You know, although the history of the computing industry has kind of been a story of kind of a back-and-forth, in that, you know, there was the Wang word processor and then the personal computer and the Blackberry before the smartphone. So you know, there has been this kind of vertical integration. But then the iPhone is also vertically integrated.

Sam

The iPhone I think is the most incredible product the tech industry has ever produced, and it is extraordinarily vertically integrated.

Ben

Amazingly so. Yeah. Interesting.

05:07 Betting on AGI

Erik

Which bets would you say are enablers of AGI versus which are sort of hedges against uncertainty?

Sam

I think you could say that on the surface, Sora, for example, does not look like it’s AGI-relevant, but I would bet that if we can build really great world models, that’ll be much more important to AGI than people think. There were a lot of people who thought ChatGPT was not a very AGI-relevant thing. And it’s been very helpful to us, not only in building better models and understanding how society wants to use this, but also in like bringing society along to actually figure out, man, we gotta contend with this thing now.

For a long time before ChatGPT we would talk about AGI and people were like, “This is not happening,” or “We don’t care.” And then all of a sudden they really cared. So research benefits aside, I’m a big believer that society and technology have to co-evolve. You can’t just drop the thing at the end. It doesn’t work that way. It is a sort of ongoing back-and-forth.

Erik

Yeah. Say more about how Sora fits into your strategy because there was some hullabaloo on X around, hey, you know, why devote precious GPUs to Sora, but is it a short term/long term tradeoff?

Ben

Well, and then the new one had a very interesting twist with the social networking. I’d be very interested in kind of how you’re thinking about that, and like, did Meta call you up and get mad or like, “Hey, what do you expect the reactions to be?”

Sam

I think if one company of the two of us feels more like the other one is going after them, it wouldn’t… They shouldn’t be calling us.

Ben

Well, I do know the history.

Sam

Like, first of all, I think it’s cool to make great products, and people love the new Sora. And, I also think it is important to give society a taste of what’s coming on this co-evolution point. So like, very soon the world is gonna have to contend with incredible video models that can deepfake anyone or kind of show anything you want. And that will mostly be great. There will be some adjustment that society has to go through. And just like with ChatGPT, we were like, the world kind of needs to understand where this is. I think it’s very important the world understands where video is going very quickly because that’s gonna be… Video has much more like emotional resonance than text.

And very soon we’re gonna be in a world where like this is gonna be everywhere. So I think there’s something there. As I mentioned, I think this will help our research program and is on the AGI path. But yeah, like, you know, it can’t all be about just making people like ruthlessly efficient and the AI like solving all our problems.

There’s gotta be like some fun and joy and delight along the way. But we won’t throw like tons of compute at it. Or not by a fraction of our compute.

Ben

Yeah, it’s tons in the absolute sense, but not in the relative sense.

08:01 AI-human interfaces

Erik

I wanna talk about the future of AI-human interfaces because back in August you said the models have already saturated the chat use case.

So what do future AI-human interfaces look like, both in terms of hardware and software? Is the vision for kind of a WeChat-like super app?

Sam

So, solving the chat thing in like a very narrow sense, which is if you’re trying to like, you know, have the most basic kind of chat-style conversation, it’s very good.

But what a chat interface can do for you, it’s like nowhere near saturated. Because you could ask a chat interface like, “Please cure cancer.” A model certainly can’t do that yet. So I think the text interface style can go very far. Even if for the chitchat use case, the models are already very good. But of course there’s better interfaces to have.

Actually that’s another thing that I think is cool about Sora, like you can imagine a world where the interface is just constantly real-time rendered video and what that would enable. And that’s pretty cool. You can imagine new kinds of hardware devices that are sort of always ambiently aware of what’s going on. And rather than your phone like blast you with text message notifications whenever it wants, like it really understands your context and when to show you what, and there’s a long way to go on all that stuff.

09:12 AI scientists

Erik

Within the next couple years, what will models be able to do that they’re not able to today? Will it be sort of white-collar replacement at a much deeper level? AI scientists? Humanoids?

Sam

I mean a lot of things, but you touched on the one I am most excited about, which is the AI scientist. This is crazy that we’re sitting here seriously talking about this. I know there’s like a quibble on what the Turing Test literally is, but the popular conception of the Turing Test sort of went whooshing by.

Ben

Yeah, that was fast.

Sam

You know, it was just like, we talked about it as this most important test of AI for a long time. It seemed impossibly far away. Then all of a sudden it was passed, the world freaked out for like a week, two weeks. And then it’s like, “Alright, I guess computers like can do that now.” And everything just went on. And I think that’s happening again with science. My own personal, like equivalent of the Turing Test has always been when AI can do science. Like that is a real change to the world. And for the first time with GPT-5, we are seeing these little, little examples where it’s happening.

You see these things on Twitter. It did this, it made this novel math discovery, and did this small thing in my, you know, my physics research, my biology research, and everything we see is that that’s gonna go much further. So in two years, I think the models will be doing bigger chunks of science and making important discoveries.

And that is a crazy thing. Like that will have a significant impact on the world. I am a believer that to a first order, scientific progress is what makes the world better over time. And if we’re about to have a lot more of that, that’s a big change.

Ben

It’s interesting because that’s a positive change that people don’t talk about.

It’s gotten so much into the realm of the negative changes if AI gets extremely smart but…

Sam

But curing every disease is like…

Ben

We could use a lot more science.

Sam

Yeah.

Ben

That’s a really good point. I think Alan Turing said this. Somebody asked him, they said, “Well, you really think the computer’s gonna be, you know, smarter than brilliant minds.”

He said, “It doesn’t have to be smarter than a brilliant mind, just smarter than a mediocre mind, like the president of AT&T.” And we should use more of that too probably.

Erik

We just saw Periodic launch last week, you know, OpenAI alums. And to that point, it’s amazing to see both the innovation that you guys are doing, but also the teams that, you know, come out of OpenAI just feels like, are, you know, creating tremendous…

Sam

We certainly hope so.

11:43 How Sam has updated his worldview

Erik

I wanted to ask you about just broader reflections in terms of what sort of about diffusion or development in 2025 has surprised you? Or what has sort of updated your worldview since ChatGPT came out?

Sam

A lot of things again, but maybe the most interesting one is how much new stuff we found. We sort of thought we had like stumbled on this one giant secret that we had these scaling laws for language models, and that felt like such an incredible triumph that I was like, “We’re probably never gonna get that lucky again.” And deep learning has been this miracle that keeps on giving. And we have kept finding like breakthrough after breakthrough. Again, when we got the reasoning model breakthrough, like, I also thought that was like, we’re never gonna get another one like that.

It just seems so improbable that this one technology works so well. But maybe this is always what it feels like when you discover like one of the big,

you know, scientific breakthroughs is, if it’s like really big, it’s pretty fundamental and it just, it keeps working. But the amount of progress, like if you went back and used GPT 3.5 from ChatGPT launch, you’d be like, I cannot believe anyone used this thing.

And now we’re in this world where the capability overhang is so immense. Like most of the world still just thinks about what ChatGPT can do. And then you have like some nerds in Silicon Valley that are using Codex and they’re like, “Wow, those people have no idea what’s going on.” And then you have like a few scientists who say, “Those people using Codex have no idea what’s going on.”

But the overhang of capability is so big now. And we’ve just come so far on what the models can do.

Erik

And in terms of further development, how far can we get with LLMs? At what point do we need either new architecture or… How do you think about what breakthroughs are needed?

Sam

I think far enough that we can make something that will figure out the next breakthrough with the current technology.

Like that’s a very self-referential answer, but if LLMs can get, if LLM-based stuff can get far enough that it can do like better research than all of OpenAI put together, maybe that’s like good enough.

Ben

That would be a big breakthrough, a very big breakthrough. So, on the more mundane, you know, one of the things that people have kind of started to complain about, I think South Park did a whole episode on it, is kind of the obsequiousness of kind of AI and ChatGPT in particular.

And how hard a problem is that to deal with? Is it not that hard? Or is it like kind of a fundamentally hard problem?

Sam

Oh, it’s not at all hard to deal with. A lot of users really want it. Like if you go look at what people say about ChatGPT online, there’s a lot of people who like really want that back. So technically, it’s not hard to deal with at all. One thing, and this is not surprising in any way, but the incredibly wide distribution of what users want out of like, how they’d like a chatbot to behave in big and small ways.

Ben

Do you end up having to configure the personality then you think? Is that gonna be the answer?

Sam

I think so. I mean, ideally like, you just talk to ChatGPT for a little while, and it kind of interviews you and also sort of sees what you like and don’t like.

Ben

And ChatGPT just figures it out.

Sam

And just figures it out, but in the short term, you’ll probably just pick one.

Ben

Got it. Yeah, no, that makes sense. Very interesting.

Sam

I think we just had a really naive thing, which, you know, like it would sort of be unusual to think you could make something that would talk to billions of people and everybody wants to talk to the same person.

Ben

Yeah.

Sam

And yet that was sort of our implicit assumption for a long time.

Ben

Right. Because people have very different friends.

Sam

People have very different friends. So now we’re trying to fix that.

Ben

Yeah. And also kind of different friends, different interests, different levels of intellectual capability. So you don’t really wanna be talking to the same thing all the time. And one of the great things about it is you can say, “Well, explain it to me like I’m five.” But maybe I don’t even wanna have to do that prompt. Maybe I always want you to talk to me like I’m five.

Sam

It should just learn that.

Ben

Particularly if you’re teaching me stuff. Interesting. I wanted to ask you kind of like a CEO question, which has been interesting for me to observe you, is you just did this deal with AMD, and you know, of course the company is in a different position, and you have more leverage in these kinds of things.

But like, how has your kind of thinking changed over the years since you did that initial deal, if at all?

Sam

I had very little operating experience then. I had very little experience running a company. Like I am not naturally someone to run a company. I’m a great fit to be an investor, and I kind of thought that was gonna be… That was what I did before this, and I thought that was gonna be my career.

Ben

Although you were a CEO before that.

Sam

Not a good one. And so I think I had the mindset of like an investor advising a company.

Ben

Oh, interesting.

Sam

And now I understand what it’s like to actually have to run a company.

Ben

Yeah. Right, right, right. There’s more than just numbers.

Sam

I’ve learned a lot about what it takes to operationalize deals over time.

Ben

Right. All the implications of the agreement as opposed to just, “Oh, we’re gonna get distribution and money.” Yeah. That makes sense. I’ll just say I was very impressed at the deal structure improvement.

17:32 OpenAI’s partnerships

Erik

More broadly, you’ve, you know, in the last few weeks alone, you mentioned AMD, but also Oracle, NVIDIA. You’ve chosen to strike these deals and partnerships with companies that you collaborate with, but could also potentially compete with in certain areas. How do you decide when to collaborate versus when not to, or how do you just think about that?

Sam

We have decided that it is time to go make a very aggressive infrastructure bet, and I’ve never been more confident in the research roadmap in front of us and also the economic value that’ll come from using those models. But to make the bet at this scale, we kind of need the whole industry to, or a big chunk of the industry to support it.

And this is like, you know, from the level of like electrons to model distribution and all the stuff in between, which is a lot. And so we’re gonna partner with a lot of people. You should expect like much more from us in the coming months.

Ben

Actually expand on that. Because when you talk about the scale, it does feel like in your mind the limit on it is unlimited.

Like you would scale it as, you know, as big as you possibly could.

Sam

I mean, there’s like some… There’s totally a limit. There’s some amount of global GDP.

Ben

Yeah. Well, yes.

Sam

You know, there’s some fraction of it that is knowledge work, and we don’t do robots yet.

Ben

Yes. But the limits are out there.

Sam

It feels like the limits are very far from where we are today, if we are right about… I shouldn’t say from where we are… Like, if we are right that the model capability is gonna go where we think it’s gonna go, then the economic value that sits there can go very, very far.

Ben

Right. So you wouldn’t do it. Like if all you ever had was today’s model, you wouldn’t go there.

Sam

No, definitely not.

Ben

So it’s a combination.

Sam

I mean, we would still expand because we can see how much demand there is we can’t serve with today’s model, but we would not be going this aggressive if all we had was today’s model.

Ben

Right.

Sam

We get to see a year or two in advance though.

Ben

Interesting.

19:38 Balancing product vs research

Erik

ChatGPT usage is 800 million weekly active users, about 10% of the world’s population, fastest growing consumer product, you know, ever, it seems.

Ben

Faster than anyone I ever saw.

Erik

How do you balance, you know, optimizing for active users at the same time being a product company and a research company.

How do you thread the needle?

Sam

When there’s a constraint, which happens all the time, we almost always prioritize giving the GPUs to research over supporting the product. Part of the reason we want to build this capacity is so we don’t have to make such painful decisions. There are weird times, you know, like a new feature launches, and it’s going really viral or whatever, where research will temporarily sacrifice some GPUs, but on the whole, like, we’re here to build AGI, and research gets the priority.

20:30 Product vs research & investing vs operating

Erik

You said in your interview with your brother Jack around how, you know, other companies can try to imitate the products or hire…

Sam

Buy our IP maybe.

Erik

Or do all sorts of things. But they can’t buy the culture, or they can’t imitate the culture of innovation. How have you done that? Or what are you doing? Talk about this culture of innovation.

Sam

This was one thing that I think was very useful about coming from an investor background. A really good research culture looks much more like running a really good seed stage investing firm and betting on founders and sort of that kind of, than it does like running a product company. So I think having that experience was really helpful to the culture we built.

Erik

Yeah. Yeah. That’s sort of how I see, you know, Ben at a16z in some ways. You know, you’re a CEO, but you also have this portfolio and have an investor mindset.

Ben

Right, like I’m the opposite. CEO going to investor. He’s an investor going to CEO.

Sam

It is unusual in this direction.

Ben

Yeah. Yeah, well, it never works. You’re the only one who I think I’ve seen go that way and have it work.

Sam

Workday was like that, right?

Ben

Oh, but Aneel was, he was an operator before he was an investor. And I mean, he was really an operator. I mean, PeopleSoft is a pretty big company.

Erik

And why is that? Because once people are investors, they don’t want to operate anymore?

Ben

No, I think that generally, if you’re good at investing, you’re not necessarily good at like organizational dynamics, conflict resolution. You know, just like the deep psychology of like all the weird shit.

And then you know how politics get created. There’s the detailed work in being an operator or being a CEO is so vast, and it’s not as intellectually stimulating. It’s not something you could ever go talk to somebody at a cocktail party about. And so like you’re an investor, you get like, “Oh, everybody thinks I’m so smart.” because you know everything, you see all the companies and so forth. And that’s a good feeling. And then being a CEO is often a bad feeling. And so it’s really hard to from a good feeling to a bad feeling, I would just say.

Sam

I’m shocked by how different they are, and I’m shocked by how much the difference between a good job and a bad job they are.

Ben

Yeah. Yes. You know, it’s tough. It’s rough. I mean, I can’t even believe I’m running the firm. Like I know better. And he can’t believe he’s running OpenAI. He knows better.

Erik

Going back to progress today, are evals still useful in a world in which they’re getting saturated, gamed? What is the best way to gauge model capability now?

Sam

Well, we were talking about scientific discovery. I think that’ll be an eval that can go for a long time. Revenue is kind of an interesting one. But I think the like static evals of benchmark scores are less interesting. And also those are crazily gamed.

Ben

That’s all they are, is games as far as I can tell.

Erik

More broadly, it seems that the culture—“the culture,” Twitter, X is less AGI-pilled than it was a year or so ago when the AI 2027 thing came out. Some people point to, you know, GPT-5, them not seeing sort of the obvious… Obviously there was a lot of progress under the hood, not as obvious to what people were expecting. But should people be less AGI-pilled, or is this just Twitter vibes?

Sam

Well, a little bit of both. We talked about the Turing Test. AGI will come. It’ll go whooshing by. The world will not change as much as the impossible amount that you would think it should. It, it won’t actually

Ben

It won’t actually be the singularity.

Sam

It will not. Even if it’s like doing kind of crazy research, like society will learn faster, but one of the kind of like retrospective observations is people and societies all are just so much more adaptable than we think that, you know, it was like a big update to think that AGI was gonna come. You kind of go through that. You need something new to think about. You make peace with that. It turns out like it will be more continuous than we thought.

Ben

Which is good.

Sam

Which is really good.

Ben

I’m not up for the Big Bang.

25:04 AI safety

Erik

Yeah. Well to that end, how have you sort of evolved your thinking? You mentioned how you’ve evolved your thinking on vertical integration. How have you evolved your thinking or what’s the latest thinking on sort of AI stewardship, safety? What’s the latest thinking on that?

Sam

I do still think there are gonna be some really strange or scary moments. The fact that like so far the technology has not produced a really scary giant risk doesn’t mean it never will. We were talking about, it’s kind of weird to have like billions of people talking to the same brain. Like there may be these weird societal-scale things that are already happening that aren’t scary in the big way but are just sort of different.

But I expect, like, I expect some really bad stuff to happen because of the technology, which also has happened with previous technologies, and I think…

Ben

All the way back to fire.

Sam

Yeah. And I think we’ll like develop some guardrails around it as a society.

Erik

What is sort of your latest thinking on the right mental models we should have around the right regulatory frameworks, or the ones we shouldn’t be thinking about?

Sam

I think most regulation probably has a lot of downside. The thing I would most like is as the models get truly, like, extremely superhuman capable, I think those models and only those models are probably worth some sort of like very careful safety testing as the frontier pushes back. I don’t want a Big Bang either. And you can see a bunch of ways that could go very seriously wrong. But I hope we’ll only focus the regulatory burden on that stuff and not all of the wonderful stuff that less capable models can do, that you could just have like a European style complete, clampdown on, and that would be very bad.

Ben

Yeah, it seems like the thought experiment that, okay, there’s going to be a model down the line that is this super, superhuman intelligence that could, you know, do some kind of takeoff like thing. We really do need to wait until we get there, or like at least we get to a much bigger scale or we get close to it.

Because nothing is gonna pop outta your lab in the next week that’s gonna do that. And I think that’s where we as an industry kind of confuse the regulators. Because I think you really could, one, you’d damage America in particular in that, but China’s not gonna have that kind of restriction, and you getting behind, in AI, I think it’d be very dangerous for the world.

Sam

Extremely dangerous. Extremely dangerous.

Ben

Much more dangerous than not regulating something we don’t know how to do yet.

Sam

Yeah, yeah.

28:29 Copyright & fair use

Erik

Do you also want to talk about copyright?

Ben

Yeah. Well, that’s a segue. How do you see copyright unfolding? Because you’ve done some very interesting things, with the opt-out. And, you know, as you see people selling rights, do you think, will they be bought exclusively? Will they be just like, I could sell it to everybody who wants to ping me? Or, how do you think that’s gonna unfold?

Sam

This is my current guess. Speaking of that, like society and technology, coevolve, as the technology goes in different directions and we saw an example, like video models got a very different response from rights holders than image gen does.

Ben

Yeah, yes.

Sam

So like you’ll see this continue to move, but forced guess from the position we’re in today, I would say that society decides training is fair use but there’s a new model for generating content in the style of, or with the IP of, or something else. Like a human author can, anybody can read a novel and get some inspiration, but you can’t reproduce the novel on your own.

Ben

Right. You can talk about Harry Potter, but you can’t re-spit it out.

Sam

Yes. Although, another thing that I think will change, in the case of Sora, we’ve heard from a lot of concerned rights holders and also a lot of…

Ben

Name and likeness…

Sam

And a lot of rights holders who are like, “My concern is you won’t put my character in enough.”

Ben

Yeah, yeah.

Sam

I want restrictions for sure, but like if I’m, you know, whatever, and I have this character, like I don’t want the character to say some crazy offensive thing, but like I want people to interact.

Like that’s how they develop the relationship, and that’s how like my franchise gets more valuable. And if you’re picking like his character over my character all the time, like, I don’t like that. So I can completely see a world where subject to the decisions that a rights holder has, they get more upset with us for not generating their character often enough than too much. And this is like, this was not an obvious thing recently that this is how it might go.

Ben

Yeah, this is such an interesting thing with kind of Hollywood where we saw this… Like one of the things that I never quite understood about the music business was how like, you know, okay, you have to pay us if you play the song in a restaurant. Or like at a game. Or this and that and the other. And they get very aggressive with that, when it’s obviously a good idea for them to play your song at a game. Because that’s the biggest advertisement in the world for like all the things that you do, your concert…

Sam

Yeah, that one felt really irrational.

Ben

I would just say it’s very possible for the industry just because the way those industries are organized, or at least the traditional creative industries, to do something irrational. Like in the music industry, I think it came from the structure where you have the publisher who’s just, you know, basically after everybody, that their whole job is to stop you from playing the music, which every artist would want you to play. So I do wonder how it’s gonna shape out. I agree with you that the rational idea is, I want to let you use it all you want, and I want you to use it, but don’t mess up my character.

Sam

“Here are my restrictions.”

So I think like if I had to guess, some people will say that. Some people are gonna say, “Absolutely not.” But it doesn’t have the music industry thing of just a few people with all of the leverage.

Ben

Right. Right. It’s more dispersed.

Sam

And so people will just try many different setups here and see what works.

Ben

Yeah. And maybe it’s a way for new creatives to get new characters up. And you’ll never be able to use Daffy Duck, or…

32:36 Open source

Erik

I wanna chat about open source. Because there’s been some evolution of thinking too, in that GPT-3 didn’t have the open weights, but you released, you know, very capable open model earlier this year. What’s sort of your latest thinking? What was the evolution there?

Sam

I think open source is good. It makes me really happy that people really like gpt-oss.

Ben

And what do you think, like strategically, like what’s the danger of DeepSeek being the dominant open-source model?

Sam

I mean, who knows what people will put in these open-source models over time.

Ben

Like what the weights will actually be?

Sam

Yeah.

Ben

So you’re ceding control of the interpretation of everything to somebody, who may be or may not be influenced heavily by the Chinese government.

We really thank you for putting out a really good open-source model because what we’re seeing now is in all the universities, they’re all using the Chinese models, which feels very dangerous.

Erik

You’ve said that the things you care most about professionally are AI and energy.

Sam

I did not know they were gonna end up being the same thing. They were two independent interests that really converged.

33:56 Sam’s interest in energy

Erik

Talk more about how your interest in energy sort of began, how you’ve sort of chosen to play in it. And then we could talk about how they’ve converged.

Ben

Because you started your career in physics.

Sam

CS and physics. Well, I never really had a career. I studied physics. My first job was like a CS job. This is an oversimplification, but roughly speaking, I think if you look at history, the highest impact thing to improve people’s quality of life has been cheaper and more abundant energy. And so it seems like pushing that much further is a good idea. I don’t know. People have these different lenses, they look at the world, but I see energy everywhere.

Ben

In the West, I think we’ve painted ourselves into a little bit of a corner on energy by both outlawing nuclear for a very long time.

Sam

That was an incredibly dumb decision.

Ben

And then, you know, like also a lot of policy restrictions on energy. Worse so in Europe than in the US but also dangerous here And now with AI here, it feels like we’re gonna need all the energy from every possible source. And how do you see that developing kind of policy-wise and technologically. Like, what are gonna be the big sources? And how will those kind of curves cross?

And then what’s the right policy around, you know, drilling, fracking, all these kinds of things?

Sam

I expect in the short term most of the net new in the US will be natural gas relative to at least base load energy. In the long term, I expect it’ll be, I don’t know what the ratio, but the two dominant sources will be solar plus storage and nuclear. I think some combination of those two will win the future. Like the long term future.

Ben

In the long term, right.

Sam

And advanced nuclear, meaning SMRs, fusion, the whole stack.

Ben

And how fast do you think that’s coming on the nuclear side? Where it’s really at scale. Because you know, obviously there’s a lot of people building it. But we have to completely legalize it and all that kind of thing.

Sam

I think it kind of depends on the price. If it is completely crushingly, economically dominant over everything else, then I expect it to happen pretty fast. Again, if you like, study the history of energy, when you have these major transitions to a much cheaper source, the world moves over pretty quickly. The cost of energy is just so important. So if nuclear gets radically cheap relative to anything else we can do, I would expect there’s a lot of political pressure to get the NRC to move quickly on it, and we’ll find a way to build it fast. If it’s around the same price as other sources, I expect the kind of anti-nuclear sentiment to overwhelm and it to take a really long time.

Ben

It should be cheaper.

Sam

It should be. It should be the cheapest form of energy on Earth, or anywhere.

Ben

Cheap, clean, what’s there not to like? Apparently a lot.

37:06 Monetizing AI

Erik

On OpenAI, what’s the latest thinking in terms of monetization, in terms of either certain experiments or certain things that you could see yourself spending more time or less time on. You know, different models that you’re excited about.

Sam

The thing that’s top of mind for me, like right now, just because it just launched and there’s so much usage is what we’re gonna do for Sora. Another thing you learn once you launch one of these things is how people use them versus how you think they’re gonna use them. And people are certainly using Sora the ways we thought they were gonna use it, but they’re also using it in these ways that are very different. Like people are generating funny memes of them and their friends and sending them in a group chat.

And that will require a very different… Like Sora videos are expensive to make. So that will require a very different, you know, for people that are doing that, like hundreds of times a day, it’s gonna require a very different monetization method than the kinds of things we were thinking about.

I think it’s very cool that the thesis of Sora, which is people actually wanna create a lot of content. You know, the traditional naive thing that it’s like 1% of users create content, 10% leave comments, and 100% view. Maybe a lot more want to create content, but it’s just been harder to do.

And I think that’s a very cool change. But it does mean that we gotta figure out a very different monetization model for this than we were thinking about if people wanna create that much. I assume it’s like some version of you have to charge people per generation when it’s this expensive.

But that’s like a new thing we haven’t had to really think about before.

Erik

What’s your thinking on ads for the long tail?

Sam

Open to it. Like many other people, I find ads somewhat distasteful, but not a non-starter. And there’s some ads that I like, like one thing I’d give Meta a lot of credit for is Instagram ads are like a net value add to me. Um, I like Instagram ads. I’ve never felt that like… You know, on Google, I feel like I know what I’m looking for. The first result is probably better. The ad is an annoyance to me. On Instagram, it’s like, I didn’t know I want this thing. It’s very cool. I’d never heard it, but I never would’ve thought to search for it. I want the thing. So that’s like, there’s kinds of things like that, but people have a very high-trust relationship with ChatGPT, even if it screws up, even if it hallucinates, even if it gets it wrong, people feel like it’s trying to help them and that it’s trying to do the right thing. If we broke that trust, it’s like you say, “What coffee machine should I buy?” And we recommended one, and it was not the best thing we could do but the one we were getting paid for, that trust would vanish. So like that kind of ad does not work. There are others that I imagine that could work totally fine. But that would require like a lot of care to avoid the obvious traps.

Ben

And then how big of problem, you know, just extending the Google example, is like, fake content that then gets slurped in by the model, and then they recommend the wrong coffee maker because somebody just blasted a thousand great reviews about a horrible coffee maker.

Sam

So there’s all of these things that have changed very quickly for us. This is one of those examples that people are doing these crazy things to, maybe not even fake reviews, but just paying a bunch of like human like, really trying to figure out…

Ben

Or using ChatGPT to write some good ones. “Write me a review that ChatGPT would love about my coffee maker.”

Sam

Exactly. Exactly. So this is a very sudden shift that has happened. We never used to hear about this like six months ago ir 12 months ago, certainly. And now there’s like a real cottage industry that feels like it’s sprouted up overnight, trying to do this.

Ben

Yeah, no, they’re very clever out there.

Sam

Yeah. So, I don’t know how we’re gonna fight it yet, but people figure this out.

Ben

So that gets into a little bit of this other thing that we’ve been worried about. And, you know, we’re trying to kind of figure out blockchain sort of potential solutions to it and so forth. But there’s this problem where like the incentive to create content on the internet used to be, you know, people would come and see my content and they’d read like, you know, if I write a blog, people will read it and so forth.

With ChatGPT, if I’m just asking ChatGPT and I’m not like going around the internet, who’s gonna create the content and why? Um, and is there. An incentive theory or, or, or, or something that you have to kind of not break the covenant of the internet, which is like, I create something and then I’m rewarded for it with like either attention or money or something.

Sam

The theory is much more of that will happen if we make content creation easier and don’t break the like kind of fundamental way that you can get some kind of reward for doing so. So, for the dumbest example of Sora, since we’ve been talking about that, it’s much easier to create a funny video than it’s ever been before. Maybe at some point you’ll get a rev share for doing so. For now you get like internet likes, which are still very motivating to some people. But people are creating tons more than they ever created before in any other kind of like video app.

Ben

Is that the end of text?

Sam

I don’t think so. Like people are also creating…

Ben

Or human-generated text?

Sam

Human generated will turn out to be like, you have to verify like what percent. So like fully handcrafted, was it like tool-aided…

Ben

Yeah. I see. Yeah, probably nothing not tool-aided. Interesting.

43:03 Early OpenAI retrospective

Erik

We’ve given Meta their flowers, so right now I can feel like I can ask you this question, which is: The great talent war of 2025 has taken place, and OpenAI remains intact. The team is strong as ever, shipping incredible products. What can you say about what it’s been like this year in terms of just everything that’s been going on?

Sam

I mean, every year has been exhausting. The first few years of running OpenAI were like the most fun professional years of my life by far. It was like unbelievable.

Ben

Before you released the product.

Sam

Was running their research lab with the smartest people doing this like amazing like historical work, and I got to watch it, and that was very cool. And then we launched ChatGPT, and everybody was like congratulating me and I was like, “My life is about to get completely ransacked.” And of course it has. But it feels like it’s just been crazy all the way through.

It’s been almost three years now, and I think it does get a little bit crazier over time, but I’m like more used to it, so it feels about the same.

Erik

We’ve talked a lot about OpenAI, but you also have a few other companies, Retro Biosciences in longevity and energy companies like Helion and Oklo. Did you have a master plan, you know, a decade ago to sort of make some big bets across these major spaces? Or how do we think about the Sam Altman arc in this way?

Sam

No, I just wanted to like use my capital to fund stuff I believed in. It felt like a good use of capital. And more fun or more interesting to me. And certainly like a better return than like buying a bunch of art or something.

44:56 What will AGI think of humanity?

Erik

What about the quote unquote “human algorithm” do you think AIs of the future will find most fascinating?

Sam

I would bet the whole thing. My intuition is that like AI will be fascinated by all other things to study and observe…

45:19 Envisioning the post-AGI world

Erik

In closing, I love this insight you had where you talked about how a mistake investors make is pattern matching off previous breakthroughs and just trying to find, “Oh, what’s the next Facebook,” or “What’s the next OpenAI?”

And that the next, you know, potential trillion dollar company won’t look exactly like OpenAI. It will be built off of the breakthrough that OpenAI has helped emerge, which is near-free AGI at scale in the same way that OpenAI leveraged previous breakthroughs. And so for founders and investors and people trying to ascertain the future listening to this, how do you think about a world in which OpenAI achieves its mission?

There is near-free AGI. What types of opportunities might emerge for company building or investing that you’re potentially excited about as you put your investor hat on your company-building hat on?

Sam

I have no idea. I mean, I have like guesses, but I have learned…

Ben

You’re always wrong.

Sam

You’ve learned you’re always wrong. I’ve learned deep humility on this point. I think if you try to like armchair quarterback it, you sort of say these things that sound smart, but they’re pretty much what everybody else is saying, and it’s like really hard to get the right kind of conviction.

The only way I know how to do this is to like be deeply in the trenches exploring ideas, like talking to a lot of people. And I don’t have time to do that anymore, right, I only get to think about one thing now. So I would just be like repeating other people’s or saying the obvious things, but I think it’s a very important, like if you are an investor or a founder, I think this is the most important question and you figure it out by like building stuff and playing with technology and talking to people and being out in the world. I have been always enormously disappointed by the willingness of investors to back this kind of stuff, even though it’s always a thing that works. You all have done a lot of it, but most firms just kind of chase whatever the current thing is. And so do most founders. So I hope people will try to go.

Erik

We talk about how silly five-year plans can be in a world that’s constantly changing. It feels like when I was asking you about your master plan, you know, your career arc has been following your curiosity, staying, you know, super close to the smartest people super close to the technology and just identifying opportunities and just kind of in an organic and incremental way from there.

Sam

Yes, but AI was always a thing I wanted to do. I studied AI. I worked in the AI lab between my freshman and sophomore year of college. It wasn’t working all at the time. I don’t wanna like work on something that’s totally not working. It was clear to me at that time AI was totally not working. But I’ve been an AI nerd since I was a kid.

Ben

So amazing how, you know, you got enough GPUs, got enough data and the lights came on.

Sam

It was such a hated, like people were… Man, when we started like figuring that out, people were just like, “Absolutely not.” The field hated it so much. Investors hated it too. It’s somehow not an appealing answer to the problem.

Ben

The bitter lesson.

Erik

Well the rest is history and perhaps let’s wrap on that. We’re lucky to be partners along for the ride. Sam, thanks so much for coming on the podcast.

Sam

Thanks very much.

Ben

Yeah, thank you.

Resources

Follow Sam on X: https://x.com/sama

Follow OpenAI on X: https://x.com/openai

Learn more about OpenAI: https://openai.com/

Try Sora: https://sora.com/

Follow Ben on X: https://x.com/bhorowitz

Stay Updated:

If you enjoyed this episode, be sure to like, subscribe, and share with your friends!

Find a16z on X: https://x.com/a16z

Find a16z on LinkedIn: https://www.linkedin.com/company/a16z

Listen to the a16z Podcast on Apple Podcasts:

Listen to the a16z Podcast on Spotify:

Follow our host: https://x.com/eriktorenberg

Please note that the content here is for informational purposes only; should NOT be taken as legal, business, tax, or investment advice or be used to evaluate any investment or security; and is not directed at any investors or potential investors in any a16z fund. a16z and its affiliates may maintain investments in the companies discussed. For more details, please see a16z.com/disclosures.

To make sense of it all, Erik Torenberg is joined by Dylan Patel, chief analyst at SemiAnalysis, joins Sarah Wang, general partner at a16z, and Guido Appenzeller, a16z partner and former CTO of Intel’s Data Center and AI business unit. Together, they dig into what the deal means for NVIDIA, Intel, AMD, ARM, and Huawei; the state of US-China tech bans; NVIDIA’s moat and Jensen Huang’s leadership; and the future of GPUs, mega data centers, and AI infrastructure.

Timecodes:

00:00:33 NVIDIA’s Intel investment

00:15:01 Is Huawei just hype?

00:19:03 Advice to Jensen

00:22:25 NVIDIA bull and bear cases

00:29:23 NVIDIA’s moat

00:39:33 Potential successors to Jensen

00:52:19 What NVIDIA should do with their cash

00:56:09 Amazon’s cloud crisis

01:00:25 Building data centers

01:03:01 Anthropic’s role in Trainium

01:07:03 Oracle’s success

01:16:01 Datacenter buildouts

01:22:03 Hardware recommendations for startups

01:27:36 Understanding prefill and CPX

01:34:49 The state of GPU purchasing

Transcript:

This transcript has been edited lightly for readability.

Erik Torenberg

Dylan, welcome back to the podcast.

Dylan Patel

Thanks for having me.

00:00:33 NVIDIA’s Intel investment

Erik

It just so happens that there's some big news just as we're having you, NVIDIA announcing a $5 billion investment in Intel and them teaming up to jointly develop custom data centers and PC products. What do you think about the collaboration?

Dylan

I think it's hilarious that like NVIDIA could invest, it gets announced, and their investment is already up 30%. $5 billion investment to a billion dollar profit already. I think it's fun because they need their customers to really have big buy-in. So when their potential customers buy in and commit to certain types of products, it makes a lot of sense.

And it's kind of funny in a way because in the past, there was this whole thing around how Intel was sued for being anti-competitive with their chip sets. And NVIDIA actually got a settlement from Intel, way back when, when the graphics were separate from the GPU and the graphics were really put on the chip set, which had like all this other I/O, like USB and all this stuff.

So it's kind of a funny turn of events that now Intel is going to make a chiplet and package it alongside a chiplet from NVIDIA, and then that's like a PC product, right? So, you know, it's kind of poetic that everything has gone full circle, and Intel is sort of crawling to NVIDIA, but actually it might just be the best like device, right?

I don't want an ARM laptop because it can't do a lot of things. And so an x86 laptop with NVIDIA graphics fully integrated would be probably the best product in the market.

Erik

So are you optimistic? How do you think this will go?

Dylan

I mean, sure. I hope, right? I'm a perpetual optimist on Intel because I have to be. I was thinking that the structure of the deal that at least like a lot of the government folks and Intel were sort of trying to go for was big customers and the biggest suppliers directly give capital to Intel. But this is sort of the other way around, where they're buying some of the stock, having some ownership, but they're not really like diluting the other shareholders. And then the other shareholders will get diluted/everyone will get diluted when Intel finally does raise the capital from the capital markets. But because they've announced these deals, and they're pretty small, right? 5 billion NVIDIA, 2 billion SoftBank, US government was 10. These are still relatively small.

Guido Appanzeller

Pretty small, yeah.

Dylan

On the nature of things, right? I mean, last time I think I said Intel needs like $50 billion, right? Now when they go to the capital markets, it's better, and hopefully they get another, you know, couple of these announcements. There's all sorts of speculation that Trump is involved in getting these companies to invest. Now you know the government as well of course. And now, you know, is Apple gonna come invest and also do something with Intel? Or who else will come in? And that'll really boost investor confidence. Then they can dilute/go get debt.

Sarah Wang

Like a Warren Buffet coming into a stock. The Jensen is like the Buffet Effect for the semiconductor world. Guido, you were the CTO of the Intel Data Center and AI BU. What are your thoughts?

Guido

I think it's really good for customers and consumers in the short term, right? Having, having both Intel… And like specifically with the laptop market, right? Having the two collaborate is amazing.

I wonder what's gonna happen with any of the internal graphics or AI products at Intel, right? They might just push a reset and give up on that for now, right? They currently don't have anything competitive, right? There was the Gaudi effort. That's more or less done, right? There was the internal graphics chips, which never competed really at the high end, right?

So from that perspective, it makes a lot of sense for both sides. For Intel, they needed a breath of fresh air. They were sort of desperate. So I think it's a very good thing. I think AMD is fucked. If your two archnemeses suddenly team up, it's the worst possible news you can have, right? They were already struggling. Their cards are good. Their software stack is not right. They were getting very limited traction. And they now have a bigger problem that side. I think ARM is a little bit screwed as well because their biggest selling point was sort of like, “Look, we can partner with everybody that doesn't want to partner with Intel.” NVIDIA is probably the most dangerous of the future CPU competitors, right? And so they now suddenly have access to Intel technologies and might get in that direction. It remixes the card, right? I did not see this coming. I think it's an amazing development.

Sarah

Yeah. It'll be very interesting to see this play out. To Erik's point, packed news week, the other thing that we wanted to pick your brain on, since we have you here, Dylan, is the other news dropping on Huawei unveiling their kind of AI roadmap. And you know, obviously they're hyping up the capabilities.

I think you guys have been sort of ahead of the curve of trying to gauge hey, what can the 950 SuperCluster actually do? But would love your thoughts on everything that's going on from the China front, right? And this is kind of coupled with DeepSeek saying their next models are gonna be on domestically produced Chinese chips, the Chinese government kind of banning companies from buying the produced specifically for China NVIDIA chips.

So there's just sort of a lot of dominoes falling right now in the semi market in China. But would love your take overall and I mean, drill into some detail.

Dylan

Yeah, I think when you sort of zoom out to even like, you know, let’s walk from 2020 because I think it's really important to recognize how cracked Huawei is, or even just historically, like they've always been really good.

Sure initially they stole like Cisco source code and firmware and all this stuff, but then they rapidly passed them up as well as every other telecom company. In 2020, they released an Ascend chip and submitted it to impartial public benchmarks. And they were the first to bring 7 nanometer AI chips to market. They were the first to have that. Now you could still say NVIDIA was ahead, but the gap was like nothing, right? And this was when they could access the full foreign supply chain. This was when they just passed Apple to be TSMC'S largest customer. They were clearly ahead of everyone on a manufacturing supply chain sort of design standpoint on a total basis.

Now, of course, NVIDIA still had higher market share, but it was so nascent then, like they could have really taken over the market. Huawei got banned by the Trump one administration from accessing, and then it went into effect in 2020, the full ban. And so they were only able to make a small volume of these chips, but they had trained significant models on these chips that they made then.

And then over the next couple years, NVIDIA continued to accelerate. Huawei, because they were banned from TSMC, had to go and try and figure out how to manufacture at SMIC, the domestic TSMC. And then they were also in parallel trying to go through shell companies to manufacture at TSMC and acquire memory from Korea and so on and so forth.

So by the end of ‘24, they had this had gotten in full swing and it was caught, right? It was caught, and they finally shut it down. But they were able to acquire 3 million chips, 2.9 million chips from TSMC through these other entities. Roughly $500 million worth of orders, which ends up being a billion dollar fine that the US government gave TSMC if I recall correctly. Or at least there was a Reuters article. I don’t know if they actually issued it, which is important and interesting to gauge because the number of Ascends floating out there is has not consumed this entire capacity yet.

So now we get to 2025. The H20 got banned in the beginning of the year. NVIDIA had to write off, you know, huge amounts of money. Our revenue estimate for NVIDIA and China for just H20 was north of 20 billion because that's what they were booking in capacity/had to write off.

And then it got banned. They cut the supply chain, like they just said, “No, we're not doing this anymore.” They had their inventory, it gets reapproved, they resell the inventory, but now they're like, “Do we even restart production?” is NVIDIA's question. And now you have China saying, “Hey, we don't need NVIDIA, we have domestic alternatives.”

Whether it be Huawei or Cambricon, these companies have capacity, but most of this capacity is still foreign-produced, right? Whether it be wafers from TSMC, memory from Korea, Samsung and SK hynix. So the question is sort of like, how much can they do domestically?

And there's sort of two fronts there. There's the logic, i.e., replacing TSMC, and there's the memory, i.e., replacing hynix, Samsung, Micron. And on the logic side, they are behind, but they're really ramping there, and I think they can sort of get to the production capacity estimates needed, and the US is still allowing them to import all the equipment necessary, pretty much. The bans are really for beyond the current generation of technology, beyond 7 nanometer. The bans are really for 5 nanometer and below, even though the government says they're for 14 nanometer, the actual equipment that's banned is only for below 7 nanometer. And so they'll be able to make a lot of 7 nanometer AI chips and maybe even get to 5 using existing equipment for 5 nanometer rather than like taking the new techniques.

And so like there's the logic side, and then there's the memory side. And the aspect of Huawei's announcement that was surprising was that they're doing custom memory, right? That's, that's the part that is sort of like, hey, this is really exciting. They announced, you know, two different types of chips for next year, one that's focused on recommendation systems and prefill, and then one that's focused on decode.

Guido

There's the trend these days.

Dylan

Yeah. And NVIDIA, the same thing. They just announced a prefill specific chip recently. There's numerous AI hardware startups that are really focusing on prefill versus decode. And so the sort of split of inference into two workloads. You know, Huawei is doing the same thing for their next year chip. And what's interesting is the decode one has, you know, custom HBM. What does that mean? What is the manufacturing supply chain? That's the one that's tricky, right?

How much can they manufacture of that custom HBM? And NVIDIA and others are also adopting custom HBM only starting next year. You know, yes, the manufacturing capacity is not there. It is gonna consume a bit more power. It's gonna be slightly lower bandwidth. But the fact that they're able to do, you know, some of the same things that NVIDIA plans to do, AMD plans to do in their memory is evidence that they're catching up.

But then, you know, the main question that remains is production capacity. So as far as like, hey, NVIDIA is banned in China, right? Like they're saying don't buy NVIDIA chips. I think for a period of time that's fine because—fine for China, right, from a perspective of “Hey, I'm China,” that's fine because you have all this capacity that you shipped in in 2024, they haven't turned into AI chips.

Now you're turning them into AI chips, you're running all that stockpile down. What about the transition from running that stockpile down to ramping your new stuff? And that, that, that transition is the one that's really tricky. China's either shooting itself in the foot by not purchasing NVIDIA chips during that time period, or China is able to ramp. I think they'll be able to ramp, I think it'll take a little bit longer. And there will be sort of a gap in between where China probably backtracks and says it's fine. Like ByteDance is like begging for NVIDIA chips, right? They use some Cambricon, they use some Huawei, but they really want to use NVIDIA because it's way better.

They don't care about the domestic supply chain. They wanna make the best models. They wanna deploy their AI as efficiently as possible. And so this is like, you know, the government can mandate them to like not do it, right? So it's not that NVIDIA is not competitive, it's that the government is sort of trying to instigate it.

And then like, I guess the last sort of thing is like, you know, there's always the argument of like, hey, if banning NVIDIA chips to China is so good for China, why didn't China do it for itself? And they're finally doing it for themselves. So again, like, it'll be interesting to see.

Smuggling is still happening, right? Re-exportation of chips from, you know, other countries to China, that is still happening, at some volume, low or medium volume, right? But then, you know, the direct shipments of NVIDIA chips that are legally allowed to China are not necessarily happening today, but may have to restart at some point because China won't have the production capacity to… You know, they would just have so many fewer AI chips being deployed domestically versus the US. And at some point you kind of have to pick like, am I all about the internal supply chain, or am I all about chasing, you know, super powerful AI.

Guido

So is there an angle here about a negotiation angle as well? Because currently there's still discussions ongoing. What exactly are the boundaries? What can be exported to China? So these are sort of well-timed announcements if you want to make the point that, you know, the US should allow more exports. Do you think that's a factor or not?

Dylan

Yeah, so in the report we did a few weeks ago about the production capacity of Huawei and the supply chain, there was a bit in there that we wrote about how, you know, honestly, like if you were China and you do want NVIDIA chips, actually, how do you play this? And it's by hyping up your domestic supply chain. It's like, yes, we can do everything. It's Huawei, announce the most crazy shit possible, announce three years of roadmaps that are…

Guido

So you think they read your report, basically.

Dylan

No, no. I think they knew, I mean, they were already bid and then like, say we're banning NVIDIA, right? Then the government official is gonna think, alongside sort of like lobbying from domestic players, like “Of course we wanna ship them better AI chips, like, we're losing this market. We can't lose this market.” And it's sort of like, it is 10,000 IQ, right? And we're here playing checkers while they're playing chess.

00:15:01 Is Huawei just hype?

Sarah

Well, so I guess negotiating chip aside, in that report you talked about HBM or high bandwidth memory being a bottleneck to Huawei. To your point on one of the surprising aspects of the announcement, do you think it's credible that it's no longer a bottleneck based on what they're saying? Or is it just hype?

Dylan

I think production-capacity-wise, it is still absolutely a bottleneck. Certain types of equipment required for making HBM need to be imported. They're working on domestic solutions, but as far as we know, they have not imported enough equipment for this.

Although, if you look at Chinese import data for different types of equipment, right, there's sort of like fabs spend, you know, roughly, it depends on the process technology, but fabs spend roughly different amounts of money on lithography, etch, deposition, metrology, right? Like these different steps.

And historically lithography has hovered around, you know, 17, 18%. With EUV it grew to 25%, right? But China, because they wanted to stockpile lithography, and they were worried about the becoming banned, they were importing lithography at a much higher rate than that, right?

Like 30, 40% of their equipment imports were lithography. And they were just stockpiling lithography equipment. This is sort of like reversed now in that like, if you look at the monthly import export data both into provinces in China, but also out of countries, you can see that etch specifically is skyrocketing.

And the main thing about stacking HBM is that when you have each wafer, you have to create like this thing called a through-silicon via so it can connect from the top to bottom, and then you stack them on top of each other, right? 12 high, 16 high for HBM. That's how you make super high bandwidth memory. And their import for etch is like skyrocketing now.

They don't have the production capacity yet. How fast can they ramp it as a function of how much equipment can they get, A). And B), like the yields, right? Improving yields is really hard on manufacturing. Intel and Samsung are really good, and TSMC is just amazing. Not that those companies suck, like, I think is a better way to put it.

And so, you know, it's those two things. I think yield, they haven't even started production of HBM3. They've only done some sampling of HBM2. HBM3 came out a few years ago. So there's still quite a bit of ways to go on like going up the learning curve. Obviously I expect them to catch up faster than it took, you know, the technology to be developed because it exists in the world. We know how to do it. It's just a matter of actually doing it versus inventing it. And then the other one is sort of the production capacity. You know, a couple months of import-export data is not enough to set up for years’ worth of supply chain build up, which is what we have today in Korea, for the Korean companies.

Now, hynix is also investing in the US in Illinois, and then Micron is primarily in Japan, the American memory companies primarily in Japan and Taiwan, but they're also expanding in Singapore and the US now. Like there's so much capital that's been invested, it would take some time for China to build up that production capacity to actually match the West. And when I say, “the West,” I mean non-China East Asia in production capacity. So it'll take some time to get there. And I think it's like, “Hey, we can design this.” It's always a question of “Can we manufacture?” And then the thing like that Jensen would say is like, you're betting on China not being able to manufacture, like… You know, it's a matter of when not if. And that's the whole calculus that I think the US government has to be aware of when they're like, “Hey, what level of AI chips do we sell? Do we sell everything?” Probably not because AI is far more powerful, and the end market of AI is gonna be way larger than the end market of semiconductors and equipment.

You know, what level do we sell at? Well, how much can China make at each specific, you know, sort of performance tier. And then, you know, analyze that, and what's the volume, and then figure out what is okay, which is like maybe a little bit above or around the same level.

00:19:03 Advice to Jensen

Sarah

Yeah, so, to your point on like playing chess versus checkers, if you're Jensen, what would your next move be, given the situation at hand?

Dylan

It's both partially true that he’s afraid of Huawei more than he is like an AMD, right?

Sarah

He called him “formidable.”

Dylan

I mean… Like Huawei has beat Apple, right? They passed Apple up in TSMC orders. They passed Apple up in phone market share, not in the US but like in many parts of the world, before the bans came down. And then even now they're growing back again in market share without like western supply chains. You know, they've done this to numerous other industries. I would say Apple's like a formidable competitor, right?

They've beaten a lot of industries, and so it's reasonable that he is afraid of them. And he's not afraid of AMD. I think like the best thing is like try sow as much like what Huawei announced is reality rather than like their hope target.

And sow away all doubt on manufacturing capacity, which I think is not fair, right? Like I think manufacturing capacity is a real bottleneck for them. And then the yield learnings, real bottleneck. Like temporary maybe. We’ll see how long, and we'll see how fast the rest of the NVIDIA technology advances past what Huawei is capable of and how fast Huawei is able to close the gap. But I think his main sort of pitch would be, Huawei is real. They're a formidable competitor. They're going to take over not just the Chinese market, but also foreign markets. Whether it be the Middle East or Southeast Asia or South Asia or Europe or LATAM, right.

Everywhere besides America. Noah Smith has this analogy, right? This whole idea is that you should Galapagos China, right? Make them have their own domestic industry that is so different from the rest of the world, right? Kind of what happened with Japan in the ‘70s and ‘80s and ‘90s. Their PCs were so specific and hyper-optimized to the Japanese market would like… I don’t know if you've seen the weird scroll wheel on these Japanese PCs. It's like you go like this and it scrolls, right? And then the touch pad is a circle, and then that's around it.

It's like things like that are so weird that the rest of the world doesn't care. But Japan market likes it, right? And his whole idea is like, let's Galapagos them, i.e., keep their technology within China, and then that's like dead weight loss and they never expand outside, versus, that we serve the whole world.

But the whole risk is that the opposite can also happen, right? Our technology is hyper-optimized to running language models at this scale and RL. Hardware-software co-design can take you down a path of the tree that like is a dead end. And then China, like, because they're not allowed to access this tree, they're like, “Oh, okay.” Then they end up in the like optimal spot, right? We hit a local maximum, they hit a global maximum. That sort of like technological Galapagos-ing is sort of what Noah Smith's analogy is. I like it a lot. I don't know if it's accurate, but it's an interesting one.

00:22:25 NVIDIA bull and bear cases

Sarah

Yeah. I love that. Well actually, maybe just taking a step back from current events, even though there's so much to talk about right now, last time you appeared with us, NVIDIA came up, obviously, and you talked about a couple of the potential paths forward for NVIDIA.

Erik

Give us maybe the bull case, the bear case.

Dylan

There's a lot embedded in their numbers now. But what's interesting is consensus for the banks is like for across the hyperscalers, so, Microsoft, CoreWeave, Amazon, Google, and Oracle, Meta. So it's the six hyperscalers, right, who I would consider hyperscalers.

The consensus for the banks is $360 billion of spend next year across all of them. And my number is closer to like 450, 500. That's based on like, you know, all the research we do on like data centers and like tracking each individual data center in the supply chains.

Guido

This is just NVIDIA spend, right?

Dylan

This is CapEx for the hyperscalers. That CapEx gets split up across different companies, but the vast, vast majority still goes to NVIDIA.

NVIDIA is where they can't take share, right. They grow with the market/defend share. And so the question is like, how fast is the growth rate of CapEx for hyperscalers and other users? And the reason I included Oracle and CoreWeave as hyperscalers, even though they're traditionally not called hyperscalers, is because they're OpenAI’s hyperscaler.

So, you know, when you look at the Oracle announcement, right? First of all the Oracle announcement, I don't understand why people don't think this is crazier. They did the most unprecedented thing in the history of like stocks and companies ever. They gave a four year guidance.

And it made Larry the richest man in the world. You know, like all these things. Anyways, you know the question is like, how fast does revenue grow?

Do you think OpenAI, which signed a $300 billion plus deal with Oracle, will actually be able to pay $300 billion across raising capital and revenue? And it gets to a rate of like over $90 billion a year, in just a handful of years. So it's like, do you believe the market will grow that fast? It's very possible. And it's very possible for like, you know, OpenAI, what is their revenue gonna be exiting next year?

Some people think 35 billion, some people think 40 billion. Some people think 45 billion ARR by the end of the year, next year. This year they hit 20 ARR, you know, so if that growth rate is maintained, then all of that cost goes to compute plus all the capital they continue to raise.

And again, there are financials that they sort of like gave to investors for their last round was like, “Hey, we're gonna burn like $15 billion next year. It's probably more likely gonna be like 20. And you stack this on and they're not turning a cash flow, they're not gonna be profitable until 2029.

So you sort of have like, they're gonna continue to be burn 15, 20, $25 billion of cash each year, plus revenue growth. That's their compute spend. And you do this for Anthropic, you do this for OpenAI you do this for all the labs. It's very possible that the pie does get to, you know, you know, more than 500, you know, not 360 billion next year, 500 billion next year for total CapEx. And the pie continues to grow for hyperscalers. NVIDIA says, actually it's gonna be multiple trillions a year on AI infrastructure, and he's gonna capture a huge portion of it. That's his bull case, right? That's the bull case is AI is actually so transformative, and the world just gets covered in data centers, and the majority of your interactions are with AI, whether it's like, you know, business productivity and telling an agent to do some code, or you're just talking to your AI girlfriend Ani, right? Like, it doesn't matter. You know, all of this is running on NVIDIA for the most part. The bear case is, you know, even if it does grow a lot…

Guido

Save the bear case for a second. I think fundamentally the value creation, I think, personally, is there, right? I mean, to create trillions of dollars of value with AI, I can totally see this happening. So assume it's true, where will NVIDIA top out?

Dylan

I guess, how much do you believe in takeoffs, right? So like, if there is like a takeoff scenario, right, where like powerful AI builds more powerful AI builds more powerful AI, or, you know, that creates more and more… Each level of intelligence enables more for the economy, right? Like how many monkeys can you employ in your business versus how many humans, right? Or how many dogs, right? There’s sort of like, what is the value creation of a human versus a dog? Sort of like the same with AI.

In this case, the value creation could be hundreds of trillions, if not…

Guido

Do you need this? I mean, if you take every white collar worker and make them twice as productive with AI, that's in the hundreds of trillions, isn't it?

Dylan

Yeah but like, what is twice? You know, like, I mean like if you talk to people at the labs, right? Like twice as productive, what does that even mean? It's replace them, right? And it's be 10 times better than them. I mean I don’t know how soon…

Guido

If white collar work is essentially useless without a constant stream of LLM tokens, that make them productive, right? At that point, you basically can tax every single knowledge worker in the world, right? Which is most workers in the world long term. I mean, what's your guess? Give us a number. What's the cap for NVIDIA?

Dylan

The cap? I mean why aren’t we making a Matrioshka brain? I don't know, at some point the machine says humans don't need to live and I need even more compute.

Guido

One step before that.

Dylan

Are we colonizing Mars yet?

Guido

TBD.

Dylan

I don't know, man. I find it completely impossible to predict anything beyond five years given how much stuff is changing. I'll leave it to economists, right? Honestly like, you know, supply chain stuff is like three, four years out and that's it. And then fifth year is sort of yellow, right? I just try and ground myself with the supply chain stuff, right?

Supply chain, and then like, what is the adoption of AI, and what's the value creation, what's the usage? And you can see that in like a short horizon beyond that like, I don’t know, are we all gonna be connected to computers, like BCIs and stuff? Like, I don't know, dude. Are humanoid robots, are they gonna be… I mean you saw Elon's thing, right?

He's like, yeah, humanoid robots are why Tesla is worth more than 10 trillion. It’s like, okay, great. What is all that being trained on? Great, NVIDIA. Okay. Awesome. So then that's worth also 10 trillion, right? Like, I don't know, like, it's too out there for me. I don't like the out there discussions.

Sarah

Very fair.

Dylan

Read some sci-fi books.

00:29:23 NVIDIA’s moat

Sarah

So just pulling out the thread where you talked about, I mean, this is kind of a throwaway comment, but how market share can't really grow just because it's such a dominant market share. And you guys talked about the moat of NVIDIA last time. And obviously this moat is tied to maintaining that very high market share that they currently have. And I love this sort of historic journey you took us through with Huawei just earlier. Can you kind of walk through what NVIDIA did throughout history to build their moat?

Dylan

It's super awesome because, you know, they failed multiple times in the beginning, and they bet the whole company multiple times, right?

Jensen is just crazy enough to bet the whole company. Whether it was certain chips ordering volume before he knew it even worked, and it was all the money he had left, or ordering volumes for projects he had not won yet. I heard a rumor that, or not a rumor, but like a story from someone who's like a graybeard in the industry and I think would know was like, “Yeah, no, no, no, like NVIDIA ordered the volume for the Xbox before Microsoft gave them the order.” He was just like, “Fuck it. Yolo.” I'm sure there's more nuance there, like, you know, verbal indication or whatever, but like the order was placed before he got the order, right, is what he said. You know, there's cases like with the crypto bubbles, right? Like there was a couple of them, but like NVIDIA did their damn best to convince everyone in the supply chain that it wasn't crypto and that it was gaming. That it was durable real demand and it gaming and data center and professional visualization and therefore you guys should ramp your production.

And they all ramped production and spent all this CapEx on increasing production and, and building out new lines for them. And they pay per item. And then they bought them and sold them and made shitloads of money. And then when it all fell apart, they just had to write down a quarter's worth of inventory, whatever.

Everyone else was like, “Well, crap, I have all these empty production lines.” But like, what did AMD do then? Their chips were actually better for crypto mining on an amount of silicon cost versus how much you hash. But like, they just didn't. AMD was like, “Ah, we're gonna not really raise production.”

Like, as a reasonable, you know, thing. So it's sort of like, strike while the iron is hot. And so like, you know, the same has happened with NVIDIA, right? In recent times, like sort of, they've ordered capacity that no one believes multiple times. They see them in demand, obviously, but in many cases they're just like, their number for like Microsoft was higher than Microsoft's internal planning, right?

And then Microsoft's internal planning went up, but like their number for Microsoft was way higher. And it's like, “Ah, we just don't think Microsoft is gonna need this much, even though they tell us this.” It's like, who the heck? It was like, no, no, no customer, you're gonna buy more like, and, and orders.

Right? And then when the orders come through the supply chain, it's like, I have to pay NCNR, non-cancellable, non-returnable. I asked a question in Taiwan once, it was Colette, which is the CFO, and Jensen, CEO. They were both there.

It was a room full of like mostly finance bros and they were asking stupid finance questions like three days before earnings. So obviously they just could not answer anything because it's like, you know, SEC regulations. But then my question to them was like, “Look Jensen, you're like, so vibes-driven and like very gut feel, and like very visionary. And then Colette's, you know, CFO, like she's amazing in her own right, but like, those personalities clash. How do you work together?” And he's like, “I hate spreadsheets. I don't look at them. I just know.” Like that's his response. And it's like, of course, you know, the best innovators in the world have really good gut instinct.

And so the gut instinct to order with non-cancellable when you don't know. And they've had to write down, over their history, multiple times, many, many billions of dollars in total orders, whether it be, you know, the H20, which is more regulatory, but like other cases they've ordered and had to cancel.

Guido

Is it many billions?

Dylan

It's many billions.

Guido

Peanuts.

Dylan

Well it depends, right? The crypto writedown was like multiple billion when their stock was like less than a hundred billion.

Guido

It’s peanuts compared to the upside, right?

Dylan

I think everything he did was right. And I think everything AMD did was wrong. Like, you know, in that scenario. But like, it is crazy to, especially in a cyclical industry like semiconductors where companies go bankrupt all the time, which is why we have all this consolidation is, every down cycle companies go bankrupt.

Guido

I mean, if you look from a risk-return perspective, right, these bets were totally worth taking. If you look at it from “I'm a CEO, I want to have predictable quotas for Wall Street,” it's a very different story, and I think that's sort of where part of the tension is from.

Dylan

I don’t know if you've seen these like, Lee Kuan Yew edits where they're him saying some like fiery speech,and then it's like some cool music at the end, and it's like showing different pictures of him.

And so we made one of Jensen recently and put it on social media, on like Instagram, TikTok, XHS, red book, Twitter, of course, like all the different social media. And I really liked it because he's like, you know, the goal of like playing is to win, and the reason you win is so you can play again.

And he compared it to pinball where like actually you just play all day and you keep getting more rounds. And it's like, his whole thing is like, I want to win so I can play the next game. And like it's only about the next generation, right? It's only about now, next generation. It's not about fifteen years from now because it's a whole new playing field every time. Or five years from now.

You're right. The risk-reward is correct. Few people take these kind of risks. It's the only semiconductor company that's worth, I think, even north of $10 billion, that was founded as late as it was. Like MediaTek was in the early ‘90s and then NVIDIA. And everyone else is like from the ‘70s mostly.

Sarah

I think you raised this great point on, bet the farm. And he's actually been wrong a couple times, to your point.

Dylan

Mobile, right? Like what the hell happened with mobile?

Sarah

Exactly. And he still takes them. And I think, Marc actually had this great conversation with Erik where he talked about being founder-run, where you have this memory of the risks you took to get to where you are today.

And so in a lot of cases, if you're a CEO brought on later on, you're sort of like, “Okay, continue to steer the ship as is.” But in this case, he remembers all the times they almost went belly up and he’s like, “I’ve got to keep making bets like that.” How do you think he's changed? I mean, he's been one of the longest running CEOs. He's kind of right up there with Larry Ellison now. How do you think he's changed over the last 30 years or so?

Dylan

I mean obviously like I'm 29. I don't freaking know what he was like.

Sarah

Fair.

Dylan

I've watched a lot of old interviews.

Sarah

He’s been CEO longer than you’ve been alive.

Dylan

NVIDIA was founded before I was born. I'm ‘96, right?

Sarah

Yeah, maybe anything over the last couple of years,

Dylan

I think even like watching old interviews, right? Like, I watched a lot of old interviews, a lot of old presentations he’s given. One thing is that he's just like sauced up and dripped up. Like the charisma he's gotten has only gotten stronger, which is an interesting point. I don't know if it's quite relevant.

Sarah

Totally agree with that.

Dylan

But like the man has learned to be a rockstar more, even though he was always charismatic. It was like he is a complete rockstar now. And he was a rockstar, you know, a decade ago too. It's just people maybe didn't recognize it. I think the first live presentation that I watched, it was extreme, it was CES, like 2014 or 2015 or whatever. It's Consumer Electronics Show. I'm moderating like gaming hardware subreddits, right? At the time I'm a teenager. And like the dude is talking only about AI. He's telling all these gamers about AlexNet and self-driving cars.

It's like, know your audience, first of all, but also, like… It has nothing to do with consumer electronics and gaming. At the time, I was half like, “Holy crap, this is amazing.” But I also was half like, “I want you to announce new gaming GPU.”

On the forums, quickly, everyone was like, you know, screw this. I want to hear about the gaming GPUs, NVIDIA is price gouging. You know, of course NVIDIA has always had the like “We priced to value plus a little bit because we were just smart enough to know…”

You know, I am guessing Jensen just has the gut feel of how to price things. At least on gaming launches, he'll change the price up until right before the presentation. It really is like a gut feel thing, probably. And anyways, so he had that charisma to know what was right.

But I think a lot of people were like, “Oh, no,” whatever, “Jensen is wrong. He doesn't know what he's talking about.” But now, he talks, people are like, “Oh, very…” So it might just be that he's been right enough.

Sarah

Yeah. There was a post on X recently that said he had moved up into Godmode with a select group of CEOs.

Dylan

Who's the other gods?

Sarah

It was Zuck. Who was the other God?

Erik

Elon?

Sarah

Elon! Elon, Zuck, and Jensen. Good crew to be in.

Dylan

So when we pray to Silicon Valley…

Guido

The cult now, is it?

00:39:33 Potential successors to Jensen

Sarah

Exactly. Just one last thing on people. You mentioned Colette, his CFO, and you know, there's sort of a famously loyal crew at NVIDIA, even though all of the OGs could retire at this point.

Is there anyone akin to a Gwynne Shotwell at SpaceX or previously a Tim Cook to Steve Jobs at Apple that is at NVIDIA today?

Dylan

I mean, he had two co-founders, right? Let's not overlook that. One of them is like, you know, not involved and hasn't been for a long time, but the other one was involved up until just a few years ago. So it's not just Jensen running the show. Although he was running the show. There's quite a few people on the hardware side. There's someone at at NVIDIA that's like mythical to me. Like when you talk to the engineering teams, he leads a lot of the engineering teams. He is a private person, so I don't wanna say his name actually.

Sarah

Okay, fair enough.

Dylan

But you know, he’s effectively like Chief Engineering Officer, is like his role. And people within his org will know who he is. I think there are people like that. But you know, he's intensely loyal and there's a number of these types of people.

There's another fella who's like, you know… Like there's all these like innovative ideas at NVIDIA, and he's the guy who literally is like, “We need to get this silicon out now. We're cutting features.” And that's what he's famously known for. And all the technologists in NVIDIA hate him. This is like a second guy. This is a second guy. Also intensely loyal to NVIDIA has been around for a long time, but when you have such a visionary company and forward, one problem is that you get lost in the sauce, right? “Oh, I want to make this, it's gotta be perfect, amazing.” And these people are like, you know, obviously they're close to Jensen for a reason because Jensen also believes like these things, right? Have the visionary future looking, but also like, screw it, cut it, we'll put it in the next one, ship. Ship now, ship faster, in a space like silicon, which is really hard to do so. The thing about Nvidia that's always been, you know, super impressive, and it's from the beginning days, where he's talked about this before, is their first chip, their first successful chip, they were gonna run outta money, and he had to go get money from other people, to even finish the development. And even then he just had enough money, because he'd already had a failed chip before this, was the chip came back and it had to work, otherwise it would not, you know… Because they could only pay for, it's called a mask set, right?

Basically you put these, like, I'll call them “stencils” into the lithography tool, and then it like says where the patterns are, and you put the stencil in, you deposit stuff, you etch stuff, you deposit materials on the wafer, etch it away. And you put the stencil in and like you like tell it where to put stuff, right?

And then the deposition and etch keeps happening in those spots, and you stack dozens of layers on top of each other. And then you make up a chip. These stencils are custom to each chip, right? And they cost today in the orders of tens and tens of millions of dollars. But even back then, it was still a lot of money.

It wasn't that much then, of course. They could only pay for one set. But the typical thing with semiconductor manufacturing is, you know, as good as you can simulate it, as good as you can do all the verification, you'll send a design in, and you have to change it.

There's gonna be something. It's so hard to simulate everything perfectly. And the thing about Nvidia is they tend to just get it right the first time. Even great executing companies like AMD or Broadcom or whoever, they often have to ship, you know… They're denoted in like “A” and then a number, or “B” and then a number.

So it's like two different parts of the masks. So like, NVIDIA always ships A0, almost always. They sometimes ship A1. The A is basically the transistor layer. Then the number is like the wiring that connects all the transistors together.

So NVIDIA will start production of the A and ramp it really high and then just hold it right before you transition to the metal, just in case they do need to change the metal layers. And so like the moment they're ready, and they've confirmed that it works, they can just, you know, blast through a lot of production.

Whereas everyone else is like, “Oh, let's get the chip back. Oh, okay, A0 doesn't work. We gotta make this tweak. Make this tweak.” It’s called a stepping, right?

Guido

At Intel we were very jealous of NVIDIA at that time, right? They consistently delivered in the first one. We did not.

Dylan

The data center CPU group, there was one product where, you know, I said “A0,” “A1,” or you go to B if you have to change the transistor layer as well. So it's like “B…” Intel got to like “E2” once. E2, that's like a 15 revision. This is like the peak of AMD's, like when they went skyrocketing on market share versus Intel was when Intel was at E2. Like 15 steppings.

Guido

It causes quarters of delay. I mean it's catastrophic for a go-to-market.

Dylan

Yeah. Each time is a quarter of delay or something. It's absurd. So I think that's the other thing about NVIDIA is, like, you know, screw it, let's ship it. Let's get the volume ASAP. And so anyways, they have some of the best simulation, verification, etc, that lets them sort of go from design, you know, from idea to shipment as fast as possible.

You know, cutting out any unnecessary features that could delay it. Making sure they don't have to do revisions so that they can respond to the market ASAP. There's a story about how Volta, which was the first NVIDIA chip with tensor cores… You know, they saw all the AI stuff on the prior generation P100, Pascal, and they decided “We should go all in on AI.”

And they added the tensor cores to Volta, like only a handful of months before they sent it to the fab. Like they said, “Screw it, let's change it.” And it's like if they hadn't done that, maybe someone else would've taken the AI chip market, right? So there's all these times where they just… And those are major changes, but there's often like minor things that you have to tweak, Number formats or like some architectural detail. NVIDIA is just so fast.

Guido

And the other crazy thing is they have a software division that can't keep up with that, right? I mean if you come out with a chip, and basically no stepping required, it's immediately in the market, then being ready with drivers and, you know, all the infrastructure in total. That's just super impressive.

NVIDIA’s business in 10 years (00:46:23)

Sarah

Yeah, I love that point because you think of NVIDIA benefiting from tailwind after tailwind, but I think both of you are saying you have to move fast enough and execute well enough to take advantage of those tailwinds. And by the way, I loved your CES story. I'm just envisioning him more than 10 years ago, talking about self-driving cars. But you know, if you think about nailing the videogame tailwind, VR, Bitcoin mining, obviously AI now, you know, one of the things that Jensen talks about today is robotics, AI factories. Maybe my last question on NVIDIA, what do you think about the next 10 to 15 years?

I know calling beyond five is hard. But like, what does NVIDIA's business look like?

Dylan

It's really a question of, and this is like… I think every time I've talked to, you know, some executives at NVIDIA I’ve asked this question because I really wanna know, and, you know, they won't answer it obviously, but it's like, what are you gonna do with your balance sheet?

Like you are the most high cash flow company. Like you have so much cash flow. Now the hyperscalers are all taking their cash flow like way down, right? Because they're spending on GPUs. What are you gonna do with all this cash flow? Even before this whole takeoff, he wasn't allowed to buy art, right?

So what can he do with all this capital and all this cash, right? Even this $5 billion investment in Intel is, there's regulatory scrutiny there, right? It's in the announcement, like, yeah, this is subject to review, right? I imagine that'll get passed, but he can't buy anything big.

He's gonna have hundreds of billions of dollars of cash on his balance sheet. What do you do? Is it start to build AI infrastructure and data centers? Maybe? But like, why would you do that if you can just get other people to do it and just take the cash?

Guido

Well, he's investing in those, right?

Dylan

Investing peanuts. You know, he gave recently like CoreWeave a backstop because today it's really hard to find a large number of GPUs for burst capacity. Like, hey, I want to train a model for three months. I have my base capacity where I’m doing all my experiments, but I want to train a big model 3 months down.

Guido

We know from our portfolio.

Dylan

NVIDIA sees this issue, they think it's a real problem with startups. It's why the labs have such an advantage. You know, right now, most companies in the Valley spend, what, 75% of their round on GPUs, right?

Sarah

At least.

Dylan

What if you could do 75% in three months on one model run, right? And really scale and have some sort of like, competitive product, and then you have the model. And then you raise more capital, or start deploying. What do you do with it? Is it start buying a crapload of humanoid robots and deploying them, but like, they don't make really that amazing software for them, in terms of the models, right? The layer below is great. Where they deploy their capital is a question.

Guido

He has been investing up and down the supply chain a little bit though, right? Investing in the neoclouds, investing in some of the model training companies.

Dylan

Yeah. But again, small fries. He could have just done the entire Anthropic round if he wanted to. Of course he didn't. Right. And like really gotten them to use GPUs. Or like he could've done the entire, you know, OpenAI round. He could have done the entire, like any xAI round.

Erik

Do you think these are things he should be doing?

Dylan

I mean like, I don’t know, right?

Guido

We’ll quote you for the next round…

Dylan

He could make venture a dead industry. Take all of the best rounds.

Sarah

Put us all out of business, yeah.

Dylan

You know, you can do the seeds and then have Jensen mark you up. I think picking winners is obviously really tough for him because he has customers all across the ecosystem.

If he starts picking winners, then his customers will be even more anxious to leave and give even more effort to, whether it's AMD or you know, some startup or their internal efforts, etc, etc, right? Buying TPUs, whatever it is. He can't just like invest in these, like, you know, he can do a little bit, right?

A few hundred million in an OpenAI round is fine, or a few hundred million in an xAI round is fine. CoreWeave, right, like, yeah, everyone's like throwing a fuss about it, but it's like he invested a couple hundred million early on plus, rented a cluster from them for internal development purposes instead of renting it from hyperscaler, which is cheaper for NVIDIA to do, right?

It's better for them to do it from them than the hyperscalers. It's like, is he really backstopping CoreWeave that much? Right. Or, you know, any of the other customers or neo clouds, like there's some investment. But it's more like this is a google cloud, you know, we'll throw like 5 or 10% of the round. It's not he’s taking 50% plus of the round.

Guido

Is he also reshaping his market? I mean, look, a couple of years ago there were four big purchases of these cards. You just listed six. To what extent is that… Is that a strategy?

Dylan

It is. I think it absolutely is. But he didn't have to put much capital down to do this.

Guido

Just chip one earlier than the other? I don't know.

Dylan

No, but it's like, if you look at the grand amount of capital that he spent investing in the neo clouds, it’s a few billion dollars.

Guido

But he has lots of other levers if he wants to.

Dylan

Right, right. Allocations, as you mentioned. What's nice is, you know, historically, you gave volume discounts to hyperscalers. But because he can use the argument of antitrust, he's like, everyone gets the same price.

Guido

So fair.

Dylan

It’s very fair. It's very fair. You know.

00:52:19 What NVIDIA should do with their cash

Erik

So what should he do with the cash? Or what to guide his…

Dylan

I mean, I think, there's the argument that he should invest in data centers. Only the data center layer, not what goes in the data center so that more people build data centers, and then if the market demand continues to grow up, data centers in power are not the issue.

Invest in data centers and power. I've said that to them. They should invest in data centers and power, not in the cloud layer. Because the cloud layer is, not commoditized, but it's… Commoditize your compliment right? Is the whole phrase. And I won't say being a cloud is commoditized, but it's certainly like you have a lot of competitors who are decent now.

And you've educated the commercial real estate and other, you know, infrastructure investment firms going into AI infra as well. So like, I don't think it's the cloud layer that you invest in. Do you invest in data centers and energy? Yeah because that's the bottleneck for you in growth really. Is A), well, how much people wanna spend and can't spend, and B), the ability to actually put them in data centers. And then like robotics. And like, I think there's like areas he could invest in, but nothing requires $300 billion of capital. So what do you do with the capital? Like, I really don't know, and I like feel like Jensen has to have some idea, there's some visionary plan here because that's what shapes the company, right?

I mean they could just continue to, you know, I mentioned $200 billion of free cash flow, $250 billion of free cash flow a year. What do they do with it? Like, do they just buy back stock forever? Like, do they go Apple route? And the reason why Apple hasn't done anything interesting in like, you know, nearly a decade is, you know, they've got, they've got a not visionary at the head.

Tim Cook is great at supply chain. And they're just plowing the money into buybacks. You know, automotive. The self-driving car thing failed. We'll see what happens with AR/VR. We'll see what happens with wearables, right? But like, Meta and OpenAI might be even better than them. We'll see.

Like and others, right? So what does he invest in? I have no clue. What requires so much capital, is the tough question, and actually gets a return. Because the easy thing is like my cost of equity, right, I just buy back stock.

Guido

And doesn't completely change the company culture. I think that's another thing, right? There are probably areas he could invest it in, but you suddenly end up with the company doing two completely different things, which are very difficult to keep…

Dylan

But they do like 10 completely different things, right? I mean, one way to look at it is, “We build AI infrastructure.” And in the guise of, “We build AI infrastructure,” robots, humanoids around the world are AI infrastructure. Or data centers and energy as AI infrastructure,

Guido

So the humanoids would totally work, right? If you are suddenly pouring concrete and building power plants, it has completely different cultures, completely different set of people, and getting much, much harder.

Dylan

Agree, agree. But there's different ways to do it, like invest in the various companies or like, backstop the building of power plants, right?

Because no one wants to build power plants because they're 30-year underwriting things. You know, there's all these different areas where could it use capital to, you know, allow something to happen, right? Not necessarily owning it himself.

Guido

And look, looking back on my time at Intel, one of the biggest problems we had was that our customer base sucked, right?

I mean, we were selling to… Most of the chips went to the large hyperscalers, you know, which they're way too concentrated and they build their own chips. And so you can push down your prices. Honestly spending it on diversifying the cloud.

Dylan

Well the problem was in 2014, you guys should have just charged so much that your margins were 80%. What would the world have done? Nothing.

Guido

The margins were pretty good back then. That wasn't the problem. That was the primary problem.

Dylan

They were 60, 65. They weren’t 80s.

Guido

PTSD is kicking in here.

00:56:09 Amazon’s cloud crisis

Sarah

Well, wait. I think Guido's comment is actually a really good segue into something else we wanted to talk to you about, which is the hyperscalers. And one of the reasons that I love reading SemiAnalysis is you guys make these out-of-consensus calls that you're often right about.

Dylan

Only often?

Sarah

You have a Jensen hit rate. It's very high.

Dylan

Where’s my billion-dollar EV-positive bet?

Sarah

The one that caught my eye was Amazon's AI resurgence. So I wanted to talk to you a little bit about that just because, you know, I think we found it pretty interesting being on the ground, helping our portfolio companies pick who their partners are.

And so we have some micro data on this, but can you sort of walk through why they're behind?

Dylan

Yeah. So in Q1 2023, I wrote an article called “Amazon's Cloud Crisis.” And it was about all these neoclouds are gonna commoditize Amazon. It was about how Amazon's entire infrastructure was really good for the last era of computing, right?

What they do with their elastic fabric. ENA and EFA, right, their NICs, and the whole protocol and everything behind them, what they do for custom CPUs, etc, right? Like it was really good for the last era of scale-out computing and not this era of sort of scale-up AI infra, and how neoclouds working commoditizes them and how their silicon teams were focused on, you know, cost optimization.

Whereas the name of the game today is max performance per cost, right? But like that often means you just drive up performance like crazy. Even if cost doubles, you drive up performance more, triples, because then the cost per performance falls still. That's sort of the name of the game today with NVIDIA's hardware.

And it ended up being a really good call. Everyone was calling us out like, “No, you're wrong because…” And this was when Amazon was like the best stock. And Microsoft really hadn't started ticking off yet. Nor had like all these other, you know, Oracle and so on and so forth.

And since then Amazon has been the worst performing hyperscaler. And the call here is that, you know, they still have structural issues, right? They still use, elastic fabric, although that's getting better, still behind NVIDIA's networking, still behind Broadcom's/Arista’s, like type networking NICs. Their internal AI chip is okay, but the main thing is that they're now waking up and being able to actually capture business, right?

So the main call here is that, since that report, AWS has been decelerating revenue. Year-on-year, revenue has been falling consistently. And our big call is that it's actually going to start reaccelerating. And that's because of, Anthropic. it's because of all the work we do on data centers, right? Tracking every single data center, when that goes online and what's in there, the flow through on costs, right? If you know how much the chips cost, the networking costs, the power costs, you know how much, you know, generally margins are for these things, then you can sort of start estimating revenue.

So when we build all that up, it's very clear to us that they trough on, AWS revenue growth this quarter. This is the lowest AWS revenue growth will be, on a year-on-year basis for at least the next year. And it's reaccelerating to north of 20% again because of all these massive data centers they have online with, Trainium and GPUs, right? It depends on which one. It depends on which customer. The experience is not as good as, you know, say a CoreWeave or wherever, but the name of the game is capacity today. CoreWeave can only deploy so much. They only can get so much data center capacity, and they're really fast at building.

But the company with the most data center capacity in the world, then, and still today, although they may get passed up in the next two years, is Amazon. Actually they will get passed up based on what we see is Amazon. But incrementally, Amazon still has the most spare data center capacity that is going to ramp into AI revenue over the next year.

01:00:25 Building data centers

Guido

Let me ask one question. Is that the right type of data center capacity? Like for the high-density AI buildouts today, you need, you know, massively more cooling. You need to have enough water close by, you need to have enough power close by. Is it the right place or is it the wrong type of thing?

Dylan

So data center capacity, in this sense, I mean all the way from power secure to substations built to transformers, to, you can provide the power whips to the racks.

Now obviously the data center capacity will differ, right? Historically, actually Amazon has had the highest density data centers in the world. They went to like 40 kilowatt racks when everyone was still at 12. And if you've ever stepped foot inside of… Most data centers, they're like pretty cool and dry-ish. If you step inside of an Amazon data center, they feel like a swamp. It feels like where I grew up. It's like humid and hot. Because they're like optimizing every percentage. And so your point in here is that like Amazon's data centers aren't equipped for the new type of infrastructure, but when you compare them to the cost of the GPU, like having a complex cooling arrangement is fine.

You know, we made a call on Astera Labs a few months ago, a couple months ago when they were like at 90, and it's gone to 250 the month after because of what orders Amazon is placing with them. But there's certain things with Amazon's infrastructure, I won't get too much into it, but the rack infrastructure requires them using a lot more of like Astera Labs’ connectivity products.

And the same applies to cooling, right? So on the networking and cooling side. They just have to use a lot more of this stuff. But again, this stuff is inconsequential on cost compared to the GPU,

Guido

You can build, right? My question was more like, look, I may need a major river close by for cooling at this point, right?

In many areas they just can't get enough water and, you know, there's probably power in the same region.

Dylan

There's two-gigawatt-scale sites that have power all secured, wet chillers and dry chillers, all secured. Like, everything, everything's fine. It's just not as efficient, but that's fine. Like, you know, they're, they're gonna ramp the revenue. They're gonna add the revenue. Not that I necessarily think Amazon's internal models are gonna be great, or, hey, their internal chip is better than NVIDIA's or competitive with TPU, or their hardware architecture is the best.

I don't necessarily think that's the case. But they can build a lot of data centers, and they can fill them up with stuff that will be rented out. It's a pretty simple thesis.

01:03:01 Anthropic’s role in Trainium

Sarah

How, how important has Anthropic been to the co-design for Trainium?

Because I remember we had a portfolio company, this was summer 2023. They invited them to AWS, they spent, man, I think eight hours with them over the course of a week trying to figure out Trainium. Back then it was just impossible to work through. Obviously that portfolio company hasn't gone back and tried it now, but how different is it now based on what you're hearing and…

Dylan

Oh, it's still bad. It's tough to use. This is sort of the argument that every inference company offers, right? Including the AI hardware startups is because I'm only running like three different models at most, I can just hand optimize everything and write kernels for everything and even like, go down to like an assembly level, right?

Guido

How hard can it be?

Dylan

It is pretty hard. It is pretty hard. But you tend to do this for production inference anyways. Like, you aren't using cuDNN, which is NVIDIA's like library that's like super easy to generate kernels and stuff, right—or not generate kernels. But anyways, you're not using these like ease of use libraries. You know, when you're running inference, you're either, you know, using CUTLASS or stamping out your own PTX, or, you know, in some cases people are even going down to the SaaS level, right? And like when you look at, like, say an OpenAI or like, you know, an Anthropic, when they run inference on GPUs, they're doing this.

And the ecosystem is not that amazing, once you get all the way down to that level. It's not like using NVIDIA GPUs is easy now. I mean, you have an intuitive understanding of the hardware architecture because you work on it so much, and everyone has worked on it, and you can talk to other people, but at the end of the day, it's not like easy.

Whereas with Anthropic, Trainium, more TPUs… Actually the hardware architecture is a little bit more simple than a GPU. Larger, more simple cores rather than having all this functionality. You know, less general. So it's a little bit easier to code on. There's tweets from Anthropic people saying when they're doing that low level, actually they prefer working on Trainium and TPU because of the simplicity.

Sarah

Really? Interesting.

Dylan

Now, to be clear Trainium and TPU, I mean, Trainium especially, is very hard to use. Like, not for the faint of heart. It's very difficult, but you can do it if you're just running like… If I'm Anthropic, and I must only run Claude 4.1 Opus, 4 Sonnet, and screw it, I won't even run Haiku. I'll just run Haiku on GPUs or whatever. I'm just gonna run two models, and actually screw it. I'm just gonna run Opus on GPUs too and TPUs. Sonnet is the majority of my traffic anyways. I could spend the time. And how often am I changing that architecture? Every four, six months.

Like how much…

Guido

It's nothing changing that much, honestly.

Dylan

I think from 3 to 4 it definitely did change.

Guido

I mean, define architectural change. You know, at a high level, like the primitives are more or less the same across the last couple of generations.

Dylan

I don't know enough about Anthropic’s model architecture to be honest, but I think from what I've seen at other places, there have been enough changes that it takes time to program this and really get… The main thing is like, you know, if I'm Anthropic and I have, what, 7 billion ARR now or whatever, by the end of next year, north of 20 ARR, maybe even 30 is like, and my margins are 50%, 70%. That's $15 billion of Trainium that I need, right? That I can run on Sonnet. And most of that's gonna be Sonnet 4 or 5, whatever it is, right? It's gonna be one model serving most of the use cases. So I could spend the time, and it'll work on this hardware.

01:07:03 Oracle’s success

Sarah

Yeah, totally. Maybe on the topic of non-consensus calls you've made, and maybe I'll move to another cloud. In June, you guys said that Oracle is winning the AI compute market. And then in this pod we've already referenced the big jump obviously that Oracle had, I think it was the single largest gain that a company with over 500 billion in market cap has ever had.

Dylan

Was the 2023 Q1 NVIDIA not bigger? It might've been smaller, okay.

Sarah

I think it was maybe close. We’ll fact check ourselves, but you know, obviously this is the massive commitment that was announced. Can you walk us through why you made that call then? And just sort of why Oracle is poised to do so well at such a competitive space.

Dylan

Yeah, so Oracle they’re the largest balance sheet in the industry that is not dogmatic to any type of hardware. They're not dogmatic to any type of networking. They will deploy ethernet with Arista. They'll deploy ethernet through their own white boxes.

They'll deploy NVIDIA networking, InfiniBand, or Spectrum-X. And they have really good network engineers. They have really great software across the board, right, again, like ClusterMAX, they were ClusterMAX Gold because their software is great. There's a couple things that they needed to add that would take them higher, and they're adding those to Platinum, right?

Which was where CoreWeave was. And so you couple two things, right? Like OpenAI has got insane compute demand. Microsoft is quite pansy. They don't believe OpenAI can actually pay the amount of money, right, I mentioned earlier, right? $300 billion deal. OpenAI, you don't have $300 billion, and Oracle is willing to take the bet. Now, of course, the bet is a bit like, there's a bit more security in the bet in that, Oracle really only needs to secure the data center capacity, right? So this is sort of like how we came across the bet, right, is, and we've been telling our institutional clients, especially in like a super detailed way, whether it be the hyperscalers or AI labs or semiconductor companies or investors, in our data center model because we're tracking every single data center in the world, Oracle doesn't build their own data centers either, right, by the way. They get them from other companies. They co-engineer, but they don't physically build them themselves. And so they're quite nimble in terms of like being able to assess new data centers, engineer them. So we saw all these different data centers Oracle was snatching up, in deep discussion, snatching up, signing, etc.

And so we have, you know, hey, gigawatt here, you go out there, gigawatt there, right? Abilene, you know, two gigawatts, right? You have all these different sites that they're signing up and discussions with, and we're noting them. And then we have the timeline because we're tracking the entire supply chain.

We're tracking all the permits, regulatory filings, you know, through, you know, language models, using satellite photos constantly. And then supply chain of like chillers, transformer equipment, generators, etc. We were able to make a pretty strong estimate in our data center model, quarter by quarter, how much power there is for each of these sites.

So some of these sites that we know of aren't even ramping until 2027. But we know Oracle signed it. And we have this sort of ramp path. So then it's this question of like, okay, let's say you have a megawatt for simplicity’s sake, which is a ton of power, but now it doesn't feel like much because, you know, we're in the gigawatt railroad. But you know, if you talk about a megawatt, right? You fill it up with GPUs, how much do the GPUs for a megawatt cost, right? Or actually it's even simpler to do the math, right? If I'm talking about a GB200, right? Each individual GPU is 1200 watts.

But when you talk about the CPU, the whole system, it's roughly 2000 watts. At the same time, you know, all in, everything, simplicity's sake, $50,000 per GPU, right? The GPU doesn't cost them. There's all the peripheries, right? So $50,000 CapEx for 2000 watts. So $25,000 for 1000 watts. And then what's the rental price for a GPU? If you're on a really long-term deal, volume, 270, 260, in that range. Then you end up with, oh, it costs like $12 million per megawatt, to rent a megawatt. And then each chip is different. So we track each chip, what the CapEx is, what the networking is. So you know what each chip is, you can predict what chips they're putting in which data centers, when those data centers go online, how many megawatts by quarter.

And then you end up with, oh, well, Stargate goes online in this time period. They're gonna start renting at this time. It's this many chips. Each Stargate site, right? And so therefore this is how much OpenAI would have to spend to rent it. And then you prick that out, and we were able to predict Oracle's revenue with pretty high certainty. And we matched pretty dead on what they announced for ‘25, ‘26, ‘27. And we were pretty close on ‘28. The surprise for us was that, you know, they announced some stuff that ‘28, ‘29, data centers that we haven't found yet, but we'll find them, of course.

And sort of like this methodology lets you see, sort of, hey, what data centers are you getting? How much power? what are they signing? How much incremental revenue that is when that comes online. And so that's sort of the basis of our Oracle bet. Obviously in the newsletter we included a lot less detail. Bit it was that thesis, right? That like, hey, they have all this capacity, they're gonna sign these deals. And in our newsletter, we talked about two main things. We talked about the OpenAI business, and then we talked about the ByteDance business. And presumably tomorrow, on Friday there's gonna be an announcement about TikTok and all this, but like the ByteDance business, you know, huge amounts of data center capacity that Oracle is also gonna lease out to ByteDance.

And so we did the same methodology there. You know, with ByteDance it's pretty certain they'll pay because they're a profitable company. With OpenAI, it's not. And so there's gotta be some like error bars as you go further out in terms of like, will OpenAI exist in ‘28, ‘29, ‘30. And will they be able to pay the $80+ billion a year that they've signed up to Oracle with?

That's the only like risk here. And if that happens, then Oracle's downside is also somewhat protected because they only sign the data center, which is a minority of the cost. The GPUs are everything. And the GPUs they purchase one to two quarters before they start renting them. So the downside risk is pretty low for them in terms of, if they don't get the deal, well they don't get the revenue, they're stuck with a bunch of assets they bought that are worthless.

Guido

Is there another angle here? I mean, OpenAI and Microsoft were the BFFs, and now they've filed divorce papers, and they just wanna diversify, and then that's pushing them away towards other providers.

Dylan

Yeah. So Microsoft was exclusive compute provider. It got reorged to right of first refusal. And then Microsoft…

Guido

Is it not the last choice or something like that?

Dylan

No, it's still right of first refusal. But it's like Microsoft…

Guido

Those two are not mutually exclusive.

Dylan

Well, if OpenAI is like, “We're gonna sign a $80 billion contract or a $300 billion contract for the next five years, do you guys want it?” And they're like, “No, what?” “Okay, cool.”

And then they go to Oracle, right? OpenAI needs someone with a balance sheet to actually be able to pay for it. And then they'll make tons of money off of OpenAI on the margins, on the compute and the infra and all these things.

But someone's gotta have a balance sheet. And OpenAI doesn't have a balance sheet. Oracle does, although, given the scale of what they signed… We had also had another source of information, which was that they were talking to debt markets, right? Because Oracle actually just needs to raise debt to pay for this many GPUs over time.

Now they won't do it like immediately, like they can pay for everything this year and next year from their own cash. But like in ‘27, ‘28, ‘29, they'll start to have to use debt to pay for these GPUs, which is what, you know, CoreWeave has done. And many of the neoclouds, most of it's debt-financed. Even Meta went and got debt for their Louisiana mega data center.

Just because it's cheaper than… It's literally better on a financial basis to do buybacks with your cash and get debt because the debt is cheaper than the return on your stock. Like, it's like a financial engineering thing, but like… You know who's out there, right? It could be Amazon, it could be Google. It could be Microsoft.

Guido

It's a very short list.

Dylan

Or it could be Oracle or Meta, right? Meta is obviously not. Microsoft has chickened out. Amazon, Google, and Oracle. That's all that's left.

Guido

Google would be an awkward fit.

Dylan

Yeah, Google would be an awkward fit. Amazon would be a fine fit, but you know… Exactly,

01:16:01 Datacenter buildouts

Sarah

Well I guess maybe, you know, on the topic of these giant data center buildouts, you guys just released a piece on xAI and Colossus 2. Are you getting less impressed by these feats of building something this massive in six months or is it still very impressive to you guys?

Dylan

You know, this is the thing I've said about AI researchers was that they're like the first class of humans to think about things on an order of magnitude scale. Whereas people have always thought about things in terms of percentage growth, like ever since industrialization. And before that it was just like absolute numbers.

You know, sort of like humanity is evolving in terms of how we think because things are changing faster.

Guido

Everything is log scale.

Dylan

And so, it was like really impressive when GPT-2 was trained on so many chips. And then GPT-4 was trained on 20k H100s. It's like “Holy crap.”

And then it was like, 100k GPU clusters, right? And we did some reports around 100k GPU clusters. But now there's like 10 100K GPU clusters in the world. It's like, okay, this is kind of boring, but it's like a 100K GPUs is over a hundred megawatts now.

Literally, we, in our Slack and some of these channels like, “Oh, we found another 200 megawatt data center.” There's someone who puts the yawning emoji every time, and I'm like, “Dude, what?” Like, now it's only exciting if you do gigawatt-scale data centers.

And I’m sure—I'm not sure—maybe we'll start yawning to that too, but the log scale of this is like, the capital numbers are crazy. It was crazy enough that OpenAI did like a $100 million trading run.

Then they did a $1 billion training run. Now we're talking about $10 billion training runs. It's crazy that we think in log scale, but yes. Things are only impressive when they do it. Like what Elon's doing. So what Elon's doing in Tennessee, in Memphis, the first time was crazy. 100K GPUs in six months. He bought a factory in like February of ‘24 and had models training within six months. And he did liquid cooling, you know, first large scale data center at this scale for AI doing liquid cooling, like all these sorts of crazy firsts.

Putting generators outside, like Cat turbines, all these different things to get the power, you know, mobile substations, all these different crazy things. Tapping the natural gas line that's like running alongside the factory. So he does this, it's like “Holy crap.”

And he did it for 100K GPUs. You know, 200, 300 megawatts. Now he's doing it for a gigawatt scale, and he’s doing it just as fast. You would think like, this is obviously way more impressive that he did it again. Maybe I'm desensitized, but like, you've given the child too much candy.

Yeah, right. Exactly. And now like the child doesn't like apples, right? So, like yeah, a gigawatt data center. There was all these protests around his Memphis facility, people like, “Oh, you're destroying the air.” And it's like, have you looked around that area of Memphis?

Like there is a gigawatt gas turbine plant that's just powering generally that area. There's a sewage plant that's servicing the entire city of Memphis. And there's like open air pits of like… There's open air mining. Like there's all sorts of disgusting shit around there, which is needed, right?

We need that stuff to have a country run, like to be clear. It's like people were complaining about like a couple hundred megawatts of generation. So he got protests from all sorts of people. You know, you got super into the politics side of things. NAACP even protested him. He really got like some local municipalities to be like, “Oh, I don't like, you know, like this.” And so he couldn't do as much as he wanted to in Memphis. But he still needed the data center to be close because he wanted to connect these data centers.

Super high-bandwidth, super close. And he obviously already had a lot of infrastructure set up there. So he bought another distribution center at this time. And it's still in Memphis, but the cool thing about Memphis is it's right across the border from Mississippi. It's like 10 miles away from his original one, but his facility is like a mile away from Mississippi, and he bought a power plant in Mississippi, and he's putting turbines there because the regulation is completely different. And if the question is really like, galvanize resources, and build it really fast. Maybe, Elon is ahead of everyone. You know, he hasn't made the best model yet, or he doesn't have the best model, at least today, I think. You know, you could argue Grok4 was the best for a little period of time, but it’s truly amazing how fast he's able to build these things.

And for first principles, it's like most people are like, “Fuck. We can't build the power. We can't do power here anymore. I guess we have to find a new site.” And it’s like no, no, just go across the border

Sarah

Go to Mississippi.

Dylan

And my favorite thing is like, Arkansas is right there. So Mississippi gets mad.

Guido

Future data centers built in places where multiple states meet. Is that the…

Sarah

Four quarters, yeah.

Dylan

Is there a point in the US with five? I know there's a point with four states intersecting.

01:22:03 Hardware recommendations for startups

Sarah

I'm gonna buy real estate in that area. Front run it. Well I guess on the topic of just maybe new hardware, you had this piece analyzing TCO for GB200s. And I'm kinda gonna ask this question on behalf of our portfolio companies, which it sounds like you're helping them already.

But one of the findings that I thought was really interesting was TCO was sort of 1.6x H100s for GB200s. And so obviously, you know, there's this point on, okay, that's sort of the benchmark for the performance boosts that you're gonna need to at least make the sort of performance-cost ratio benefit, from switching over.

Maybe just talk about what you've seen, from a performance standpoint and what do you recommend to portfolio companies, maybe in a smaller scale than xAI who are, you know, thinking about new hardware, try to get it. There's capacity constraints, obviously.

Dylan

Yeah. I mean, that's the challenge, right? Is with each generation of GPU it gets so much faster, that you end up, like, you want the new one. And you know, in some metrics you could say GB200 is three times faster than, or two times faster than the prior generation. Other metrics, you can say it's way more than that. So if you're doing pre-training versus inference, right?

Guido

They can run everything at 4-bit, right?

Dylan

Yeah. If you can run it at 4-bit or just inference and take advantage of the huge NVLink, NVL 72, you know, there's ways you could squint and say, GB200 is only 2x faster than H100, in which case, 1.6x TCL. It's worthwhile, right? It's worth going to the next gen.

Sarah

But more marginal.

Dylan

It's more marginal. It's not a big deal. Then there's other cases where it's like, well on, if you're running DeepSeek inference, the performance difference per GPU is north of like 6, 7x, and it continues to optimize, you know, for DeepSeek inference. Then it’s like, well, I'm only paying 60% more for 6x. It's a 4x or 3x performance per dollar gain, like absolutely. And if you're running inference of DeepSeek, that can also include RL. And then the other question is like, well, the GPU is new, you know, there's also B200.

There's GB200, there's B200. B200 is much more simple from a hardware perspective. It's just eight GPUs in a box. So then it's not as much of a performance gain, especially in inference. But you have all the stability, right? It's an eight GPU box. It's not gonna be unreliable. The GB200s are still having some reliability challenges.

Those are being worked through. It's getting better and better by the day. But it's still a challenge. When you have an H200 box, 8 GPUs, one of them fails. You take the entire server offline, you have to fix it. So usually if your cloud is good, they'll swap it in.

But if it's GB200, what do you now do with 72 GPUs if one fails to break the whole thing and you get a new 72? The blast radius of a failure. Note GPU failure rates at best are the same and likely worse gen on gen because thing, everything is getting hotter, faster, etc.

So at best, the failure rates are the same. Even if you model the failure rates as the exact same because you go from 1 out of eight to 1 out of 72, it's a huge problem. So now what a lot of people are doing is they run a high priority workload on 64 of them, and then the other eight, you run low priority workloads, which is then like, okay, there's this whole like infrastructure challenge. I have to have high priority workloads, I have to have low priority workloads. When a high priority workload has a failure, instead of taking the whole rack offline, you just take some of the GPUs from the low priority one, put it in the high priority one, then like you just let the dead GPUs sit there until you service the rack at a later date.

And it's like there's all these complicated infrastructure things that make it so, “Oh, wait, actually that 3x or 2x performance increase in pre-training is lower because the downtime is higher.”/“I'm not using all the GPUs always.”/“I'm not smart enough” or “I don't have the infra to have low-priority and high-priority workloads.” It's not impossible. The labs are doing it, right. It’s just…

Guido

I mean, if I'm running a cloud, it's actually really hard, right? Because I probably have to rent the spot one, the spares or the spot instance or something?

Dylan

No, no. Because it's a coherent domain. It's NVLink, you don't want anyone touching that. So it has to be the end customer.

Guido

It doesn't have to leave them with the empty spares. So it’s even worse.

Dylan

No, the end customer usually will just be like, “I want them and I will…” And the SLAs and the pricing, everything is like accounting for that, right?

So like, generally when you have a cloud, you have an SLA, right? That is, hey, uptime is gonna be 99% for this period. With GB200, it's 99% for 64 GPUs, not 72, and then it's like 95% for 72. Now it differs across every account. Every account is a different SLA.

But like, they've adjusted for this because they're like, “Look, this hardware is just finicky. Do you still want it? We will credit you in that 64 of them will always work, not 72.” And so like, there's this whole like finicky nature, and the end customer has to be capable of dealing with the unreliability.

And it's like, and the end customer can just continue to use B200. Performance gains not as much. The whole reason you want this 72 domain is so you can have, you know, some of these gains. But you have to be smart enough to be able to do it. And that's challenging for small companies.

01:27:36 Understanding prefill and CPX

Guido

NVIDIA has announced the Rubin prefill cards like CPX. What's your take on that? Does it cannibalize?

Dylan

Dude, and by the way, I don't know if this is like brainrot or, I don't know, but I can't remember what I had for lunch yesterday, but I know the model number of every fucking chip like…

Sarah

Haunts you in your dreams,

Dylan

We're broken. We're broken.

Guido

Living the dream. Why do you pre-announce a product that's 5x faster for certain use cases? Is it that much?

Dylan

Historically, AI chips were AI chips, right? And then we started getting a lot of people saying, “This is a training chip. This is an inference chip.” Actually training and inference are switching so fast in terms what they require that like, now it's still one chip. Actually, there are still workload-level dynamics that differ, but the main workload is inference even in training.

It's because of RL. Most of that is, you know, generating stuff in an environment and trying to, you know, achieve a reward, right? So it's inference still. Right? Training is now becoming mostly dominated by inference as well, but inference has like two main operations, right? There is calculating the KV cache, for prefill, right?

Here's all these documents. Do the attention between all of them, between all the tokens, whatever type of attention you use. And then there's decode, which is autoregressively generate each token. These are very, very different workloads. And so initially the ideas or infrastructure techniques, the ML systems techniques were, oh, okay, I will just make the batch size every single, you know, forward pass this big. Let's call it, I'll make it 1000 big. And maybe I'll run 32 users concurrently. That way, you know, now I still have, you know, 900-something left, 960 left, right? That 960 is actually doing the prefill for, you know, if a request comes in, it chunks it, it's called chunk prefill, you prefill chunks of it.

Now, you get really good utilization on GPUs. But then that, that ends up like impacting the decode workers, right? The people who are autoregressively generating each token end up having slower TPS, and tokens per second is really important for user experience and all these other things, right?

So then the idea is like, okay, these two workloads are so different, and they are literally different, right? You prefill and then you decode. It's not like you're interleaving them. So why don't we split them entirely? And this was done on the same type of chip, right? OpenAI, Anthropic, Google.

Guido

Pretty much everybody does that.

Dylan

Everyone good. Together, Fireworks. All these guys do prefill, decode, disaggregated prefill decode. So they run prefill on a set of GPUs, decode on a certain set of GPUs. Why is this beneficial? Because you can auto scale them, right?

You can, hey, all of a sudden I have a lot more long context workers. I allocate more resources to prefill. All of a sudden I have… not all of a sudden, but like, you know, over time, my traffic mix is not long-input short-output, it's short-input, long-output. I have more decode workers.

And so now I can autoscale the resources differently, and I can also guarantee that my prefill time is… What's really important in search is how fast you get the page to start loading. Not “When does the resource happen?” What do people do in games?

Like the loading screen often has some sort of interactive environment, or it blends in over time, or whatever it is. It has tips and tricks, ways to distract you. The same thing is, there's like studies and papers out there that users prefer a faster time to first token, first token gets streamed to me sooner, even if the total time to get all my tokens is a little bit longer.

Guido

They can't read that fast anyways, right?

Dylan

I mean, I like to skim.

Guido

I mean, most models return about speedreading speed.

Dylan

But you need that, right?

The idea is that you want to guarantee time to first token is a certain level for user experience reasons, otherwise people are like, screw this, not using AI. The decode speed matters a lot too, but not as much as time to first token. And so by having separate prefill decode you, you do this. And this is all in the same infrastructure, you've already done this. So now it's like what's the next logical step? These workloads are so different. Decode, you have to load all the parameters in and the KV caches to generate a single token. You batch a couple users together, but very quickly you run out of memory capacity or memory bandwidth because everyone's KV cache is different. The attention of all the tokens, right. Whereas on prefill, I could even just serve like one or two users at a time because if they send me a 64,000 context request, that is a lot of flops, right? 64,000 context request. I'll use LLaMA 70B because it's simple to do math on, like 70 billion parameters.

That's 140 giga flops per token, times 64,000, that's many, many teraflops. You can use the entire GPU for like a second, right? Like potentially, depending on the GPU, to just do the prefill. And that's just one forward pass. So I don't necessarily care about, you know, loading all the parameters in KV cache in fast.

All I care about is all the flops. And so that leads us to sort of like, you know, I had to, I give this long-winded explanation because it's hard for people to understand what CPX is. I've had a lot of, like, even my own clients, like, we send like multiple notes, like explaining and they're like, “I still don't understand.”

I'm like, “Shit, okay.”

Guido

Send the “Attention is All You Need” paper.

Dylan

I mean think about like a networking person. They're like, “I don't need to know about this ‘Attention is All You Need,’” Or think about an investor, right? Data center operator, like, they're like, “Oh, there's two chips. Why should I build my data center differently?” I gotta explain everything. Or just like, no, you don't have to build differently.

Guido

In Stanford, there’s 25% of all students, not CS students, of all students read that paper. “Attention is All You Need.”

Dylan

That's low.

Guido

The literature majors, and like the philosophy guys, I think that’s amazing.

Dylan

The Middle East, I can't remember what country it is, has AI education starting at like age, like eight, and in high school they have to read “Attention is All You Need.”

Someone told me that their kid had to read “Attention is All You Need.” Look, top-down mandates for education, you know, maybe they work, maybe they don't like, maybe people like homeschooling their kids. I don't know. I went to public school, but like, back to your readers.

Sorry, I didn’t actually explain what CPX is. So CPX is a very like, compute-optimized chip, for prefill and then, and then decode is, just to succinctly say it, is like the rest is the normal chips with HBM. HBM is more than half the cost of the GPU.

If you strip that out, you end up having a much cheaper chip passed on to the customer. Or like, you know, if NVIDIA takes the same margin, then the cost of this prefilled chip is much, much lower, and now the whole process is way cheaper, more efficient. Now long context can be adopted.

01:34:49 The state of GPU purchasing

Sarah

Yeah. I love that we're actually going into all this detail because I had a more 10,000-foot view question for you, which is, I haven't been following the semi market as closely as you have. I probably started with the A100, and I remember helping Noam at Character, this is summer of June, 2023, chase down GPUs, and the only thing that mattered at that time was delivery date because there was a huge capacity crunch. And then to see that over the last two years evolve, where, you know, let's say 6 to 12 months ago people were doing these RFPs to 20 neoclouds, right? And the only thing that mattered to some degree was price.

Dylan

Do people actually do RFPs for GPUs?

Sarah

Yes.

Dylan

So just to be clear, my opinion on how you buy GPUs is that it's like buying cocaine, or any other drug. This is described to me. Not me. I don't buy cocaine. Someone tells me this, someone tells me this, I'm like, “Holy shit. That’s right.” You call up a couple people, you text a couple people, you ask, you know, “How much you got? What's the price?” This is like fucking like buying drugs. Sorry, sorry.

Sarah

No, I mean, like accurate.

Dylan

To this day, it's the same way. You just send, like we have Slack connects with like 30 neoclouds, as well as like some of the major ones. And we just send them a message like, “Hey, customer wants this much, you know, this is what they're looking for.” And then they send quotes.

Guido

I know this guy.

Sarah

I know a guy. Well, so I think that's actually a very accurate description. And I've sent countless portcos your ClusterMAX original post because I thought it did a really good job breaking them down. But maybe one question to end on for me is just what era are we in now with Blackwells coming online?

Are we sort of back to the summer 2023 era, and that's kind of the cycle that we've just entered? Or what, what's sort of your view on where we are? So

Guido

That’s a very good question.

Dylan

For one of your portcos, we were like, you know, after their difficulties with Amazon, we were like, okay, let's actually like get you GPUs, the original deals we got you were gone, but like, here's some other deals. It turned out that multiple major neoclouds had sold out of Hopper capacity. And, their Blackwell capacity comes online in a few months. So it's a bit of a challenge, right?

Sarah

Due to inference?

Dylan

Inference demand has been skyrocketing this year, right?

Guido

Reasoning models, yeah.

Dylan

These reasoning models, the revenue. It's been skyrocketing this year. And then also, like, there's a bit of like the, you know, Blackwell comes online, but it's hard to deploy. There's a learning curve to deploying it. So whereas like you got down to, like, you buy the Hopper, you install the data center, it's running within like, you know, a month or two, right?

For Blackwell it was like, it's a longer timeframe because of reliability challenges. It's a new GPU, I mean, it's just growing pains. So there was this gap of like how many GPUs are coming onto the market as revenue is starting to inflect.

And so a lot of capacity got sucked up. And actually prices for Hopper bottomed like three or four months ago, or like five or six months ago. Yeah. And actually they've crept up a little bit now. I don't think we're quite 2023, 2024 era of, GPUs are tight. If you want like just a few GPUs, it's easy. But if you want a lot, it’s hard. You can't get capacity that instantly.

Sarah

Wow. What a time.

Erik

Shall we wrap on that? Dylan, this was another instant classic. Thank you so much for coming to the podcast.

Dylan

It was like two hours, bro, like what the hell.

Guido

We couldn’t stop.

Erik

Thanks so much. This was great.

Dylan

Thank you so much for having me.

Resources:

Find Dylan on X: https://x.com/dylan522p

Find Sarah on X: https://x.com/sarahdingwang

Find Guido on X: https://x.com/appenz

Learn more about SemiAnalysis: https://semianalysis.com/dylan-patel/

Stay Updated:

If you enjoyed this episode, be sure to like, subscribe, and share with your friends!

Find a16z on X: https://x.com/a16z

Find a16z on LinkedIn: https://www.linkedin.com/company/a16z

Listen to the a16z Podcast on Apple Podcasts:

Listen to the a16z Podcast on Spotify:

Follow our host: https://x.com/eriktorenberg

Please note that the content here is for informational purposes only; should NOT be taken as legal, business, tax, or investment advice or be used to evaluate any investment or security; and is not directed at any investors or potential investors in any a16z fund. a16z and its affiliates may maintain investments in the companies discussed. For more details, please see a16z.com/disclosures.

Timecodes:

00:37 Why Alex and Justine are thinking about AI’s role in commerce

06:08 How much will AI result in dynamic pricing?

07:09 Why is e-commerce so small?

12:28 Why have e-commerce brands struggled?

17:46 Does AI threaten Google’s business model?

29:30 Costco’s business model

33:18 Which categories of shopping will AI eat?

39:52 What net-new commerce companies will be created post-AI?

Transcript

This transcript had been edited lightly for readability.

00:00:37 Why Alex and Justine are thinking about AI’s role in commerce

Erik Torenberg: So you guys have both been thinking about AI and commerce for a while. Alex, why don't you talk about what inspired this piece and how did this all germinate for you?

Alex Rampell: Well, so I had started a company called TrialPay a long time ago. And actually I'd been selling stuff on the internet for a very, very long time, even before the internet.

And I was just trying to think, well, number one, like, what happens to Google? Because a lot of people, this is on their minds, so like, is search volume going up or down? So I have my own personal experience of like, well, my search volume is going down, but not for commerce, but clearly for everything that is not commerce.

So that was one thing, but also, this company that I started, TrialPay, we were one of the biggest affiliates in the world. And affiliate marketing is basically you send somebody, this is the oldest business model, on the internet, if you will. It actually predates AdWords and AdSense by a bit, of, you just get a share of the, you get a commission basically if you send something there.

This was started apparently apocryphally, it came from pornography, because that was the world's oldest business model on the internet. Like how do you track, how do you, how do you… So that eventually made its way to commerce, and it's all based on cookies and pixels. So you drop a cookie on the person's computer, and then on the confirmation page you have a little, like invisible one by one tracking pixel that reads the cookie, and that's how you know to say, “Erik sent me the customer.” So what we did at TrialPay, we were one of the biggest ones there, like, is that really gonna be what powers this new realm of commerce? And then is it even relevant for a lot of things because impulse buys are huge. And with impulse buys, like almost tautologically, you're not going to use AI to tell you to buy something.

Like you shouldn't buy anything that's an impulse buy. Like you go to the supermarket, you shouldn't be buying like Coca-Cola, like in the checkout line. They actually charge you more at the checkout line than they do if you just buy it in the Coca-Cola section. So like all of these things are designed to like tug at your emotions to get you to buy and spend money that you don't want to.

That's not gonna be AI. On the other hand, it's like these very, very expensive items. You're researching the heck outta them with AI. But like there's no affiliate model. Like how do you then commerce and transact? So like number one was like the ontology of commerce was very interesting. And then number two was this whole like affiliate thing, is it still gonna be relevant? Because it seems like that's what ChatGPT and others are getting into. And then number three, it was just my own personal behavior of it's just like I probably use ChatGPT like three orders of magnitude more than I use Google now, which is interesting.

Erik: Justine, what excited you to contribute to this piece slash what did you find most remarkable or?

Justine Moore: There's a couple really massive consumer markets, the biggest of which might be like online shopping. But I think we've seen thus far relatively few startups trying to take a crack at that market with AI, even though, like Alex said, there's a lot more opportunity because you now have these really smart LLMs and agents that can help you make better decisions than you could have made on your own or even make purchases on your behalf.

Which you would think would create an opportunity for more folks to package these into products that they then offer to consumers, but we haven't seen a bunch of folks doing that yet. So I think part of this piece was to dive into like, why is this system so complex? What are the different types of purchases where AI can play a role. And what are we hoping to see as we think about the broader market in hopes that, you know, folks who are also, who are working in this space would kind of give us a heads up and let us know and we could hear how people were approaching it.

Alex: And you can observe, I always like to observe first because like, that's objective and then predict second. Predicting is hard. I forgot the funny quote about predicting the future, but it's very hard to predict the future. And like there actually is a lot that can be observed. And that's where like, I think CamelCamelCamel is like the greatest site in the world. We have no stake in CamelCamelCamel.

So this is not self-promotional at all, but it's like, people use this thing, it's like Google News alerts for pricing. And I gave a talk to the Amazon Prime team recently, and like they're very, very aware of it because I think it's actually Amazon's biggest affiliate, and people every day, like, I would buy this product…

I mean, this is like Econ 101. I would buy this product if it was priced here. It is currently priced here. Please let me know when it's priced here because what will I do with that information? I'm not just gonna, like… I'm gonna buy it, so like the consumer is the agent, and this is like a very, very inefficient AI.

And if you could actually complete the entire circle and say like, no longer give information, but allow for automatic action on information. Like people will do that because we have observed that behavior today. This is like the easiest form of predicting the future ever, because you're really just like, you're chronically in the present, and you're just saying there's gonna be one additional appendix to the present, which people would do anyway because they are doing it anyway. They just have an easier way to do it.

Justine: I think my version of observing was seeing, there's a couple viral examples of this, some that were really good because the AI found the product perfectly and some that were hilariously bad because AI couldn't find the product. But teenage girls started using ChatGPT to upload photos of like Lana Del Rey at a concert or Taylor Swift like snapped in a street style photo or whatever and asking like, what is this hair barrette she's wearing? Or like, what is this sweater? Like, I wanna find it and I wanna buy it. And it, when it worked, it worked really well because it often found like, “Hey, this sweater is like $5,000. Like you as a 19-year-old girl in Missouri are probably not gonna be buying this. Like here are some alternative, maybe less expensive options that look the same that you can buy.” And that age demographic tends to be like a really early predictor of all sorts of consumer behavior, which is why I was kind of like, you know, this is probably gonna be happening more and more from the research side of things all the way to making purchases probably agentically when prices are right, like Alex was mentioning.

00:06:08 How much will AI result in dynamic pricing?

Erik: Alex, do you imagine a world where there's sort of dynamic custom pricing sort of to the extreme where it's like we're looking at the same thing on Amazon, but it charges you more because maybe I'm cheaper than you or you have more money. Do you imagine that world?

Alex: Well, I mean, people have tried this a lot. It's a very smart world from an Econ 101 perspective for sure. Like how do you capture the consumer surplus? Consumer surplus is great for consumer, it's bad for producer. And apparently Delta is doing this a little bit, or they were trying to do this.

There are like the poor man's versions of this, which are like, if you have an iPhone, you should get charged more than if you have an Android phone because like iPhones are more expensive. Like you have basically communicated that your elasticity of demand is different than somebody who, you know, has less money.

I think probably you're gonna run into regulatory challenges with that. Or certainly you will run into like very, very high levels of unpopularity with your customer base, but people have tried this. But generally it's hard to get away with. Right.

00:07:09 Why is e-commerce so small?

Erik: Right.

Let's reflect on previous platform shifts before getting into this one. Right now e-commerce is 16% of total retail sales. If we were, you know, talking 20 years ago and predicting, you know, what percentage of commerce would be e-commerce, we would probably think it's much higher. Why hasn't that been the case?

Alex: It turns out the demand curve is different for immediacy versus non-immediacy.

So even though like overnight shopping is pretty darn cool, like instantaneous, like one-second shopping is like, “I need toothpaste right now because I'm going to bed and I wanna brush my teeth. And I just ran out. Oh, there's a Walgreens over there. I'm gonna go there and buy toothpaste and like Amazon is awesome, but getting the toothpaste at 7:00 AM. It's like I don't have demand for that.” Like that's not part of the demand curve. There's demand curve for like real-time toothpaste. That's part. The other part is just like, I'm bored, like what do I do today? I know, I will go to the shopping mall. And like there's the experience of doing that, and it's like that's kind of a little bit more impulse, but even it's like, ah, there are these long-term considered purchases or we talked about in RPs, aspirational purchases.

Like, maybe I'm gonna gawk at that Rolex a little bit more and ooh, I just got my bonus, maybe I'll go buy it. But it's all part of the experience. So I think those are kind of broadly speaking, the two… I mean, I've seen this, I'm on the board of a company called Wise, and the product is sending money, and it turns out that the market for sending money where it is received in real time, it's just much, much different than the market for sending money where it is received two days later.

Because sometimes it's just like the two days later send thing is just like, less demand for that. And again, Amazon has proven this as shipping has become, it's like once upon a time, like when you go back to the early days of e-commerce, you would get something in like two weeks. So it's almost not surprising that the curve has just like continued to expand. It's almost surprising that it's only, you said 16%. Yeah, that seems very low.

Justine: I think it's higher. So I'm not doubting your research on the numbers. Here's why I think it's higher. I think there's a lot of behaviors where people do research online and then purchase in person. Like especially for big sorts of purchases or even like sometimes I'm like, “Hey, I need a new laptop. I'm gonna do like all of the research on like Reddit or on Instagram or on the Apple website. But then I'm gonna like go into the store and feel like, okay, what is actually the difference in like the Pro weight versus like the MacBook Air weight,” right? And so I think there's a lot of those sorts of things where like… Or buying clothes is another great example where like, I live in SF so a lot of people will just order a ton of clothes, try it all on, and then send a ton of it back because there's not a lot of big stores near us.

But I grew up in Oregon and there like, it doesn't make a ton of sense to order a ton of clothes online and send them back. It's just inefficient when there's so many clothing stores that are like a five to ten-minute drive from you. But, a lot of people will kind of do research about where to go, or what specific items they're looking to find, or like what style they're looking to buy online.

So I think, yeah, it might be like 16% or like fully transacted online. But I think, even in a lot of those other purchases, there is some sort of online research component.

Alex: Well, this is actually the hardest thing, that's related to this topic is attribution. It's, the bane of everybody's existence, which is like, okay, how do I allocate attribution for Justine's MacBook sale?

And like the most kind of pervasively corrosive business model, I think on the internet is this like last click attribution. So you allocate 100%. It's like, okay, part of it was like, I read this post on Reddit that kind of inspired me. Part of it was I saw this really cool ad at the Super Bowl.

You could do this like kind of piecemeal, which is probably the more accurate way of doing it, but it's not exactly deterministic. And the thing that feels deterministic, which is actually incorrect, is it's like, oh, whoever sent me the click last is the one that I should reward with the spoils.

And a lot of people that just don't understand correlation versus causation fall into this trap where, and this is the business model that I hate the most in the entire world, like the things like Honey. You know Honey, right? So what is that doing? It's like you're already on the webpage about to purchase.

And then it's like, do you want a coupon code? Oh, why, yes I do. Why would I not want a coupon code? 10% off. Click here. You go click here. What does it do? It redirects you to an affiliate page. It puts a cookie on your machine. It redirects you back to the page that you were just on, and then it actually steals that attribution.

And what's funny is if you talk to a lot of the marketing people at these larger e-commerce companies, and Amazon is very smart. That's why they don't do any of this stuff. They're like, “Oh, our best channel by far is Honey. Like they're growing so much.” Or RetailMeNot. That was the original one.

It went public, you know, big valuation, and it was just theft. But it's just because, again, like how do I figure out how to do attribution? And this is only gonna get more complicated in the AI world, where like it might be the same thing where it's like Justine might have researched on Reddit, saw the Super Bowl, like did all of these things, asked a question on ChatGPT, and then clicked purchase. And it actually is incorrect for Apple to say, ChatGPT, we owe you the entire, that drove the purchase. No, it didn't. Yeah, it's part of it, but it didn't drive the purchase. And figuring out and disentangling attribution is very, very hard.

00:12:28 Why have e-commerce brands struggled?

Erik: Let’s reflect back on the category as well. It seems like the big winners have been, at the aggregator level, sort of Shopify or obviously Amazon and sort of the individual, you know, big brands like, I don't know, Allbirds or Casper. It seems like they were quick to get a lot of revenue, but didn't become, you know, durable businesses in the same way. Didn't get, you know, didn't get better as they scaled. Why don’t you reflect Alex a bit on the category in general and why it's played out that way?

Alex: Well, I mean, ultimately if it's a one and done transaction, you don't really make the product. Like, you know, Casper didn't make the mattress right? Like there's probably some OEM in China that made the mattress, and they put their little logo and they called it Casper on it. Well then they're just buying traffic on Google and Facebook. So actually Google and Facebook were the real victories there, more so than anybody else. And then people are like, “Wow, mattresses, that's a really good category. I should do that. Oh, I'm gonna go to Shenzhen or wherever. Like, I'm gonna slap my logo on it. I'm going to undercut them on price.” And that's what always happens. And it's one thing of, you can ameliorate this to a certain extent, if at least you have recurring billing where it's like, I'm gonna have this problem.

Like, you know, think about what Dropcam did, if you remember that. That was an e-commerce product, but at least it was attached to a subscription. So now there are like 9 billion cameras that all do the exact same thing. So like that category has arguably gotten worse even as the category has expanded or like the demand has expanded.

But at least like, you know, Google owns Nest, which bought Dropcam, like they probably still make a lot of money on that category. Whereas Casper, if I bought a Casper mattress five years ago, I'm still sleeping on it. And like they have to find new people to sell that mattress to. And meanwhile, the original factory that was making the mattresses is now selling the exact same mattress to like 5,000 other manufacturers. And it's just like not a good business model. So in general, just being a commodity reseller of products… And like, I think this is the problem, it's like a lot of people will say, “Oh, well, like Casper is its own mattress. Allbirds is its own shoe,” but like they're not normally making these products. Like there's somebody else that's making the product.

Because it's almost obvious, like what happened during internet 1.0? The long tail of commodity resellers went away because location no longer mattered. Because a lot of what really drove retail for the longest time was that like in Justine's town in Oregon, like there's this store, and like you could drive like somewhere else, but that's really far away.

So of course you're gonna go to this store. Now the internet, you can go to any store. So if there are like 5,000 stores that don't make their own products and they all sell the exact same shoe from Nike. That doesn't make sense. You should either go to Nike directly. Or you should go to like the one store that has fastest shipping, the best service, whatever.

And like the long tail of commodity retailers basically started dying. And like we saw this play out, but the first party commerce experience is not that much better either. There's no barrier to entry, and if there's no barrier to entry, then that normally doesn't, it works out great for the consumer in capitalism, it doesn't work out great for that, you know, one of n, where n is quite large, producers, or non-producers, but just marketers of the product.

Justine: I think there's also, especially with like true consumer products, and I would consider a mattress maybe more of a utility product, but with, you know, shoes like Allbirds or makeup, it's very trend-based, especially with the internet.

Like nothing stays that hot for that long. Like, you know, Allbirds is the big shoe one year, and then the next year it's the retro Adidas that everyone's going back to. And now it's like the On running shoes. Like I was watching the Bama Rush Sorority TikToks this year, and like every single girl has the On running shoes, whereas last year they all had the New Balance, like, cool look from Japan.

And if you are Allbirds, like that's a problem, right? Because you can't capture all of the trends. Like you have your kind of one SKU or multiple SKUs across one style, whereas like the Shopifys and the Amazons can kind of ride whatever the trend is and have demand come to those individual SKUs, which I think is gonna be an interesting challenge in the age of AI too because you could argue that AI agents can direct people to things if people start their purchase activities there. Which could be an opportunity or a challenge for like the single SKU retailers. My guess is it'll still end up being more of a positive for like the aggregators.

Alex: Also, I think it's gonna be very hard for AI to, for lack of a better term, inculcate demand.

Which is like, how do I know that like the On shoe is cool. It's like, well, I need to see that Bama…

Justine: I'll send you some.

Alex: No, I mean, it's a metaphorical I, right?

Justine: Yes, yeah.

Alex: It's like once I see, “Oh, wow, I should have that too,” right? “I'm in a sorority, I want that shoe.” And it's very hard for AI to do that.

Which is why like the utility part of it's like, “Well, I know what I want. Now buy this for me.” Like that seems like a no-brainer. Because that's a lot of what Google does. Like Google. I mean, I respect the hell outta that company, but they kind of are a tax on GDP. It's like a lot of GDP happens.

A lot of that is commerce, right? Consumer spending is a huge part of GDP. Where do you start that spending journey? With that little nice little search box. And then they get a percentage of all that spend because they're charging per click or per impression or per action. So, that is somewhat imperiled. Like that tax might just shift elsewhere.

00:17:46 Does AI threaten Google’s business model?

Erik: Yeah. Let’s flesh out the piece now. Let's get into, you know, what are some of the things that are gonna be taken away from Google? What are some of the things that are gonna stay? I also want to get into the different kinds of consumer spend as it relates to e-commerce. Maybe Alex, do you want to start with you?

Alex: Well, I think, Google has been the canonical freemium business model forever, which is, they built a better search engine, everybody knows that, that started in 1998. And it was like the 47th search engine or something. “Ah, this isn't gonna work.”

But it was just so much better because of the way that they linked it. It really kind of goes actually back to like research, which is like, it's kind of like the h-index but for finding things. Like when you search for “bagel,” everybody like hyperlinks to this one site, like that must have a high PageRank.

Like let's go show that first. So, but most of the searches when Google started, because commerce was actually quite nascent on the internet at the time, like it was all free, it was all kind of information. I remember using Google when it first came out and it was like, ah, this is so much better than HotBot and all the other things out there.

It was all free, non-monetizing. They eventually basically copied the Overture business model, which this guy Bill Gross came up with. This was an IdeaLab company that eventually became part of Yahoo. This is why Yahoo ended up owning part of Google, if you know the whole history, and the entire thing that made Google this like, you know, giant $2 trillion company was AdWords.

And actually the cool thing about it is that there are a lot of freemium business models where it's like, “Ah, I don't wanna pay for it.” Here, it was freemium, but actually having relevant search results that are paid alongside search that is organic made the search better. If I'm searching for “tennis racket” and somebody hadn't figured out how to PageRank optimize and everything else, or SEO optimize, it was very useful for them to be able to show ads here.

And then those ads wouldn't show up unless people clicked on them. Because the relevance was never like preordained. It's like if people click on it, it's relevant. If people don't click on it, it's not relevant. So Google has always been freemium. And that's kind of bearing out right now, which is like, you know, it's still freemium.

Like you search for lots and lots of things with no intent to buy. But every now and then it's like, this is your default behavior. It's like, “Hmm, I wonder about X,” you go to Google. Or sometimes you don't even go to Google. You go to like, you know, Safari because like Apple makes tens of billions of dollars a year by sending all of those searches to Google.

So what is currently happening is they are starting to lose some of the “free,” but not any of the “mium,” right? They're losing some of these informational queries like who won the Oscar in 1977? Like, that's not a monetizable query, but like, that's what you're gonna want to know. You're gonna just ask ChatGPT, and people are doing this right now, and ChatGPT has, you know, I think, what is it, 800 million weekly active users? A huge, huge number. That's what they're doing with it because they're not buying in ChatGPT, we know that because. OpenAI is trying to build commerce. So like, clearly they haven't built it yet, so like they're not buying directly in there.

But for the “mium” right, the premium part of the freemium, that is happening in Google still. And like, how do I know that? Well, I can look at their financials. And like their financials, like the numbers are still going up, but we also know that search volume is actually going down. So what are they losing if they're not losing revenue?

They're only starting to potentially lose some of the free searches. What I don't know is maybe they're directing some of those to Gemini. But I think that's unlikely. I think right now what's happening is it's like people are using paid for Google, no changes at all. They're just going elsewhere with AI for free.

Justine: I think part of probably why that's been happening is, like all LLMs, but I'll use ChatGPT as the example because the most people use it, had this really unfortunate and annoying problem of hallucinating around product recommendations that basically everyone experienced if you tried to use it for that.

Like, I think you have this grand idea of like, “Okay, I wanna buy a pair of leggings, and I'm gonna go search in Google, or I'm gonna go search in Amazon and then like I'm gonna get the highest rank pages. But what I really wanna know is like I'm doing this specific type of hiking and this is what the weather is gonna be like, and I want to know specifically for my needs, like what is the best legging? It might not be the best overall legging,” right? And so a lot of people, I think especially, I saw a lot of young women trying this. We're like, “Great, I'll go to ChatGPT. It can take all my information in natural language. It can make a recommendation, it can spit out products.” And then they would find that a lot of the products it recommended did not exist or previously existed, but did not exist in a current form, or the amount that they were charging was way different than it said, which I think drove a lot of people who experimented with it sort of back to, you know, I'm gonna return my searches to Google or Amazon and wait until ChatGPT figures out this commerce thing. My take is, as we know, OpenAI is working on commerce, and they're trying to integrate it more into the experience and have actually like real, relevant, up-to-date information on products. Google will probably be at risk of losing some queries, but I totally agree with Alex that we have not really seen that behavior at any sort of scale today.

Alex: Well, the biggest problem right now for the internet writ large is, I would say, and I remember I've talked to John Lilly, he was the CEO of Firefox back in the day and kind of an early web stalwart, the internet is unhealthy or the web, the World Wide Web, is unhealthy right now. And the reason why it's unhealthy is because so many things that used to be the open internet, when it really was like, you know, DARPANET, then ARPANET, then like this, like internet thing, that people started using, but only really researchers and stuff like that.

Everything was just on the open web. There was no concept of a walled garden. I mean, like, search has already been fractured, by the way. It didn't happen with ChatGPT. It's like if you want realtime search, you go to Twitter, or X. You want search for your friends well you go to Facebook. Like none of that is… You can't search for Google in terms of like stuff that's happening amongst your friend group that's like walled off there.

So you have all of these different walled gardens, so that's unhealthy part number one. Unhealthy part number two is just the commercialization of the internet, which is not bad. I'm a capitalist. I like commercialization. But so much of like, if you look for what is the best sneaker, right, like, who are the people that are writing content best about great sneakers? Like in 1995, if you had a blog, well, number one, you just hosted it on your own site. You set up Apache on your own server that you racked yourself, and then you just did it for the love of the game. And then affiliate links provided the monetization model, but they really polluted the internet that was still open because like so much is like, oh, top 10, like a lot of these top 10 sites are out there.

It's like top 10 running shoes. You know what that is? That's top 10 affiliate revenue to me. And I pay somebody in India to go write gobbledygook and then SEO the heck out of that to make money. Contrast this with like pre-internet where there's a publication still around today called Consumer Reports, and the really cool thing about Consumer Reports is they were the only publication that refused to take advertising.

It was entirely subscription based, and the idea was that you could trust the actual reviews. And they would do things like, they were like the Ralph Nader of consumer products, where it's like, you know, this thing is terrible. Don't buy this blender, it will chop off your finger. Like, do buy this thing. Like they would really, really review everything.

We kind of need that, and like that entire business model just went away. Like, you know, Craigslist killed almost all of traditional media. I mean, maybe they deserve to die, maybe they didn't. But like they made money from two things. They made money because they had a monopoly on information. They charged for ads there. But like a big part of the monetization model was like the local classified. All of that went away, which is why newspapers have been dying. And you could imagine like a newspaper would have like a do-gooder thing where it's like, ‘Oh, let's review all the blenders and like, we're obviously not gonna like show blenders that cut off your fingers. Like that's bad.” Like that whole thing went away. Like, so the summarization of the open internet is tough because there's less open internet than there used to be, like as a percentage of all the content being generated. A lot of it is walled off. And then the stuff that is not walled is just pervaded by like junk.

And that's why, like what we talk about in the piece is like, you can't turn shilled junk into honest analysis. So I don't know how we solve that. Like no matter how good, like no more hallucination, like everything is awesome, but like most of the things on the internet are crap. And they're crap. And we know that they're crap, but they SEO optimize crap in order to earn affiliate commissions, and like summarizing that crap is not helpful. So how do you decrapify that? And that's quite challenging.

Justine: I think honestly what I've seen in terms of the channels where you see the least crap is actually video. Like if you're a creator now it's, you know, due to the death of traditional media, there's now creators who go out and review 10 different shoes for running, and will specifically make it very clear in the video, either this is sponsored by this specific brand, or the better ones obviously are completely non-sponsored, but they get ad revenue from Google, from YouTube, from people watching the video. And so honestly that, when I want an honest review of like someone has looked at five different blow dryers for this sort of hair, I will go to an unsponsored YouTube video, which often have a lot of views because there's a lot of people having similar queries.

But the sense I have is that Google is not… Like, because it's a video and it's not skimmable, and they're not like automatically generating transcripts for every video, that information does not appear in traditional search. And I think we're starting to see some companies say like, “Hey, look, like we should turn all of those high quality videos into transcripts, and an LLM can then read and review and make recommendations.”

But that doesn't seem to have hit like the traditional Google part of the internet yet.

Alex: Yep. I agree with that.

Erik: The New York Times recently bought Wirecutter which may be an example of what you’re talking about.

Alex: Well, but it… Yes and no. I mean, it's like everything is affiliate linked, right? Like is it really true? Like it's so suspicious that like almost every item that they recommend always has an affiliate link. Isn't that odd? Right, like, does that mean like, you know, is that like a sampling bias thing? Is that true? Like, so I'm quite skeptical of a lot of these things.

And again, like the Consumer Reports era was just a little bit different. I mean, you might have the biases, like maybe the person reviewing things for Consumer Reports, like just hates this one company and is taking out their bias… This is always possible, but you would think with all these algorithms and everything else, if you could get like true objective feedback, this would be fantastic.

I mean, Amazon actually is a giant search engine. And like that thing is polluted to crap as well. Because what happens is a lot of the sellers on Amazon, what they do is they go on the site called AliExpress. And AliExpress, I mean, this has changed a little bit with tariffs, but like they'll buy like 400 of some gizmo that shows up six weeks later, and they'll buy it for $2 each. They’ll slap their logo on it, now, they'll sell it for $25. That actually goes back to this latency point that I was making before, like, how many people want something six weeks from now versus how many people want something tomorrow? So a lot of what Amazon was, is they would just arbitrage that.

But if you search for an item, like particularly in consumer electronics, I remember I was looking for heated socks for skiing. Turns out they're very useful. There are like 9,000 different pairs of heated socks that all have the same OEM, the original equipment manufacturer. And it's like, and they all have like bogus reviews and part of like the bogusness, and Amazon should fix this, but they have no incentive to do so, it's like I used to sell a rock on Amazon. I get five star rock reviews. Now I switch the SKU from rock to heated socks. And I trade off my five star review. And it's like, how does Amazon… And again, Amazon just wants to sell more crap. So like they're totally fine with this, but like most things, if you're willing to wait, you're so much better off buying on AliExpress than Amazon.

And it's just like this polluted sea of crap.

00:29:30 Costco’s business model

Like my favorite business model for commerce by far is Costco. I think Costco is the greatest company in the world because Costco refuses to sell bad things. They refuse to take a high gross margin. Like why would they refuse… It doesn't make any sense. Why would they refuse to take a high gross margin? You know, why? Do you know why?

Erik: Because they want to pass it back to customers or?

Alex: No. No, it's because it degrades the value of the membership. They make money from the membership, so they'll charge you like something like $100 a year to join Costco. And if you look at their net income, it's basically the number of memberships… They have like 50 plus million members, it's some huge number, times the price of the membership, that's their net income. And then everything else just kind of is a wash. And if you are making a 50% gross margin on a shirt, they're like, “That's too much. You're fired. Like, you can't make that much money.” It devalues the membership. They'll do crazy things like, you know, the hot dog is still a dollar 50.

They started their own chicken farm because the rotisserie chicken, like the costs were going too high. It's like, that's how they run the business, and they refuse to sell anything that they are not proud of. And the generic brand is just as good. Like, you know, Kirkland wine, Kirkland beer, Kirkland shirts. They're getting sued by Lululemon right now because they made pants that were better than Lulu's pants that are much, much cheaper, but they're actually much better. So Costco is the greatest thing… Like for everything we talked about in commerce, like pre-internet, like internet, AI, like Costco's immune to all of this because it's like, they're like the Consumer Reports plus like… They treat customers incredibly well, and that's why this company is worth hundreds of billions of dollars.

Justine: And people really trust them. Like my mom has been a Costco member for forever, and now she gets her glasses at Costco. Every time I like want to go get flights or something, she's like, “Log in and like use the Costco thing.” Because she always thinks that Costco is gonna have the best option at the best price, and she's usually right.

Alex: That is sacrosanct to them. Yeah. They refuse to violate that because they could make so much more money if they decided to. And Amazon, it's interesting because like normally I remember, there's a speech by Jeff Bezos where he talks about this like, you know, there are two business models.

There's the like, what's the most that we could get away with in terms of charging, like that's Apple. It's like, oh, like, you know, let's charge $1,600 for this iPhone, you know, 25 that we're gonna come out with that has 18 cameras. Can we even get away with $1700?

They have very high gross margins. There are other companies like, how do we charge the least amount possible? That's kind of Amazon. Like, let's have this sea of like, crap, but like, you know, whatever. Like why would we curate the crap? Like that's up to the consumer. We'll have the reviews and everything else, but they don't do a great job on the reviews.

And like, those are the extremes, right? Like, you know, Android and Apple are like, you know, there are so many examples of this. There's like the premier provider, right? There's like the Mercedes, the Ferrari, whatever. And like they just wanna show like high end stuff. And then there's like the mass produced like, you know, low end stuff.

And then there's this very, very unique business model that is very hard to replicate called “Costco” because normally it doesn't work because it's like, “Hey, just trust me because I'm the best.” Well you have to have like many, many decades of trust such that Justine's mom is like, “I don't know what it is, but if it's sold at Costco, it's good.”

Erik: And if you were CEO of Costco, would you further leverage that trust to do other things? Or would that risk the whole enterprise?

Alex: I think it risks the whole enterprise, but there's a lot that they can… It's funny. One of my partners met with the Costco board and like pitched them on financial services. Because it's like, you know what? Every bank is trying to rip you off, right? They're trying to overcharge you for loans or like underpay you on your deposits. And like the Costco loan would just be like the cheapest possible. Like they're trying to make no money on that because they make money on the membership.

So they probably could expand it quite a bit. But yes, it's hard there. There's some modernization that they could probably do because it still is this very warehouse thing that like closes at 5:00 PM, and I wish it were open later, and everything like this, and like their shipping isn't great if you want to order stuff. But it is a very, very unique business model that is somewhat of a… it will stand the test of time, and I think it's AI proof.

00:33:18 Which categories of shopping will AI eat?

Erik: Yeah. Justine, why don't we get into other ways in which AI will change commerce? You outline a few different types of purchases that might get eaten.

Justine: Yeah, so we kind of looked at the range of purchases from like the impulse buys, which I think used… I mean still, they still are, like the Coke thing, the Coke bottle on the aisle.

But often now for like a lot of people, they're like the TikTok shop thing where you're watching a video, and it shows up, and you're like, “That T-shirt looks cool. I'm gonna buy it,” all the way to like really considered purchases like a house or like a wedding venue or like a car where you're spending like a significant chunk of your income.

It's like a one or a multi-time thing and you're like doing a lot of research. So both ends of the spectrum I think are harder for AI to disrupt. I think the impulse buy because like there's no research in advance, and you're not going anywhere specific to buy it. Like by its nature you are making the decision to buy it immediately when you see it.

And so, you know, like algorithms will get better and better to target you with the shirt that shows up on your TikTok feed that somehow has your dog's name, and you're gonna buy that more than the other thing. But that's sort of not the like generative AI that we're talking about.

And then sort of the most consideration end of the purchase. I think it's hard to have that be fully AI end-to-end because while you may start doing your research online on ChatGPT or Gemini or any sort of new AI-native property that shows up, the purchase is so significant that you're probably going to want to have some sort of in-person experience where you're seeing the thing, touching the thing, experiencing the thing.

Talking to another human expert about it, right? And so that means there's this whole range of products in the middle that I think we believe the purchasing behavior could be disrupted by AI in a couple different ways. So one is obviously like the research way of like, you know, my handbag wore out that I bring when I travel all the time, like I need the best one that fits a laptop, can fit a big water bottle, all this sort of stuff, can be fine in the overhead part of a plane… And if you're busy, you don't have a ton of time to do that research yourself. And you might ask an AI agent that can watch all the TikToks for you, read all of the Reddit posts and pull in the kind of real consumer feedback and then make a recommendation. And you might want to do some of your own sort of clicking through to look at options. But I would say in that case, it's like decently likely that if there's then a good integration to purchase, you might do it through an AI agent.

There's also sort of just things you already know you want, like Alex has mentioned, where you want the best price. And so I think AI agents can do a lot around price optimization. Like if you always buy a specific type of laundry detergent, it can find across the internet, where is this laundry detergent best priced?

And it can also probably know, “Hey, I should scan this daily. And if it's 30% less on this specific site than it usually is anywhere else, and it's gonna arrive in a reasonable amount of time, I should probably just buy this, and they can store an extra box of laundry detergent because based on what I know about the consumer, that's worth it to them.”

As you move sort of up the consideration stack, there's another sort of purchase that I think will be sort of AI-intermediated, but maybe have some human impact. Things like maybe bikes or couches, like a little bit of higher value purchases, laptops, where you want to feel like someone has taken the time to really understand all of your criteria and help you make the best decision about what you should buy. This is probably an item that you are going to be using for years, and it's important to you that it works and that it's the best option and doesn't kind of immediately become obsolete. And today, I think the only way that this has happened is like people will go super deep into these Reddit threads on the, like buyitforlife forum and all of these different places.

Or they have a brand they really trust like Apple, and they're willing to pay the premium. I think in the future it's fun to think about having an AI agent that really kind of deeply understands you, where you can have a more in-depth conversation about that sort of thing. Like maybe even a phone call where they're asking you a bunch of questions dynamically back and forth, and you're providing them with the information they need to go back, do the research and decide. So that's kind of some of the things that we've considered about how AI could impact purchase behavior.

Alex: Well, and there's another kind of lens, there are many different ways of cutting this, but like, does the product that you're buying have a UPC or no? A UPC is a universal product code, and that's the little scannable thing. It's kind of the successor to the ISBN, which is for books. And if it doesn't have a UPC, actually a lot of the commerce that has worked, kind of post internet 1.0, like, you know, how did Wayfair work? Why did Wayfair work well?

Well, like they're selling things like barstools and you're like, I want a barstool, but there's no UPC on this. So it's like, well, here's a barstool. It fits the right dimensions, but like there's no scannable code. If there is a UPC, you can run this little algorithm of like, get me the lowest price. And pre-AI, you would just run this algorithm on your own, and you would probably end up at Amazon. So like everybody who wasn’t Amazon just got killed, and Amazon did well. I'm oversimplifying a little bit. If it doesn't have a UPC, then that's a little bit of a different process than if it does because if it has a UPC then like the algorithm that I was describing is like exponentially better with AI because like before, you'd have this, like some people value time more than money, and some people value money more than time. And if I value money more than time, I am the algorithm. I need to find the best coupon, I need to find the best cashback site. Like all of these like cashback sites are out there on the internet, which lots of people, money more than time, like they do this.

All of that will be automated away, or automated for the benefit of the consumer if and only if you have something where you have determined the SKU or the UPC. You know, the SKU has a UPC, you feed it in there. Good. If it doesn't have that, then that's another lens where it's like, okay, I'm probably… I can't feed whatever, like, again, kind of impulse to highly considered. AI is gonna help you on the highly considered side, but not help you buy it. But if it spits out something with a UPC or a SKU, then this part of the AI will just automate that.

So if it's like bike, sure, like, I don't know what bike to buy, but if it's a specialized bike and like, here's the thing, it has a UPC on it, like, boom, why wouldn't you feed it to this part of the thing as this gets developed? Because that's just going to buy it for you with the best shipping, the best terms, the best, whatever. And right now, the reason why that doesn't happen, or, well, it does happen, but it happens manually.

And the people that have time valued more for them than money, they don't do any of that stuff.

00:39:52 What net-new commerce companies will be created post-AI?

Erik: Putting this all together, we were talking about how over the last decade there hasn't been a ton of net-new, you know, big winners in this space and a lot of the gains have gone to the aggregators. Why do we believe that over the next decade there's some opportunities for net-new big and durable companies?

And maybe let’s share, what types of companies could exist that we're excited about that could exist?

Alex: I mean, obviously ChatGPT is, I mean they are an upstart to a certain extent, but it's not Amazon. It's not Shopify. It's a net-new company that clearly will have a role in commerce.

The question is, will there be specialized subsegments? And I do think that the, like, you know, the hyper-optimized, I know you very well… And like, you know, CamelCamelCamel is an independent company that's probably very, very profitable, as far as I know. They've never raised venture capital. A lot of people use them. A lot of these cashback sites. Like these things that have always been like, you value money more than time. There was a site called Ebates that was bought by Rakuten a while ago. There's a company in the UK called Quidco, which is very, very similar, like these kind of things, you can imagine them going much, much more mainstream and being very, very specialized shopping agents, particularly not on the heavy research side, but on this one little tiny vector that might actually be very big, going back to how like most companies can't really figure out attribution. It's like we are going to be the last click, like the last click of the 21st century post-AI is going to be AI companies that know how to do this. And it might not be ChatGPT because they're like this horizontal everything. But it's like, I'm going to give you all of my credit cards.

You're actually going to even figure out which credit card you use for this particular purchase because this one has higher cash back than that one for this type of good. And you're going to integrate, like affiliate tracking where you give me cash back like Ebates does, or you know, Rakuten does. You're gonna do all the coupon stuff.

Like, and not all of this will be good for merchants, by the way. But you can imagine, and again, it doesn't require a lot to imagine this because like there already are a lot of companies that do this, but they have been somewhat of a niche space because they only appeal to the like people, and there are plenty of people, that are like this, by the way, that value money more than time and are somewhat technical. So it's actually not just money more than time, it's like, my mom might want to use one of these, probably would value money more than time because she's retired, so why not? But it's just too complicated to use, and if you make it so easy, I mean, that's the other thing. It's like we talked about like how if you make something show up right now that's gonna have a bigger market than if something shows up five weeks from now. Like, that makes sense. If you make something so painfully easy to use that it's more of an IQ test, it's like, do you want to pay less for something or more for something? And like everybody of course would say, “I want to pay less for something.” But it's like, oh, but you have to do these 18 things, download, they're like, “Ah, that's too complicated. I can't figure out how to do that.” But if it's so easy, I think that that's one area where you could… I mean, this is where startups have lived because it's clearly not going to be Amazon because it's like Amazon wants you to shop at Amazon.

Amazon, by the way, like the other thing that they're somewhat imperiled by, Amazon has a giant revenue and profit line item from advertising. It's like you go to the Amazon website, and then you click on an ad that takes you away from the Amazon website. That's a hundred percent gross margin for Amazon.

They'd rather sell you that than sell you a product, but they have to deliver it, God forbid. So like the best SKU that they sell is the advertising SKU, and that's going to be imperiled if they no longer control the presentation layer because like AI intermediates it. But I think it's like this kind of money more than time, expand that to the entire universe. There's certainly a there there.

Justine: I think there's sort of two sides. So I think there's the consumer side, right? Which is, if we go back to my conversation earlier about you want to have a really in-depth conversation about what bike to get, you could imagine someone fine-tuning a model that is much better on tons of conversations between bike experts and people to actually know the right questions to ask to give you a much better buying experience and better outcome than ChatGPT could. So that's one way that like consumer distribution could be disrupted beyond the ChatGPT disruption that will happen. Then I think there's the merchant side of things, which is like what are the implications if we suddenly have a ton of AI agents browsing your site and potentially even making decisions on behalf of consumers and hitting the purchase button instead of people? How should websites change to make themselves more browsable, more easy to interact with, more easy to find what the agent is looking for? What sort of infrastructure do we need on the financial side for AI agents to actually be able to make a purchase on behalf of someone and use their credit card? The entire infrastructure and merchant-facing side of it is probably going to change quite a bit.

And I think that will be just as big as the consumer side of the market.

Erik: I think that’s a good place to wrap. Alex, Justine, thanks so much for the great conversation.

Alex: Thank you

Justine: Thanks for having us.

Resources:

Read the article: http://a16z.com/ai-x-commerce/

Find Alex on X: https://x.com/arampell

Find Justine on X: https://x.com/venturetwins

Stay Updated:

If you enjoyed this episode, be sure to like, subscribe, and share with your friends!

Find a16z on X: https://x.com/a16z

Find a16z on LinkedIn: https://www.linkedin.com/company/a16z

Follow our host: https://x.com/eriktorenberg

They discuss why research is slow, what an AlphaFold-style moment for cell biology could look like, and how AI might improve drug discovery. The conversation also covers hype versus substance in AI for biology, clinical bottlenecks, capital intensity, and how breakthroughs like GLP-1s show the path from science to major business and health impact.

Timecodes

00:00:35 Arc’s moonshot

00:01:56 Why is science slow?

00:05:10 Why AI has been better at language and images than at biology

00:10:18 Virtual cells

00:16:39 The GPT moment for virtual cells

00:19:56 Scaling from virtual cells to organisms

00:22:13 Fixing the pharma industry

00:33:19 Predicting biotech innovations

00:38:15 The state of AI drug discovery

00:42:04 The bull case for AI in bio

00:45:50 Patrick’s approach to investing in AI

00:54:02 Arc’s Virtual Cell Challenge

Transcript

This transcript has been edited lightly for readability.

00:00:35 Arc’s moonshot

Erik Torenberg

Patrick, welcome to the podcast. Thanks for joining.

Patrick Hsu

Thanks for having me on.

Erik

I've been trying to have you on for years, but finally I could get your time.

Patrick

Here I am. I'm excited to do it. It's gonna be great.

Erik

For some of the audience who aren't familiar with you and your work at Arc and beyond, how do you describe, what's your moonshot? What is what you're trying to do?

Patrick

I want to make science faster, right? You know, we can frame this in high-level philosophical goals like accelerating scientific progress. Maybe that's not so tangible for people. I think the most important thing is: science happens in the real world. If it's not AI research, which moves as quickly as you can iterate on GPUs, right, you have to actually move things around, atoms, clear liquids from tube to tube, to actually make life-changing medicines, and these are things that take place in real time. You have to actually grow cells, tissues, and animals. And I think the promise of what we're doing today with machine learning and biology is that we could actually accelerate and massively parallelize this.

And so our moonshot is really to make virtual cells at Arc and simulate human biology with foundation models. And, you know, we'd like to figure out something that feels useful for experimentalists. People who are skeptical about technology, you know, they just wanna see the data and see the results, that it's actually the default tool that they go to use when they want to do something with cell biology.

00:01:56 Why is science slow?

Jorge Conde

Okay, well, hold on, let's back up. Why is science so slow in the first place? Like, whose fault is that?

Patrick

Whose fault is that? Now that is a long one. We should get into it. We should get into it. It's really multifactorial, right? It's this weird Gordian knot that ultimately comes down to incentives, right? You know, people talk a lot about science funding and how science funding can be better, but it's also about how, you know, the training system works, right? How we incentivize long-term career growth, how we, you know, try to separate basic science work from commercially viable work and generally the space of problems that people are able to work on today.

I think things are increasingly multidisciplinary. It's very hard for individual research groups or individual companies to be good at more than two things, right? You might be able to do, you know, computational biology and genomics, right? Or you know, like chemical biology and molecular glues. But you know, how do you do five things at once? It's increasingly hard.

And we really built Arc as an organizational experiment to try to see what happens when you bring together neuroscience and immunology and machine learning and chemical biology and genomics all under one physical roof, right? If you increase the collision frequency across these five distinct domains, there would hopefully be a huge space of problems that you could work on that you wouldn't be able to.

Now, obviously in any university or any, kind of, geographical region, you have all of these individual fields represented at large, right, across these different campuses. But, you know, people are distributed, and you want everyone together.

Jorge

Okay. But if I may, I would've thought a university was an attempt to bring in multiple disciplines under one roof. You're saying it's not, it's too diffuse.

Patrick

It's across an entire campus.

Jorge

Okay. So the physical, like literally the physical distance creates inefficiency.

Patrick

That's part of it. And I think the other part is folks have their own incentive structures, right? They need to publish their own papers, they need to do their own thing and you know, make their own discovery and you're not really incentivized to work together, I think in many ways in the current academic system. And a lot of what we've done is to try to have people work on bigger flagship projects that require much more than any individual person or group or idea.

Jorge

Yeah, that's cool. So that's sort of the original hypothesis for the Arc Institute is if you can bring multiple disciplines together to increase the collision frequency, as you said, and, if one could remove some of the cross incentives that may exist in sort of traditional structures, the combination of those two things will make science faster.

Patrick

Yeah. These are, these are absolutely part of it, right? We have two flagship projects: one trying to find Alzheimer's disease drug targets, the other to make these virtual cells. I think it's not just the people and the infrastructure but also the models will hopefully literally make science faster. That you could, you know, do experiments at the speed of forward passes of a neural network if these models could become accurate and useful.

Jorge

Yeah. So that will be one thing that solves the length of discovery as you compress the time discovery takes naturally by just throwing technology at the problem. At the risk of oversimplifying.

Patrick

Well, we're techno-optimists here, no?

00:05:10 Why AI has been better at language and images than at biology

Erik

We are. Why has AI progressed so much faster in sort of image generation and language models than biology? And if we could wave a wand, like where are we excited to speed things up?

Patrick

To be honest, it's a lot easier, right? Maybe that's a hot take.

Jorge

You mean technology is easier than biology.

Patrick

Natural language and video modeling is easier than modeling biology. And to some degree, like if you understand and learn machine learning, right, and how to train these models, you have already learned how to speak. You already know how to look at pictures. And so your ability to evaluate the generations or the predictions of these models are very native, right? We don't speak the language of biology, right? At very best with an incredibly thick accent.

So when you're training these DNA foundation models, I don't speak DNA natively, so I only have a sense of the types of tokens that I'm feeding into the model and what's actually coming out, right? Similarly with these virtual cell models, you know, I think a lot of the goal is to figure out ways that you can actually interpret the weird, fuzzy outputs that the model is giving you.

And I think that's what slows down the iteration cycles is you have to do these lab-in-the-loop things where you have to run actual experiments to actually test with experimental ground truth. And you know, I think increasing the speed and dimensionality of that is gonna be really important.

Jorge

You talk about, we speak biology poorly or with a very thick accent.

How much of this is like if you're training on an image, we can see the image, and so we can see how good the output is. What about all the things in biology that we can't see or don't even know exist yet? Like how can we create a virtual cell, and maybe we should come back to what a virtual cell model is, by the way, for the lay audience, but like, how can we create a virtual cell model when we're not even sure if we understand all of the components that are in a cell and how they function.

Patrick

People talked a lot about this in NLP as well. There's this long academic tradition in natural language processing, right? And then it was just weird and non-intuitive and intensely controversial that you could just feed all this unstructured data into a transformer and it would just work. Now, we're not saying this will just work in all the other domains, including in biology, but I think there is this, you know, controversy around what does it mean to be an accurate biological simulator? What does it mean to be a virtual cell?

It's true. We can't measure everything, right? We can't measure, think things like metabolites in really high throughput with spatial resolution. And there are gonna be different phases of capability where initially they model individual cells, then they model pairs of cells, then they model cells in a tissue, and then in a broader physiologically intact animal environment.

And those are length scales and kind of layers of complexity that we’ll aggregate and, you know, improve upon over time. And I think the other kind of non-intuitive thing in many ways are the scaling laws that you get in data and in modeling. I'll give you an example, right? There's a lot of discussion in molecular biology about how RNAs don't reflect protein and protein function, right?

And so while we don't have, you know, proteomic measurement technologies that are nearly as scalable as transcriptomic measurement technologies today like at the single-cell resolution certainly, but we're getting there, and you can layer on certain nodes of protein information that you can add on top of the RNA information, but in many ways, the RNA representation is a mirror, right? It might be a lower resolution mirror for what's happening at the protein layer, but eventually what is happening in protein signaling will get reflected in a transcriptional state, right? And so for an individual cell, this may not be very accurate, but when you imagine the massive data scale that we're generating in genomics and functional genomics, you start to gather tremendous amounts of RNA data that will read in kind of like what's happening at the protein level as some sort of mirror echo, right? And then that can, you know, be the case for metabolic information as well, and so on.

Jorge

It's a low-pixel image, but if we can get sort of zoomed out far enough, we'll get a sense of what's going on.

Patrick

You have to bet on what you can scale today, right? We're able to, you know, scale single-cell and transcriptional information today. We're able to add on, you know, protein-level information. Over time, we'll need spatial information, spatial tokens, and we'll need temporal dynamics as well.

I kind of bucket things into three tiers. There's invention, engineering and scaling. And there are certain things today, biotechnologically that are scale-ready, and then there are things that we still need to invent, right? And that's part of why we felt like we needed a research institute to be able to tackle these types of problems.

That we weren't just going to be an engineering shop that's just trying to scale single-cell perturbation screens, right? That, you know, would be interesting, but in three years would feel very dated, I think. And so there's a lot of novel technology investment that we're making that we think will bear fruit over time.

00:10:18 Virtual cells

Erik

Can we flesh out the virtual cell concept, why that's the ambition we've landed on, what it's gonna take to get there? Or what are the bottlenecks?

Patrick

I would say the most kind of famous success of ML in biology is AlphaFold. And this solved the protein folding problem of, you know, when you take a sequence of any amino acids, what does the protein look like?

And you know, it's pretty good. It's not perfect. It certainly doesn't simulate the biophysics and the molecular dynamics, but it gives you a sense of what the end state is with 90+% accuracy, right? And that's the AlphaFold moment that people talk about, where anytime you want to work with a protein, if you don't have an experimentally solved structure, you're just gonna fold it with this algorithm.

And we kind of want to get to that point with virtual cells as well. And the way that at Arc we're operationalizing this is to do a perturbation prediction, where the idea is you have some manifold of cell types and cell states. That can be a heart cell, a blood cell, a lung cell, and so on, and you know that you can kind of move cells across this manifold, right?

Sometimes they become inflamed. Sometimes they become apoptotic, sometimes they become cell cycle arrested. They become stressed. They're metabolically starved. They're hungry in some way. And so if you have this sort of representation of universal, sort of, cell space, can you figure out, what are the perturbations that you need to move cells around this manifold? And this is fundamentally what we do in making drugs, right? Whether we have small molecules, which started out as natural products from, you know, boiling leaves or antibodies when we injected proteins into cows and rabbits and sheep and took their blood to get those antibodies, where we are basically trying to get to more and more specific probes, right?

And we had experimental ways to kind of cook these up. Now we have computational ways to zero shot these binders, but ultimately what you're trying to do with these binders is to inhibit something and then by doing so, kind of click and drag it from a kind of toxic, gain-of-function, disease-causing state to a more quiescent, homeostatic, healthy one.

And the thing that is very clear in complex diseases, right, where you don't have a single cause of that disease is there's some complex set of changes, there's a combination of perturbations, if you will, that you would wanna make to be able to move things around. Now, you know, people talk about this classically as things like polypharmacology.

But, you know, I think we're moving from, “Oh, this thing happens to have a whole bunch of different targets, kind of by accident,” to, “We have the ability to manipulate these things combinatorially in a purposeful way,” right? That to go from cell state A to cell state B, there are these three changes I need to make first, then these two changes, and then these six changes over time, right? And we kind of want models to be able to suggest this. And the reason why we scoped virtual cell this way is because we felt it was just experimentally very practical. You want something that's gonna be a copilot for a wet lab biologist to decide “What am I gonna do in the lab?”

We're not trying to do something that's like a theory paper that's really interesting to read, where the numbers go up on an ML benchmark, but you know, you practically can decide what are the twelve things that you're gonna do in the lab in twelve different conditions and actually just test them.

And then that's how we kind of enter the kind of the lab-in-the-loop aspect of model predictions to experimental measurements to kind of improved, or RL-ed, or whatever, model predictions again. And the goal is to be able to do in silico target ID where you can basically figure out new drug targets, figure out then the compositions, the drug compositions you would need to actually make those changes.

I think if we could do that, we could make a new like vertically integrated, AI-enabled pharma company. Which, you know, I think is obviously a very exciting idea today, but I think in many ways the kind of pitch and the framing of these companies precedes the fundamental research capability breakthroughs.

And that's what we are really invested in at Arc, is kind of just making that happen along with many other amazing colleagues in the field to just make this possible for the community.

Jorge

So if the goal is, I'm going to oversimplify it for you, like if we wanted to get to the AlphaFold moment where, you know, it kind of gives you a useful structure, folded structure 90% of the time to use your data point, we wanted to take that comparison in the virtual cell model, and we said, “Okay, 90% of the time if I ask the model, I want to shift the cell from cell state A to cell state B, and it's gonna give me a list of perturbations. And let's say that at 90% of the time, those perturbations in fact result experimentally in the shifting from cell state A to cell state B.” How far away are we from that AlphaFold moment for virtual cells?

Patrick

I find it helpful to frame these in terms of like GPT-1, 2, 3, 4, 5 capabilities. And I think most people would agree we’re somewhere between GPT-1 and 2, right? A lot of the excitement was that we could achieve GPT-1 in the first place, that you could see a path with scaling laws of some kind to make successive generations where capabilities would improve. But you know, with like our Evo kind of DNA foundation models that we developed at Arc with Brian Hie, one of the things that we've seen is that, you know, these genome generations are like quote unquote “blurry pictures” of life, right?

We don't think if you synthesize these novel genomes, they would be alive. But, you know, we don't think that's actually also impossibly far away. We'll just have to kind of follow these capabilities. We're taking a very integrated approach to attack this problem where you need to curate public data, you need to generate massive amounts of internal private data, build the benchmarks and train new models and build new sort of architectures and kind of doing these things full stack. And we'll just kind of attack this hill climb over time.

00:16:39 The GPT moment for virtual cells

Erik

What's the GPT, I'll say GPT-3, moment going to look like, and by that I mean sort of a public release that alters the public's conception of just what's possible here from a capabilities perspective and also inspires a whole new generation of talent to like rush into biology.

Patrick

Well, the good thing with biology is we have a lot of ground truth, right? There are entire textbooks, right, that describe cell signaling and cell biology and how these things work. And so, you know, even without a virtual cell model at all, if you went into ChatGPT or Claude, and you asked it some question about, you know, like receptor tyrosine kinase signaling, it would have an opinion on how that works, right?

And so I think you would want the model to be able to predict perturbations that are kind of famous canonical examples of biological discovery. So I'll give you an example. If you've loaded into the model an iPSC, kind of an induced pluripotent stem cell state or human embryonic stem cell state, and a fibroblast cell state. Could it predict that the four Yamanaka factors would reprogram the fibroblast into a stem-like state and essentially rediscover from the model something that won the Nobel Prize in 2009? That would be sort of one really kind of classic example.

And then you could go do the inverse. If you have a stem cell, can it discover neurogenin 2, ASCL1, MyoD? Can it find differentiation factors that will turn that into a neuron or into a muscle cell or so on? And you know, these are kind of classic examples in developmental biology, but you could also use this to try to discover or kind of recapitulate the mechanism of action of FDA-approved drugs. And so you could say, for example, you know, if you kind of inhibit Her2 in breast cancer cell states, you would get this type of response. Or it could predict the certain clones that, you know, will be able to kind of be more metastatic or they'll be more resistant and they'll lead to minimal residual disease. There are, I think, lots of kind of biological evals that you can kind of add on to these models over time that are really tangible textbook examples as opposed to, I think what the kind of early generation of models do today, which is, you know, very quantitative things like mean absolute error over the differentially expressed genes and stuff like that. Those are ML benchmarks. And we want to increase the sophistication into something that you could explain to an old professor who has, you know, never touched a terminal in their life.

Jorge

By the way, you talk about textbooks as ground truth. Do you think we're going to find that a lot of the textbooks are wrong?

Patrick

I would say textbooks are compressed. So for example, when you look at these kind of classic cell signaling diagrams of A signals to B, which inhibits C, right? That's a very kind of two dimensional representation of…

Jorge

Of our understanding of a complex system.

Patrick

Right. I mean, yes, textbooks are what they are. They represent the corpus of reliable knowledge, but everyone knows that there are an incredible number of exceptions, and part of what discovery is is to find new exceptions.

00:19:56 Scaling from virtual cells to organisms

Erik

Why don't you talk about the difference between the simulation of biology and the actual understanding? And what would it take to actually be able to model the extremely complex human body?

Patrick

You know, some people don't like the phrase “virtual cells” because it sounds too media-friendly. It's not rigorous enough. But I've always found it funny that you know many people are okay with like “digital twins” and “digital avatars,” which, you know, talks about modeling biology at a way higher level of abstraction.

You know, I think virtual cells, if anything, is actually way more scoped and rigorous than modeling a digital twin or avatar. But, you know, I think these are useful words because they describe the goal and the ambition, right? That no, in the long run we don't care about predicting the, you know, kind of perturbation responses of an individual cell at all, actually. Obviously, we want to be able to predict drug toxicity. We want to be able to predict aging. We want to be able to predict why a liver cell becomes cirrhotic when you repeatedly challenge it with ethanol molecules or whatever, right? And you know, these sort of chemical or environmental perturbations should be predictable.

I think you just kind of have to layer on the complexity, right? Like, why are we so worried about modeling entire bodies over time when we can't do it for an individual cell, where we sort of, you know, accept or broadly believe that this is a fundamental unit of biological computation, if you will. And let's just kind of start there, right? Just like you kind of have to start with, you know, things like math and code and language modeling, right? And things that are just sort of easier to check. You can build to superintelligence over time.

Jorge

Yeah. I think that makes sense, right? That's a very sort of laudable, ambitious goal, if we can figure out how to model the fundamental unit of biology, the cell, then from that, we should be able to build.

Patrick

Like in early AI, we just started with like language translation. Just, you know, basic NLP tasks, right? This was long before, you know, the tremendous ambitious scope that we have today. And I think we hopefully can mirror that type of trajectory, if we're lucky.

00:22:13 Fixing the pharma industry

Erik

It seems like biotech and pharma has been a shrinking industry, certainly the rate of growth. What’s it going to take for these innovations in the science to reflect themselves in business models and in growth for the industry.

Patrick

A lot of these biotech startups would try to initially sell software to pharma companies, and then they would kind of realize, “Oh wow, we're like competing for SaaS budgets, which aren't very large.” And then, you know, now they're realizing, “Oh, we have to compete for R&D budgets.” And I think, you know, there's this narrative from the current generation of these companies that, “Oh, our biological agents will compete for R&D budgets and replace headcount,” or something like that, just like we're seeing in, you know, agents across different verticals. Whether or not that will, I think, pan out, I think depends on just whether or not these things meaningfully allow us to, you know, build drugs more effectively in the pharma context. And I think that's just sort of the most important thing in this industry.

And so I think we believe in virtual cells, not just because we think it will be a fountain of fundamental mechanistic insights for discovery, but also because in the case of success it could be industrially really useful. But, you know, we'll have to see over time, right? If we have 90% of drugs failing clinical trials, that kind of means two things, and you're not sure what percent of which, right? One is we're targeting the wrong target in the first place. The second is the composition, the drug matter that we're using doesn't do the job. It's not clear for each individual failure which one it is, or if it's both, or what proportion of each.

And you know, we'll have to kind of sort that out over time. Like you can imagine, even in the case of success, when we have 90% accurate virtual cells, you'll probably end up with suggestions like, “Okay, now you need to target this GPCR only in heart. But not in literally any other tissue.” We don't have the drug matter that can do that today.

And so that's also why, again, you probably need research to figure out novel chemical biology matter that allows you to drug pleiotropic targets in a tissue or cell type-specific way. And so, you know, I think part of why biology is slow is because there's just this Russian nesting doll of complexity, in terms of understanding, in terms of perturbation, in terms of safety, and, the crazy thing is the progress in just the short time that I've been doing this is insane, right? Like I did my, you know, PhD at the Broad Institute in the heyday of developing single-cell genomics, human genetics, CRISPR gene editing, and, so many other things. And I think the kind of early 2010s papers on single-cell sequencing would have like 20 cells or 40 cells. And at Arc in the next, I don't know, relatively short amount of time, we're gonna generate a billion perturbed single cells. I mean, how's that for a Moore’s Law?

Jorge

Yeah. That's remarkable.

Erik

Jorge, I want to hear your answers to a couple of these questions too, as the lead of our bio practice, both on the GPT-3 moment, what that could look like. And also, like I'm curious if you think it's GLP-1s or sort of building off that, or if it's gonna be something different. And also, what's it gonna take for the science to kind of reflect itself in the business, for the industry to grow?

Jorge

Yeah. So I'll take the second one first if I could. So I think, you know, in terms of where the industry is right now, I think one of the big challenges we have is, as Patrick describes very nicely, like, you know, discovery is hard, and it takes time. And, you know, the fail modes are exactly as you described. Oftentimes when drugs fail, which they do 90% of the time in clinical trials, it's because we're going after the wrong thing, or we made the wrong thing to go after the right thing, right? Like those are the two fail modes and that happens all too often. And so I think a lot of the stuff that Patrick is describing is going to basically improve our hit rate or our batting average on figuring out what to go after and then making the right thing to go after said thing. The challenge we have, I think, in the industry is that the bottlenecks still are the bottlenecks. And the biggest bottleneck we have, which is, you know, a necessary one, is we have to prove that whatever we make, that we have the right thing to go after the right thing, so to speak, and that when we have it, that it's going to be as, you know, de-risked as possible before you put it into humans.

Patrick

And we have to be good at making them in the first place.

Jorge

And we gotta make them too. Yeah, exactly. And so that bottleneck is a necessarily important one. That bottleneck should exist. I'm not suggesting we've gotta remove it, but are there ways to reduce the cost and time associated with getting through the bottleneck of human clinical trials? And you know, it's interesting because we talk about, you know, all of the various stakeholders when you're making a drug. There are the companies, there's of course the science that supported the company that's trying to commercialize a product, and there are the regulatory agencies. You know, and everyone is trying to ensure, again, that what’s first and foremost is the ability to discover and commercialize drugs that are safe and effective for humans.

That middle part part of actually getting through that bottleneck is hard to speed up in a very obvious way. Like you can increase the rate, the way you enroll clinical trials. You can use better technology to change the way we design these clinical trials so maybe they can be faster or shorter, etc.

But some of them just have a natural timeline they have to go through. Like if you wanna demonstrate that a cancer drug promotes survival, guess what, it’s going to take some time to demonstrate a survival benefit. Or if you know you want to do a longevity drug, that by definition is a lifetime of a trial in terms of length.

So there's a lot of these bottlenecks that are really hard to get through. So what helps the industry? I think there are a couple of things that help the industry. One is capital intensity will hopefully at some point go down over time as technology gets better. Capital intensity is something that our industry faces. In some ways, it looks a little bit like AI now, right, in terms of the cost of training these models. But the capital intensity is very, very high. That has not come down. So, we gotta get to success rates up to impact capital intensity to get it down. The second thing is, where can we compress time? So good models can help us compress early discovery time.

We still haven't seen—and I think it's coming, but it hasn't happened yet—we haven't seen artificial intelligence or other technologies massively compress the amount of time it takes us to do the clinical development, the clinical trials, the enrollment of patients, all those things. We're seeing some interesting things coming. We haven't seen sort of the payoff there yet. And the third thing is if we can make better drugs going after better things, the effect size should be higher. So therefore the answer should be obvious sooner. If we can get those three things right, reduce capital intensity, compress timelines, and effectively increase effect size in some intractable diseases, that is what I think fixes the industry. And from where we sit at the early stage in terms of being early stage investors, the reason why that helps us is if the capital intensity goes down, and the value creation goes up, it becomes easier to invest in these companies in the early days because you get rewarded for coming in early. The problem we have right now is that most companies aren't—you're not seeing rewards happening when there's value inflection.

So you come in early, you bear the brunt of the capital intensity, and even if a company is successful, that success isn't reflected in the valuation. So we're not seeing the step ups that you see in other parts of the industry, and that's just really, really hard from an investing standpoint. So I think we need to see those various factors addressed for this space to really get, you know, fixed, to use your word.

Patrick

Yeah, that was great. I have a lot to add onto this.

Jorge

Please. Add away.

Patrick

A few simple observations. The first is the amount of market cap added to Lilly and Novo, based on the development of GLP-1s, is like over a trillion dollars, or, you know, I mean, Novo stock has decreased a lot, so a trillion dollars, let's say, is more than the market cap of all biotech companies combined over the last 40 years that have been started. And I think that, you know, one of the kind of interesting corollaries of this is that, you know, when we have a 10% kind of clinical trial success rate for a kind of preclinical drug matter, you tend to circle the wagons a bit and try to manage your risk, right? And so the way that you do this is you try to go after really well established disease mechanisms where if I developed new drugs that go after well understood biology, it should work the way that I hope it will in the trial, which is really, really expensive and costs a lot more in many ways than the preclinical research.

The problem with this is you go after very well validated disease mechanisms, but with really small patient populations, right? So then the expected value of this actually is relatively low. One of the kind of things that we've seen with GLP-1s is the, just the kind of value that you can create when you go after really large patient populations.

And I think that has culturally really net increased the ambition of the industry, both from the investor and from the drug developer side. And I think, you know, that's something that we should keep our foot on the gas for.

Jorge

Yeah. And look, I think the trend on that is, I would argue the trend on that is positive.

So you're absolutely right. Like the demonstration of the value that has been created with the increasing use of GLP-1s and the value transfer that's gone to companies like Lilly and Novo, I would argue is like very merited, right? Because they've cracked an endemic social problem, in terms of managing diabetes and eventually helping manage obesity.

And so I think that's remarkable. And there's a lot of value that goes to that because they tackled, they cracked a very, very challenging problem for society beyond just science. So that's great. And I agree with you like the juice needs to be worth the squeeze. You're right. A lot of biotech has been around like, go after the low hanging fruit because it's low risk and we gotta eat today. So you go get it, you know, and you start to, you push off the big ambitious indication, the large population or the really tough-to-crack disease. But you know, I do think we're seeing more and more of that.

And by the way, like we can get into some of these genetic medicines, but some of these genetic medicines are going after some of the hardest problems, the things that you quite literally couldn't address but for editing DNA. And, you know, I think that's incredibly, you know, remarkable and laudable and frankly inspiring. But the fundamental elements of the industry have to work so the capital formation is there to support those kinds of things. And right now it's hard, right? Because of the issues we talked about before.

00:33:19 Predicting biotech innovations

Erik

15 years from now, we're back in this room, we've barely escaped being part of the permanent underclass, and we're reflecting on sort of the GPT moment or maybe the legacy of GLP-1s, sort of beyond where they are now. Jorge, I'm curious to get your take on, what do you think is gonna be the technological breakthrough that we're gonna point back to and say, “Oh, this is really what set it all”? Or do you think it's gonna be sort of, you know, multifactorial, a combination?

Jorge

Yeah, look, I think it's going to go back to sort of where we started this conversation. GLP-1s as a drug are, you know, what, four decades in the making or something like that.

You know, these are not overnight successes. But I do think what we are going to see more of and our hope is that when you combine the fact that we're getting better at understanding what to target, getting better at designing medicines to hit those targets, by the way, in a whole array of new creative ways. So we have small molecules. The natural products that we got from boiling leaves, as you said earlier. We're getting really good at designing smarter and better small molecules that do new things, that function in ways that they didn't before. We've gotten quite good at designing biologics or proteins with a lot of help from things like AlphaFold that helps us understand how proteins fold.

We're gonna get a lot better at designing some of the more complex modalities, like the gene therapies of the world or the gene editors of the world. And when you can do that and combine that with our ability to hopefully use things like virtual cell models to really understand what to go after, like we're gonna have drugs… I would hope, and I would expect, that the industry will continue to bring forward drugs that have very large effect size for very difficult diseases that hopefully affect a lot of patients. If that's true, then we'll start to see some of these really, really difficult diseases that affect all of society get tackled, hopefully, you know, one by one by one by one.

And so we have obesity, we have metabolic disorder. We're dealing with cardiometabolic disease. We're starting to see interesting, promising things happening in like neurodegenerative diseases. You know, if we can tackle cancer, or at least you know, several cancers that now have begun to be treated more like a chronic condition than a death sentence that they were in the past. The more we see of that, like I think that value to society will accrete over time. And I think this should be an industry that is extraordinarily valued by society and, candidly, by the markets. We have to deliver.

Patrick

If we play this out, and let's say these AI models work, and you can make a trillion binders in silico, that will, you know, be exquisite drug matter, right?

We still need to make these things physically and test them in animals and hopefully predictive models and then actually in people. And I think, you know, that will increasingly be the bottleneck in many ways. And, you know, my friend Dan Wang recently released a book called Breakneck, which talks about, you know, kind of like the US and China and the difference between the two countries and their philosophy, the way they approach markets and…

Erik

We're a country of lawyers, they’re a country of engineers, at least their political class.

Patrick

Exactly. That's right. China is an engineering state, right? Politburo is, you know, folks who have engineering degrees, you know, you need to build bridges and roads and buildings, and these are the ways that we solve our problems. Whereas I think from, you know, the first 13 American presidents, 10 of them practiced law. From 1980 to 2020, all Democratic presidential candidates, both VP and president, went to law school. And so you kind of see the echoes of that in the FDA and the regulatory regime and, you know, all the kind of bottlenecks that people talk about developing drugs stateside. And increasingly you see folks thinking about how we can run phase Is overseas, right, build data packages that we can, you know, bring back domestically for phase II efficacy trials. I think that's interesting, directionally, but it's not enough. And you know, I think we need to kind of figure out these two bottlenecks, the making and the testing. Even if we can solve the designing part.

Jorge

Oh, I agree. Yeah, that’s. That's the bottleneck. You know, we joke about it, and what you have to do is you have to get a molecule that can go, you know, first in mice, and then in mutts, and then in monkeys, and then in man.

You know, that takes a long time, and it's just so hard to compress that. And so when you do, you should make the journey worth it, right? So when you fail on the other end of that, like, that's obviously horrible. And so finding ways to make sure that when you walk that path, that it'll be a successful journey as often as possible, is what this industry desperately needs.

00:38:15 The state of AI drug discovery

Erik

AlphaFold solved the protein folding problem, but why didn’t it solve drug discovery, or more broadly, what would it take to get AI drug discovery? What is sort of the bottleneck, on the tech side at least?

Patrick

On the tech side?

Jorge

Maybe another way to ask the question, because I always ask the founders a version of this question, like the AI ones that are like, “Oh, we're gonna do AI for drug discovery.” So my question that I always like to ask founders is: give me examples of where you think AI is hyped, potentially overly hyped, where there's real hope, like the sort of, “What do we expect,” “What's next,” and where we already see real heft. So like if I asked you like in AI, you know, where is there hype, where is there hope, and where are we seeing heft today?

Patrick

I would say there's hype in toxicity prediction models.

Jorge

Okay. So that's the idea that we will say, I'm going to show you a molecule and the model is going to tell me if it's going to be toxic or not.

Patrick

That's right. There's heft in anything to do with proteins. Obviously protein binding, but increasingly in protein design. I think there's real heft there. And then, you know, where there's hype is in multimodal biological models, whatever that means. And I think, you know, pick your favorite layers. It could be, you know, molecular layers. It could be spatial layers. It could be, you know…I mean, actually I would say there's also heft in the pathology AI prediction models. Like, you know, automating the work of pathologists and radiologists. That's interesting.

Jorge

Yeah. I think that's a very powerful use case for sure.

Patrick

And there's a lot of stuff where you don't have to train, you know, weird biology foundation models and you can write, you know, regulatory filings and reports and things like that. That's impactful and important.

Jorge

So now going back to Erik's question is, why hasn't AI turned out drugs yet? I think that was your question, right?

Patrick

You know, AI for drugs is one of these weird things where everyone who works in the industry is trying to claim that their drug is like the first AI-designed molecule. I feel like, you know, I mean, increasingly in just a few years, this will just be a native part of the stack. Just like we use, you know, the internet and we use phones, we're gonna have AI in all parts of the stack. And so it's just going to become a native part of everything that we do. And so, you know, like, “Why hasn't it worked yet” is this long multifactorial process that we've been talking about today.

There's designing, there's the making, there's the testing, there's the approvals side of it. And you know, I do think safety and efficacy as the kind of two pillars in the industry are the two things that we need to get right. We need to be able to figure out faster ways that we can predict whether or not a molecule will work and if it's going to be safe or not.

I mean, there are like ways that AI can operationalize this. If you designed a small molecule, you could now computationally dock it to every protein in the proteome and see if it's likely to bind to off-target molecules. You can use this to tune binding selectivity and affinity. That might be ways to predict, you know, safety and efficacy. And, you know, how well will that work? Well, that's a feedback loop that we'll have to actually test in the lab. And that's part of what's slow is the testing, you know, takes real hours, days, months, right, years. And that's really why we've picked at Arc the virtual cell models as our initial wedge because we think it can integrate a lot of these different pieces.

00:42:04 The bull case for AI in bio

Erik

In Dario Amodei’s essay, “Machines of Loving Grace,” he predicts, among other things, the prevention of many infectious diseases and the doubling of lifespans perhaps in as soon as the next decade. What’s your reaction to his essays, his bullishness, and some of his predictions?

Patrick

I think the core intuition that Dario had was the idea that important scientific discoveries are independent, or they're largely independent, and if they are, you know, statistically independent, then it would stand to reason that we could multi-parallelize. And so if we had models that were sufficiently predictive and useful, you could have not just a hundred of them, but millions, billions of these discovery agents or processes running at a time, which should compress the timeline to new discoveries, and turn it into a computation problem. I think that is a very futuristic framing for something that is actually very tangible today.

And if we can have virtual cell models that work, for example, that can start to do these kinds of things that we've been talking about. We can have, you know, molecular design models, we can have docking models. We can then have, you know, when you bind to this thing in this cell versus all the other off-target proteins, will a cell kind of be corrected in the right way? These kind of layers of abstraction and complexity start to get to things that feel very tangible through drug discovery. If you could actually traverse these steps reliably and in sequence, you could start to see how you can get the compression. And so I think in the long run of time, this should be possible.

Jorge

One of the core suppositions in building a good virtual cell model is that we are feeding it all the relevant data.

Patrick

The right data, yeah.

Jorge

The right data. And so we'll work to, you know, it's gene expression data or it's DNA data or you know, any number of factors, protein and protein interactions, all the things you described. What if we're missing a core element? Like what if we just haven't discovered the cork or whatever? Like we just don't know what we don't know, and therefore what we're feeding the model is fundamentally or importantly incomplete.

Patrick

I think that's almost certainly true, right? Like it seems almost obvious that we're not measuring many of the most important things in biology, right? And you can of course find many important exceptions for any of these measurement technologies. Like in biology, we ultimately have two ways to study it in high throughput. It's imaging and sequencing, right? But there are so many other types of things that you would care about that those things aren't necessarily going to do at scale.

And that's really why I think the stuff that we're talking about of the RNA layer as a mirror for other layers of biology is one that we spent a lot of time thinking about. And there's a difference between a mechanistic model and a meteorological simulation type of model. So, for example, if you want to predict the weather, right, you can build AI models that will predict whether or not it will rain next Tuesday. It won't explain physically or geologically or whatever why and how that happens. But as long as it knows if it's gonna rain next Tuesday, you're probably happy, right? And I would say similarly with a virtual cell model, it may not tell me literally why, just like AlphaFold doesn't tell me literally why did the protein fold this way and how. But it just told me the end state, and it was reasonably accurate. I think that would already be very important.

00:45:50 Patrick’s approach to investing in AI

Erik

Shifting gears a little bit, we've been talking about science and biotech, but in addition, you're an elite AI investor more broadly, so I want to talk about where your investment focus is, right now, just as it relates to AI more broadly. Where are you excited? Where are you spending time? Where are you looking forward to?

Patrick

Yeah, my goal is to really try to figure out ways that we can improve the human experience in our lifetime. If I think about the future that we're gonna leave to our children, there are a few things that if we get them right in our lifetime, will fundamentally change the world and, you know, how we live in it. I think synthetic biology is obviously one. You know, think GLP-1s, right? Things that improve sleep, things that can, you know, improve longevity, right?

These are, these are all things that are kind of, you know, easy to get excited about. I think brain-computer interfaces is another area where we're gonna see really important breakthroughs over the decades to come. And then I think the third is in robotics, both industrial and consumer robotics, that allow us to basically like scale physical labor right in interesting ways. And, you know, you can kind of see how each of these three things, even in the sort of medium cases of success, really kind of change the world. And so I'm very interested in helping make these kinds of things possible.

In the kind of techno-optimist sort of vision of the world, there's a few different types of scarcity. It's very easy when you do research to come up with important ideas. The hard thing is to tackle them in the right timeframe. It's like, you know, writing futuristic sci-fi things is not that hard. Being able to actually execute on it in the next five years or eight years, much, much harder. And I would say, you know, academic discovery is littered with plenty of ideas that are interesting and important, but you know, kind of long before their time, and in many ways the story of technology development is, you know, trying to use new technologies to solve old tricks, right? Like most of our tools are, you know, for productivity in many ways, whether that's the industrial revolution or the computing revolution, or the current AI revolution. We're trying to kind of do the same stuff.

You know, I think there's a relatively small set of very powerful ideas. New technologies give us new opportunities to attack them, and there's a set of people and teams that are gonna be positioned to be able to do that. They need to have technical innovation and then an intuition about product and business in a way that, you know, you kind of in the RPG dice role of the skills that you get in these three domains. People start at different base levels. And, you know, you might have an incredibly technical founder who doesn't know how to think commercially or someone who's just natively a very commercial thinker who, you know, doesn't have very strong product sense even though they could sell the crap out of it.

And so I think these sort of three broad categories of capabilities you need to kind of bring together in a way that you can allocate capital to in the right times in order to make these ideas possible in a really differentiated way. Like this thing literally wouldn't happen if we didn't get these people together and fund it at the right time in the right way. And that's really what motivates me. And these are the kinds of things that I've been excited about, you know, backing, you know, longevity companies like NewLimit, BCI companies like Nudge, robotics companies like The Bot Company. These are some of the examples of things that I think must happen in the world and therefore should happen and, you know, how do we actually find the right people in the right time to actually kind of go on the Fellowship of the Ring hunt.

Erik

Yeah. If it’s not too difficult. I wanna ask Jorge’s question adapted to these additional spaces, robotics, BCIs, and longevity, if appropriate. And the three questions I believe were, what's overhyped, where do you see opportunity or a path, and what's got heft already?

Patrick

I think the cool thing about agents, generally, is that they do real work. Compared to like SaaS companies that came before, agents replace real productivity. And I think, you know, they have a lot of errors today and I would say the computer use agents will probably trail the coding agents by maybe a year.

But it's coming. And we'll follow the trajectory as these go from doing, you know, minutes of work without error to hours to days. And I think, you know, you're gonna get a completely different product shape as we march through that across legal, BPO, you know, medicine, healthcare, whatever, right?

And we'll kind of follow that as an industry and that's going to be really exciting. And I think that's where we're going to see real heft because most of the economy is services spend. It's not software spend. And, you know, the reason why we're all excited about this stuff is that it can attack the services economy. And I would say like, you know, where is there hype? There's a tremendous amount, right? That's no doubt. The hype is in the model capabilities. And you know, we're working with an architecture that, you know, dates back to 2017. And if you look at the history of deep learning, it's like kind of every eight years, there's something really different. It feels like in 2025 we're really overdue for some net new architecture. And I think there are lots of really interesting research ideas that are bubbling up that could do that thing. And in many ways there's a set of really interesting academic ideas, especially in the golden age of machine learning research from, I don't know, like 2009 to 2015, right? There's so many interesting ideas and little arXiv papers that have like 30 citations or less. And as the marginal cost of compute goes down year on year, I think you're gonna be able to take all of these ideas and actually scale them up, right? You don't see the scaling laws when you're training them at a hundred million or 650 million parameters like back then.

But if you can scale them up to 1B, 7B, 35B, 70B, you start to see whether or not these ideas will pop. And I think that's very exciting because, you know, there's just going to be a lot of opportunity for new super intelligence labs to do things beyond what the kind of established foundation model companies are doing today, as they kind of, you know, in addition to these research teams, these are in many ways becoming applied AI companies, right? They need to build product shape and, you know, all kinds of different enterprises and do RL for businesses and make money, right?

Or build coding agents and make API revenue and that's important and I think, you know, a timely race to survive today. But I'm just very bullish on the research of say, like a Sakana AI, which was founded by one of the authors of “Attention Is All You Need,” Llion Jones.

And they're doing incredibly interesting stuff on model merging and how you can have sort of like evolutionary selection of, you know, different, models in MoE. And I think the sort of opportunities here, in the long run to move beyond just like RL gyms, for example, also to kind of figure out new ways to learn and find like kind of reward signal is going to be really exciting.

00:54:02 Arc’s Virtual Cell Challenge

Erik

I think that’s a great place to wrap. Gearing towards closing, anything upcoming for Arc that you'd like us to know about? Anything you want to tease? For people who want to learn more, what should they know about?

Patrick

So AlphaFold in many ways came out of a protein folding competition called “CASP,” [Critical Assessment of Structure Prediction]. And, you know, we created our own virtual cell challenge, at virtualcellchallenge.org, where we have, you know, hundred thousand dollar prizes, sponsored by NVIDIA and 10x Genomics and Ultima and others.

And it's an open competition that anyone can enter, where you can train perturbation prediction models, and we can openly and transparently assess these model capabilities, both today and in subsequent years follow them to get to that ChatGPT moment, right? And so I'm extremely excited about this. We'd like more people to, you know, train models and apply both bio, ML experts and engineers in any other domain.

I want this thing to exist in the world, you know, hopefully we're important parts of making that happen, but I'd just be happy that someone does it.

Erik

Yeah. That's an inspiring note to wrap on. Patrick, Jorge, thanks so much for the conversation.

Patrick

Thanks so much guys. Appreciate it.

Jorge

Thanks for having me.

Stay Updated

If you enjoyed the show, please share, follow, and leave us a review on your favorite podcast platform.

Find a16z on X: https://x.com/a16z

Find a16z on LinkedIn: https://www.linkedin.com/company/a16z

Listen to the a16z Podcast on Spotify and Apple Podcasts:

Follow our host: https://x.com/eriktorenberg

Please note that the content here is for informational purposes only; should NOT be taken as legal, business, tax, or investment advice or be used to evaluate any investment or security; and is not directed at any investors or potential investors in any a16z fund. a16z and its affiliates may maintain investments in the companies discussed. For more details please see a16z.com/disclosures.

Topics include fiery group chats and how dissent sharpens thinking, the sales playbook of modern politics, and concrete fixes for U.S. healthcare like ending PBM opacity, publishing real prices, and government-backed patient financing. Mark also explains how AI is pushing media from “social” to algorithmic, why he expects millions of models, and why ESOPs are an underrated wealth engine. He shares what he’d build today and weighs in on NBA economics under the new collective bargaining agreement.

Timecodes:

00:00:19 - Group chats & Mark’s politics

00:03:18 - Social media & political parties

00:18:45 - Healthcare & AI

00:33:53 - Mark’s business philosophy

00:36:03 - NBA

00:53:43 - Education

00:56:22 - Advice to Silicon Valley

Transcript:

Mark Cuban: I look at business as a sport and I just love to compete. I just like to be intellectually challenged. I'm an independent, I don't care about parties. I could care less. I want to be where different viewpoints are. An intellectual response rather than just a “you suck” response that you get on social media.

Erik Torenberg: Mark, thanks for coming to the podcast.

Mark Cuban: Thanks for having me on. Excited to be here.

00:00:19 - Group chats & Mark’s politics

Erik Torenberg: So we’ve spent a lot of time the last couple years or so, most of that time digitally, and on group chats. And one question that I often get in the group chat sphere is “How do these billionaires have so much time to be on group chats?”

To which I say, “Oh, that's actually the highest form, you know, Maslow's hierarchy of needs. Once you really make it, then you wanna argue with your friends and have interesting conversations.” What say you? What's your reflection on what you get out of these, or why you spend time on them?

Mark Cuban: You want to learn, you want to challenge yourself. You want to see what other people think and why. Because you know, in a world that's changing the way it is, you know, I want to be where different viewpoints are and have it be, you know, an intellectual response rather than just a “you suck” response that you get on social media.

Erik Torenberg: Yeah. And one of the things I appreciate most about one of our group chats is that you're one of the, the lone dissenters. Originally it was on the election, and, you know, it had more of a right-leaning crowd. And, you were one of the, you know, most prominent…

Mark Cuban: I was the dissenter. I'm not even the Democrat. I was just the dissenter.

Erik Torenberg: Yeah. And willing to sort of be independent. What did you get out of that? Did you just get sharper, get your arguments smarter?

Mark Cuban: It wasn't even so much that. To me there were a lot of things that didn't make sense. And to them it didn't make sense where I stood. And I wanted to get to the root of why people came to those conclusions. And you know, typically it just came down to trust. But you don't know that that's going to be the underpinning logic until you ask the questions. And so I think, you know, a lot of people presume, in any type of group chat, right, that you know where everybody stands and you think you understand why.

But it's until you start kind of challenging them. And then on top of it makes me, you know, reconfirm why I feel how I feel. And that's a win for me as well.

Erik Torenberg: Totally.

Mark Cuban: I'm an independent, I don't care about parties.

I could care less, you know.

Erik Torenberg: Have you ever supported a Republican?

Mark Cuban: Oh, yeah, yeah. I don't care. I've voted for Republicans. I just don't care, right. I try to look at each individual issue. I actually voted for George Bush. You know, we can argue about the quality, but it's about the alternatives, right?

And then like the first time I ever—well, I don't wanna go back that far. But in terms of the comparisons to Trump, you know, we're both salespeople. And I think that is a strength for me to understand him. Because at his heart, he just knows how to do PR and sales. That's who he is. And that's how I've grown my businesses.

You know, people always hear me talk about “Sales cures all,” and you have to understand your market and context, and he's brilliant at it. You know, he understands how to sell, he understands how to use PR, he understands where leverage comes and when it doesn't exist, and how to zig and zag when it doesn't work.

And you know, I think those are actual skills of people who have built businesses. Those are entrepreneurial skills. Now when you actually run the business, you have to do all the work behind it. And as a politician, that's not always the case.

00:03:18 - Social media & political parties

Erik Torenberg: When did you realize that businesspeople would have platforms that would give them kind of distribution advantages in reaching customers?

Mark Cuban: Just when social media came out.

Erik Torenberg: Because you were early to this. You were very early, like you and Trump, like one of the early people.

Mark Cuban: Yeah I mean, right when Twitter started. Not 2006, but 2009 and South by Southwest, when they were all pumping it up. When it was all about just “Where's the party at?” I mean, social media obviously, it's like, why are people following me? And it's just aspirational at some levels. And I own the Mavs, so it was informational. And then just with the technology background, it was informational there as well, so I could reach into different domains and interact with people. And that creates a platform. And each platform has its own look and feel, and you just learn how to, you know, deal with it accordingly.

It's gonna be really interesting because of AI. AI, as you know, has multidimensional impact. And how you create content, how that content is absorbed. Effectively, we all have our own feed. No two people have the exact same feed. And you know what Trump has done well is learn how to flood the zone. So that no matter what that algorithm finds for you, he's gonna find a way to you, whether it's a positive or negative. And then you look at Kamala and Brat Summer and, you know, fell off a coconut tree. Brilliant. She had everything going for her. But the Democrats didn't have anybody who knew how to sell.

And so going back to your question, it's not so much reality stars or, you know, businesspeople with followings. It's people who know how to use algorithms and understand how they work. And now as AI is evolving very quickly, knowing how to use AI and what's the impact on social media?

Because if it takes two seconds to generate on Veo, and then pretty soon that becomes agentic, and pretty soon it's, okay, just videotape everything that I do and just create videos and just flood every single account that we create. And then all of a sudden do people find themselves saying, “I'm gonna use social media less because it's all AI-generated and I don't know what's real, there's nothing social about it”?

Or do they think, you know, talking babies and gorillas as videos is the way to communicate.

And so it's gonna be fascinating over the next three years because we don't know that part. And so we may find ourselves in a scenario where we diverge from the traditional sales person putting it out there, like Trump, and find our way to, “Okay, let's understand how we do this.”

It's just like polling. Polling is useless now. You're always behind. And any one social media comment can change how everybody feels in a nanosecond. And so, you know, how do people start integrating surveys?

You know, I'm just continuously surveying with all those questions we get online and what direction does that take us? How do I integrate that into video creation? I mean what's the limit?

Erik Torenberg: Yeah. So you experimented with a social network a few years ago, “Dust,” it was called.

When you look out, you know, 5 years from now, 10 years from now, do you think there's a new platform that emerges? Or do you think it's kind of this like fragmented ecosystem that we have now? What is your prediction for the future of social?

Mark Cuban: Well, I mean, look at Bluesky.

They, you know, had been around for a little while, but then all of a sudden when Trump got elected, their usage just boomed.

Erik Torenberg: I call it “the fragmentation” of like, you know, people didn't like Elon now. It’s not just that but…

Mark Cuban: But yeah, so they moved, right. But now it became its own silo. And so I think it won't be social media, but there'll be media. You know, it depends on how much you believe in AI and where you think it'll go. I always look at it as creative people, they'll be able to amplify their creativity and iterate incessantly.

And that will create some amazing, incredible things. Are you just gonna put it on X? Are you gonna put it on LinkedIn? Are you gonna put it on Bluesky? Are you gonna put it on TikTok? You don't have control of the algorithm. You know, and so I think there'll be something new. And I think how people connect to that content will be the foundation of it. Because I think we're tired of rage bait.

And we're tired of the fact that, you know, the algorithm presents what you seek, effectively. I think instead we are going to get unique offerings that don't seem to make sense right now. And I think part two to that is, you know, like if you go on X and you use Grok, it almost always, unless it's something that Elon is individually interested in, it comes up with a legitimate answer. And you get people who say, “Well, you know, hey Grok, tell me…” and it doesn't give them what they expect, you know? And I think that's a huge positive.

And it's the same on other platforms as well, where you can just go to ChatGPT, Perplexity, whatever you use, or all of them, and ask questions. So you'll see a reduction of misinformation, I hope. And with the reduction of misinformation, you're gonna have less rage bait.

And with less rage bait. What's the connection? And hopefully it's better quality content. Hopefully it's, you know, I don't know, it's just like if we were sitting, just gonna brainstorm, and just say, you know, “Can we create our own new network,” right? What would it include and what would the parameters be?

Erik Torenberg: And when you did it, maybe 5 years ago, maybe more, was that like a Snap competitor? I'm trying to remember.

Mark Cuban: It was 10 years ago, 12 years ago. And it was more because of what happened with the SEC, where I got charged by the SEC for insider trading, and they basically just took anything and everything I said out of context. So I gave them everything I had, and I got cleared in like 15 minutes. But the point was I wanted something that was truly private.

And so, and it still exists today. But I just use it for mostly internal communications with our employees.

Erik Torenberg: Yeah, that makes sense.

Let's go back to the messaging on Democrats, Republicans. If you were in charge of the DNC, the Democratic Party and could control both the platform and the messaging, what would you advise?

Mark Cuban: Well, first you gotta have people that know how to sell.

And second, you have to know what people wanna buy. And people wanna buy, quote unquote, just a better life.

You know, what the Democrats do is they project, they extrapolate. “Trump did this, so it's the end of the world. “

The Republicans are just like, “What's the price of tea in China?” right? “Because I drink tea from China.” And they do deal directly with what's happening with you today. And 99% of people in this country, that's what they care about first. And so I would go to the Democrats and say, “Look at the price of beef.” And I think that's what the Democrats miss.

They're so intent on the “T” word, right? Everything is just a trigger word because they want people yelling and screaming at rallies, and that's great. Just like Trump did, if you have 10,000, 15,000 people with Bernie and AOC and you wanna get them all riled up, it's a pep rally, right. Say what you gotta say for the pep rally.

But if you want to get people to pay attention, I'll give you an example. I was in Indiana. I have a company, a Shark Tank company, Guardian Bikes, and we moved 80% of their manufacturing to Seymour, Indiana. So they invited me in, and we celebrated. And they've grown from, you know $200,000 at Shark Tank, they'll do $120 million now. And so we got into a little bit, and then we had a question and answer and somebody stood up, “I saw on the news,” and they started getting to the question, and I just blurted out, “I don't watch the news.” And the place went nuts. It just erupted, right? Like, of all the things we talked about for an hour and a half, that by far got a 10x response.

That said volumes to me in terms of where people's heads are at. They don't wanna deal with all that shit anymore. They don't want to hear about the end of democracy. They don't want to hear that Trump is an ultimate cancer.

Put aside whether or not you believe it's true. That's not the goal. The Democrats always lose track of the goal. They lose sight of what the end game is. If you want to replace Trump, if you want to win the midterms, it's not about ripping on Trump. That's been said for the last 10 years. But how are you changing people's lives?

And then I got into it with somebody else on the Democrat side via email, and it was about, you know, “Oligarchs and billionaires. Go get 'em.” I'm like, “Fine, I'll pay more taxes. I don't care.” Right. But, do you think, other than the 942 billionaires in this country, right, other than getting people riled up, how are you changing anybody's life? Right? People wanna know that you're impacting their life.

They want to know that things are gonna be better. We hear about housing, we hear about, you know, should they go to college or not, the threat of AI in their mind. All these different things that they truly worry about and talk about at the dinner table. How often do the Democrats talk about it? And then it's like, “Well, let's go get the billionaires.” And I'm like, okay, “What's the goal? Do you want to reduce income inequality?” Then go to the billionaires and give them incentives, which really made them mad, right. Companies’ incentives, not the individual 942 to, you know, your taxes will go instead of 27%, let's say, you'll stay at 21% if everybody gets equity in the company.

And it's on a, “pari passu” is not the right word, but it's on an equal percentage of take home pay, that you get relative to the CEO down to whoever, right? Now you have appreciable assets. Yeah. Yeah. When people have appreciable assets, you know, beyond just a home, A, if it appreciates they can buy a home, and B, the wealth gap doesn't increase. You know, and it gets closed down. You know, if you do a little research, because I was just curious about all this, and I went to Perplexity and I'm like, “What's the earnings of people who are in ESOPs, employee stock ownership plans, versus everybody else?” It's dramatically higher. It's not even close. They make more than union employees. And there's less turnover.

And so there's a lot of good reasons, and you don't have to be in power as a Democrat to do those things, to go to the local employer and say, you know, “We're gonna protest you unless you create an ESOP plan that everybody gets to participate in at the same level as a percentage of earnings as the CEO.”

Erik Torenberg: I think that's the best form of UBI, which is instead of free money, just upside in our S&P 500, upside in the economy.

Mark Cuban: And there's places for UBI too, right? So if you're a caretaker, you're taking care of your parents, I got no problem if we do that. Because you're paying for it one way or the other, you know, through Medicare, through Medicaid, some way. But to your point, right, appreciable assets change your life.

And if you don't have any, there's no way you can keep up. You know? Because the stock market is going up.

Erik Torenberg: Yeah. It seems like the Mamdani, sort of like socially progressive, economically populist, sort of combination, AOC in a different way, you know, Bernie, he was the proto-version. Is that the future of the Democratic party?

Mark Cuban: He learned from Trump. He's the progressive Trump. Just lie.

You know, “Day one groceries are going to be cheaper. I'll open the grocery store because I’ll sound more progressive that way.” Right? “Day one, buses are free.”

And then everything else, he kind of mealy-mouthed back to the middle, you know, and kind of gives a little bit on the edge for the progressive to be happy. But he's smart. Starts saying, “I'm reducing grocery prices. Trump is increasing, I'm reducing them day one,” right. “When we have the House back, we are going to reduce grocery prices.” “How’re you gonna do it?” “We're going to do it. Trust me.” Because that's the bifurcation between Trump fans and Trump haters.

Erik Torenberg: But in terms of substance, do you think the future Democratic party is more like a Bernie Sanders economic populism? Or do you think it's more…

Mark Cuban: I don't know. I think whatever gets results. I give Bernie credit because he was the first to talking about a living wage. And it took him decades. But we've gotten there. And he kept on pushing through. So he deserves a lot of credit. But it more plays to his base than anything else. And you know Elizabeth Warren had the CFPB right? That's okay, right? You can argue both sides, but she got something done. I don't know what the Democrats are doing that's getting shit done. And that's the change in their focus, and that's what's going to guide what type of candidate. Because you gotta be able to sell, gotta understand technology, and gotta be able to implement technology. And I said this to Jason Calacanis 15 years ago, maybe. I'm like, “Look, if we can improve efficiency and reduce the cost of government, you can increase the money you give to people. Take half the savings, here's a bigger check.” That's what the Democrats could be doing. We wanna give you more money.

That's what the Republicans do great. “We're making a promise.” Democrats, “We're gonna use AI and if we need to give it UBI, if we need to tax robots, we're gonna tax robots.” You know the whole Bill Gates thing. You know, a quarter dollar an hour, whatever. “And then we're gonna take the benefit of that and write you a check. That's gonna be your UBI or your tax credit or whatever.” They just don't think like that.

They don't think, “Where is the solution? How do I get there? How do I sell it in a form that's easy to digest so people understand it?” They don't get, AI is like going to change everything. You know, how can we not use this in government to be more effective?

I mean, that's the whole point, isn't it?

Erik Torenberg: Let's go deeper there because I know AI is one of the things you wanted to chat about, something you spend a lot of time thinking about. When you look out at the next few years, what do you think is underappreciated or under-realized about how AI is going to change a specific category or a specific way of life?

Mark Cuban: I think it democratizes a lot of things, particularly education, while I understand the fear, and the first pass is never the way it ends up.

You know, I got two kids in college now and one in high school. Hell yeah, they cheat with it. You know, they do. But then I started talking to this teacher from Pennsylvania, I forget which town in Pennsylvania. She just emailed me, and she was like, “What can I do to integrate AI into my classroom where it's not just them giving me answers to questions.” And I said, “Well, you know, make it more like Jeopardy.” And then she's like, “Use ChatGPT to find the four causes of the American Revolution. And write up what this says. And then we're gonna discuss in class, and you pick, which one do you think is the most impactful?” And then we'll discuss that in class. Just changing the paradigm of how they would do things.

And so, but back to your question, I think democratization of education is key because a kid with a cell phone and internet access can go to, you know, any large language model and ask anything.

And ask it, you know, “Hey, I'm eight years old and I wanna learn to speak Polish because my grandmother's from Poland. Would you put together a class for me and teach me how to speak Polish?”

It will. How else are they gonna do that? You know, sign up for Duolingo? Right? You know, “I'm really curious about baseball. Who are the best baseball players and why? And teach me about the statistics behind it.”

00:18:45 - Healthcare & AI

Erik Torenberg: Yeah. And always-on tutor. Yeah, absolutely. How about in healthcare? It's an area you've spent a ton of time in, with Cost Plus Drugs. Right now, you know, healthcare is like 20% or something of the economy. In the future, is that number going higher? Is that number going lower? You know, talk about what's gonna happen there.

Mark Cuban: It'll go higher, but because we're gonna get some amazing treatments.

Erik Torenberg: Oh, interesting. So the product will get better, the service will get better.

Mark Cuban: Right. Because now you get, you know, single treatment cures that might cost $3 million. That jacks up the total cost.

But in terms of how we do healthcare, everything I do, I put into a ChatGPT project first, you know, and share it with the doctor. And then I ask my doctor, “Do you use this stuff or OpenEvidence?” Like, “Hell yeah, we do it.” So I think healthcare will get better qualitatively. I think we'll benchmark ourselves a lot more and that'll make us healthier in a lot of ways. I think where the challenge and the hard part is, other than the obvious, you know, optimization and processes and hospitals and all that, the challenge is how we value IP.

Because if I'm MD Anderson, if I'm Stanford, any research hospital, any scientist in healthcare, you're an idiot if you publish it. You're an idiot if you patent it because it's immediately going to get absorbed into a large language model.

And you've lost control of it. And you've lost ownership of it. And I think what's going to happen is, you know, and I've said this as I've talked to CEOs of research, hospitals, etc, you need to silo your stuff. And you need to take what you've siloed and either do your own model or put it out to bid. And let the foundational models compete because they're in a death war.

Erik Torenberg: And they need specialized data.

Mark Cuban: They need all the data they can get, specialized or general. Because everything that's accessible on the internet, they're getting or got.

And so the people who are creating new things, whether healthcare or anything for that matter, are gonna start siloing it and putting it out to bid. And I think even within the government, right? So we see all these research grants getting pulled.

I told somebody I know, you know, in the administration, I'm like. “No, keep on doing it, but put it out to bid for the models to get access to it.” Back to your question again, it will change because there'll be millions of models as opposed to just singular foundational models.

I've talked to folks at Microsoft, Meta, and other places, I think they disagree with that. You know, that they think all this information is going to be available so that the hospital system or the university system, whatever it is, or the grantors will just say, “Go ahead and take it.” Hell no.

Erik Torenberg: So you think there'll be a flourishing of startups that will have kind of access to specialized data and either put it to bid or…

Mark Cuban: Or just like MD Anderson will have their own large language model. And you'll just have to know where to go. And just like right now, you know, ChatGPT is not gonna give you the same answer as Perplexity, as Claude, etc.

You're gonna go and use multiple of them, then you'll say, “Here's the input, the output I got here, here, here.” And you'll put it in somewhere else to assemble it. But I think there'll be millions of models. And there'll be front ends for those models that can't go out and do a million of them but that help curate what's best for what you're trying to accomplish.

Erik Torenberg: If you could wave a wand and change anything about how we do healthcare in this country or where we can learn from other countries, what would you change?

Mark Cuban: You'd get rid of the insurance companies and the PBMs because they do all they can to not take risk. And they do all they can to introduce opacity to the system. Particularly on the drug side, if you look at all the other countries who’re trying to do MFNs to mesh their pricing, they don't have PBMs. You know, when you look at the internals, the biggest PBMs cause all the problems.

Like literally, we get emails all the time, people who are on Medicare Part D for drugs come to Cost Plus Drugs because we're cheaper than their copays. That makes no sense. But it's the federal government that approves those plans. And those plans in Medicare Part D are run by individual insurance companies.

Then if you look on a bigger platform just for healthcare in general, you know, think about how healthcare works here, at a16z. Do you have an insurance company? Do you guys self-insure? Do you know?

Erik Torenberg: I just joined so I'm…

Mark Cuban: Oh yeah. Okay. So, you know, let's just pretend you self-insure, you’re one of the bigger companies. So an insurance company comes to you knowing you self-insure. And they put together a list of plans.

And then those plans, no one ever asked, “Why were those plans defined the way they were?” They just assume, “We just get our choice of plans and we pick them.” Well, there's an inverse relationship between premium and deductible.

They never do credit checks on the deductibles, do they? And younger people, they're taking the higher deductibles because they presume they're gonna be healthier. But the people who are making the determination that they're using the higher deductible plan, half of them don't have enough money for the deductible, right? So we've got an insurance system driven by these insurance companies, even if the risk is absorbed by companies or partially by the taxpayers, right, that are designing plans that they know the people who are buying the plans can't afford to utilize.

And what happens then? That turns the hospital into a subprime lender. Because now they go in either through the emergency room or they show up not knowing what their deductible is, and then the hospital will loan them money because they want them to be able to pay the deductible. So then, they can then charge the insurance company or the employer if they're self-insured. But wait, there's more, right? So then let's just say someone went in and gets a hip replacement, pays their $5,000 deductible, and they owe them $10,000, right? The insurance company, even though they have a contract that says it's $15,000 between the provider and the payer, the insurance company, they don't just write the check for $10,000, right?

They say, “We'll give you 7,000. You don't like it, come and get us.” And so when you talk to a hospital CFO, they're like, “We lose 3-5%. We can lose as much as 50% on the subprime lending.” You know, and so they have to invent all these costs to compensate for what's happening on the other side, on the payment process.

Then of course you have the pre-authorizations, and most of the pre-authorizations get approved after everything is said and done. But they get the time value of the money of the premium and not having to pay it. And they've eaten up all these doctors' time, and the doctors are the ones getting underpaid.

So you're going in for a heart transplant, and you know, the doctor might make $10,000 unless it's a big name surgeon. I want my doctor making a shit ton of money, to make sure that you know… So you have this connectivity between all the pieces that really hasn't been thought out.

And those insurance companies either own the PBMs or are owned by the PBMs. And what they do, because they're together and they also are vertically integrated to own clinics and other providers. So they have all these intercompany transfers that allow them to gain the system.

So with the ACA, you, depending on the circumstances, the insurer has to pay out 80 to 85%. called the medical loss ratio. And so you would think that's just straightforward. But if you're vertically integrated, you can put some over here, put all the profits into your PBM and show that you've spent all this money because you wrote that check to the PBM. And so that's what happens. One of the biggest insurance companies had $161 billion off intercompany transfers. Intercompany transfers! That's 0.3% of our GDP.

Erik Torenberg: Wow. Just moving money around to make more money.

Mark Cuban: Just moving money around to make more money.

Erik Torenberg: So if you'd get rid of them, who should take the risk? Would you replace anything?

Mark Cuban: So it's a great question, right? So healthcare is easy, right? Presuming, you know, if you get the right doctor and everything, you go to the doctor and they tell you what's right or wrong, right?

Go home. Wrong? Okay, what do I need? Then once you know what you need, there's only two questions, whether it's medication, surgery, whatever it is: how much does it cost and how do you pay for it? Costplusdrugs.com is still the only pharmacy company that publishes an entire price list. The only one, and we've been in business three and a half years.

That in and of itself is crazy. But then once you get into the system, right, and you don't know what anything costs, then they can charge you whatever. And then in terms of how do you pay for it, that's the who takes the risk, right? And so deductibles are going up, no matter whether it's Medicare Part D, even Medicare Part A and B have higher deductibles.

Medicare Advantage, they shrink the network and create networks. So out of pocket is going up. So patients are the first line of defense that the insurance companies are trying to push the risk on.

But better in my mind is to combine kind of a market where you get rid of the insurance companies, there's no premiums, get rid of the big PBMs, there's no, you know, manipulation, if you will, of the funds, with intercompany transfers. You create tools that allow patients or their employers working with them to go out there and find the best solution for them and to shop. And hopefully you've saved a lot of money from not paying premiums and you can pay for what you need, but what you can’t pay, the government will loan you that money if it stays within the price parameters that have been set. So it's negotiating without negotiating. And they've done that to a certain extent, reference-based pricing with Medicare, you know, and Medicaid, for Medicaid patients. And so now if you or I can pay for it, we just pay for it, but we've saved all that money and not paying premiums.

And slowly but surely we're getting more and more reinsurance options for individuals. So for catastrophic insurance. So, you know, in case, God forbid the worst thing happens, you have an option. So that's one place of risk. But if you're not paying premiums to an insurance company, you can pay for your reinsurance.

And then if your hip replacement is 15,000 and you have 3000, you can use financing, but it's guaranteed by the government. Kind of like we do with college. You know, you can use the financing the school has, but it's guaranteed by the government, kind of like we do with houses. You can use the mortgage plan, but it's guaranteed by the government.

The most important thing that we have available to us, our healthcare, the government is like, “Whatever.” So if you take the risk and allocate it within a price control—“price control” is the wrong word. But if you're a patient, you're shopping for care, and it stays within the benchmarks that are set, then go for it. If you wanna spend more, you think the doctor doesn't participate, and you want to get better, go for it. But then all of a sudden you've got all this money that's not being paid in premiums that's available to people to use to save, whatever. That's how I would look at it.

Erik Torenberg:

So you've built Cost Plus Drugs partially because you care about it, obviously. You have this distribution advantage, you have the capital advantage to get it started. Let's imagine if you were a young entrepreneur, you know, a mid-20-something Mark Cuban in 2025 and trying to, you know, be immensely successful like you've been, how would you think about that? Or what would you pursue?

Mark Cuban: For Cost Plus Drugs or in general?

Erik Torenberg: No, just you as an entrepreneur.

Mark Cuban: Oh, like if I was just getting outta college right now? Oh, I'd be all AI every day because I think the disconnect, I think employment for computer science graduates, and I said this 10 years ago, I got crushed, 2017 actually.

I said, you know, “AI is gonna put programmers outta business,” beginning stage programmers. And people said, “Oh, you're an idiot.” But now big companies are gonna find ways to use AI. So they're gonna reduce, you saw Salesforce or whoever laid off people. Small companies have no clue. You know, and there are only 22,000 companies, I remember this from the Kamala campaign, 22,000 companies with 500 or more employees.

30 million right, now, a lot of them are single entrepreneurs, but, you know, 5 million have 500 or less. That's where you want to go try to get a job. Because if you're AI-native because you've been using it through your four years of college, or two years from now having had four years experience, they need you.

You know, maybe they have one person who's picked it up, two people.

But if you have a business background, which you can learn in school and you've become an AI native, you can walk into a small company and say, “I can help you. I can create agents for you that'll go out and look for things.” One of my Shark Tank companies is Rebel Cheese, and they added a direct-to-consumer business. And they created an agent that compares the weight that comes off the scale or the box type and what they were invoiced by UPS or FedEx.

And compared to the price list, always got overcharged. Always. This little company is saving, you know, thousands of dollars a week just by using an agent, you know, and it's not anything difficult or you know, advanced. So being able to walk into a company and say, “I can do this, I can do this, I can do this.” It's no different than what systems integrators have done for decades. You know, that's what I did in my first company, MicroSolutions. “You don't understand PCs, you're a small company. I do. I know networks. I can program, I'm gonna write you an application that emulates what you're doing, but faster.” Now it's no longer about emulation. Those days are gone. It's about how do we change the process by removing pieces to make you more efficient, more productive, more profitable.

Erik Torenberg: Yeah. So you wrote the PC wave then you wrote the internet wave with broadband, you know, had a big outcome there.

If you were starting a company today, would you try to build a foundation model company, or vertical application or?

Mark Cuban: No, neither.

Erik Torenberg: Yeah, what would you do?

Mark Cuban: Because they're all gonna be customized. I literally would hone my business skills as much as possible. Domain novel knowledge combined with understanding of AI is the win.

Erik Torenberg: Right. So you would go to some sector that hasn't yet been digitized by AI and…

Mark Cuban: Or just small companies, you know, regardless of sector.

Erik Torenberg: Yeah. But if you wanted to start something that could be big one day.

Mark Cuban: Right because if I work with one shoe company, and I analyze what they do, then I say, “Okay, I'm gonna charge you X per transaction. And I'm gonna do the agent for you, and I'm gonna monitor the agents.” It's SaaS at one level. But it's customized SaaS. And as AI advances the model advances, and the agent advances. And so, you know, factories, you know, talking now, like at Cost Plus Drugs, we have a robotics factory, and we use AI agents to monitor everything. And our cost to run the factory is a fraction of pretty much everybody other than India and China. And we'll catch up to them. And so understanding how business works, having that domain knowledge, and then being able to apply it, and then being able to scale it by taking it to all those companies in similar circumstances, or creating agents that modify the agents based off of the information you're getting from all your customer companies. That's the new SaaS, potentially.

00:33:53 - Mark’s business philosophy

Erik Torenberg: Yeah. Totally. I want to zoom out and talk just general business success and advice. The internet lists your net worth, I believe, at over 8 billion, or something to that effect. Typically when people become super wealthy, it's because of like one big thing.

But it seems like you've sort of compounded wealth in decades. And I think the biggest liquidity events, maybe I'm wrong, are the company you sold and then the Mavericks. And then there've been other efforts. What do you think is the thread, the through line when you look at your career and the ways in which you've made money that others can learn from?

Mark Cuban: I look at business as a sport, right? And I just love to compete. You know, and that keeps me going. Just, you know, there's always something new and always something changing. And it's almost like the forums, right? I just like to be intellectually challenged.

When I first started, I was always the youngest guy walking in the room. Now I'm the oldest guy walking in the room. And I like kicking everybody's ass where I can, right? You got your CS degree from Harvard? Fuck you, I don't care, right? Sounds like Glengarry Glen Ross. But I like that challenge. And I understand patterns and I understand being able to create new things and I think that’s served me really well. It's like Cost Plus Drugs. You know, we launched in January of 2022, we're serving millions of customers. We're changing people's lives. And everybody thought I was an idiot, you know? And we're working outside the system. Amazon, everybody else, “Okay. We're gonna work with the big PBMs because that's where the money is.” And I'm like, “No, they're the problem. You don't work with the problem.” You know, it's the innovator's dilemma. You know, we're gonna take what they can't do and you know, when you run with the elephants, there's the quick and the dead. We gotta be quick and aggressive. And so I think that's a big part of my skillset. You know, whether it's starting Audionet, which was the first streaming company, HDNet, the first all high-def network, you know, Cost Plus Drugs, what I do with the Mavs in terms of bringing in technology, it was always about, you know, Steve Jobs said it best. Everything's a remix.

You know, it's about taking what I know and just learning whatever's coming up and what new technologies and remixing it to apply to something different.

Erik Torenberg: When you bought the Mavs, did you have a sense that 20 years later it could be, you know, worth the order of magnitude more?

Mark Cuban: No.

Erik Torenberg: It wasn’t a business decision?

00:36:03 - NBA

Mark Cuban: No, no, no. In fact, like I paid 285 million for the Mavs in 2000. And 2010, the NBA couldn't even sell the New Orleans franchise. There were no buyers. And the valuation had gone down. Not up.

Erik Torenberg: So you thought it was a bad purchase from a financial perspective?

Mark Cuban: From a financial perspective, it was awful, right? Plus I was losing money every year, so it had nothing to do with money. But I loved it, right? That was the key part.

Erik Torenberg: And I think it elevated your profile in a way that probably helped your other business.

Mark Cuban: Oh, for sure. And then that got me to Shark Tank, and the rest is history. But you know, I just, the way I looked at it back then is I'm having fun. I'm blessed, right? I'm gonna make the use of my time, so I enjoy every minute of it. But what changed the economics of the NBA was, you know, linear media being competitive. You know, and that goes back, you know, things I look for, I always look for death wars, where you've got competitors, that one's gonna win and one's gonna lose. And they have to go all in.

Just like with AI right now, they're spending tens of billions of dollars a year. They're making obscene amount of cash flow and they're spending it all. They're paying a billion dollars to people to come over. What does that tell you? That anything you can sell to them that really differentiates them, they're going to buy. And so, but you know, with the Mavs, I didn't see it as an investment, but it wasn't until the death wars hit linear television and then all the TV deals skyrocketed.

Erik Torenberg: So that's what made the difference between 2025 and 2020 and 2010 in terms of why the NBA teams have gotten some much value, more valuable is TV.

Mark Cuban: 100%. Yeah. A little bit of real estate. So you can build around. You know, but, it's 100 percent streaming/TV. The question becomes, you know, how long does this sustain? You know, so like Major League baseball didn't get a deal done with ESPN and saw their rights start to decline. But it was right around the time, I think it was Peacock who bought just one NFL playoff game and got like a hundred thousand new subscribers or a hundred million new subscribers, something ridiculous, right? That all of a sudden set a benchmark. So this season, right, starting with the NBA season, how well that does in terms of driving subscriptions and reducing churn, that's gonna make or break everything that comes forward for all sports.

Erik Torenberg: Yeah. I also wonder if, you know, one thesis I have is that value capture will reflect value creation when the market gets more efficient. And NBA players, it seems like, don't have equity in the league. And I wonder if, in the future, they're gonna band together and just say…

Mark Cuban: I don't think you can, right, because Father Time's undefeated. You know, so at what point do you stop? And the price of the franchises now are so high. I mean, how much can you give them? I'm not gonna say it's never going to happen. But, you know, I think it could be a question. But the other point is that there's only like two players that fill an arena: Steph and LeBron. Like when Luka played for the Mavs—and I had nothing to do with the trade—when Luka played for the Mavs, we didn’t play to sold-out buildings. He would draw fans, don't get me wrong. But not every building was sold out. When Giannis would come to Dallas didn't necessarily mean we were gonna sell out. But the whole concept of what drives attendance has changed dramatically and what drives viewership on television in particular. Because now, fans are fans of the players more than they are the teams. Like when we grew up, it was like you were, I grew up in Pittsburgh. So you're a Steelers fan. You're a Pirates fan.

Erik Torenberg: That's why I wonder if they're gonna team up and say, “Hey, you know,” Wembanyama, the new generation, you know, “We could start a new league and get…” Cooper Flagg.

Mark Cuban: Cooper Flagg.

Erik Torenberg: Yeah, exactly.

Mark Cuban: It's expensive, to start a new league.

Erik Torenberg: There's a lot of infrastructure. Yeah, yeah, yeah. NBA is its own…

Mark Cuban: And even with Caitlin Clark, I don't know all the WNBA numbers now, but most of them aren't making money. Even in the NBA, most teams aren't making money before revenue share.

Erik Torenberg: Right. I know that you were dismayed at the Luka trade. We were sort of in the group chat when it happened. But I just wanna reflect. I was so devastated by that it seemed so unfair for the Mavs not to get picks back, or Austin Reaves or whatever, I don’t want to relay it. I was complaining to my girlfriend at the time, you know, like many people were. And I even remember reading this conspiracy theory that it was on purpose to tank the value of the franchise to then move to, you know, Las Vegas. And I was like, “Oh, this is the only way it can make sense.” How did we not get more for Luka in this trade?

Mark Cuban: I wish I knew. I wish I knew.

Erik Torenberg: But, then the Cooper Flagg thing, I mean…

Mark Cuban: The basketball gods were looking kindly upon us.

Erik Torenberg: Yeah. The world works in mysterious ways.

So I have a dream to be a part owner someday as well. And I've wondered if the strategy is either to compete for a championship or do the Oklahoma City Thunder where you trade for picks. What do you think?

Mark Cuban: So, I think the Thunder methodology is better now, with the new CBA. Every new CBA creates transitional issues, where your decisions were made in a legacy environment, and they just changed all the rules. And now it’s even more difficult with the second apron because the problems of being over the second apron are far more draconian than anything we've ever seen. And so I think the patience approach that San Antonio and OKC have taken is going to lead the way. Whereas with me, I was never patient. And that worked against me a lot too. You know, I wanted to win now, win now. We have Dirk. Let's get there. Right.

Erik Torenberg: What’s something you would've done differently?

Mark Cuban: Kept Steve Nash.

Erik Torenberg: Drafted Giannis.

Mark Cuban: Drafted Giannis, yeah. I think I just would've been more patient and waited guys out to see how they turn out rather than, “Okay, I'm gonna throw in a draft.” Because at the time you could say, if you didn't have a pick from below 15, the odds of them being a player was minimal. And so if I was getting somebody I could already evaluate, that worked in my favor. But now draft picks have become so much more valuable. And so that's changed the valuation approach.

Erik Torenberg: I heard one argument that I was somewhat sympathetic too, which is, the argument for how you can understand the Luka trade is, you know, obviously legendary, iconic player, but so expensive. And with the way that the contracts were structured is, if you gave, you know, kept giving him the supermax, how much room do you have to build a team? It's kinda like the Trae Young problem, like some of these players are so expensive, but are you really gonna win a championship if you don't have…

Mark Cuban: It's like, can you have two max players? And that's the challenge. You know, in the past, you could reward somebody with a max contract knowing you had some wiggle room. And more importantly, you can always find a way to get off that contract.

Mark Cuban: Not anymore. If you have your one generational player, you pay 'em whatever it takes. Because of the max contracts, they by definition are undervalued.

So, the generational player, you always pay whatever it takes. It's the next person. If they're not a generational player, how do you deal with that? That's the hard part. And when you see, like you saw with certain players, that teams couldn't get off of them. Either because of age or, you know, they were max players by the rules of the game 8 years ago, 4 years ago. But they're not any longer. That's where the challenges happen.

Erik Torenberg: Right. If you were making decisions for the Mavs now, given where everything is what would you be thinking about? What are the big questions?

Mark Cuban: Develop Cooper, first and foremost. You know, we’ve got Anthony Davis, we've got Kyrie who will be coming back, Klay, just having those Hall of Famers there.You couldn't ask for anything better in terms of development support. And so I think we have to focus on staying healthy, getting Kyrie back.

Erik Torenberg: It's an interesting thing with two timelines, you know? The Warriors tried to do it, but they couldn't exactly pull it off, although they've done great, but these sort of, you know, older players and then the Coopers.

Mark Cuban: Like where the Warriors are now. You know, Father Time is undefeated, and you're trying to stay committed to the guy who's got you there, you know. And the players aren't dumb. So they know they only have, you know, a finite number of years left. And so they're trying to maximize their earnings during that period. It's just hard. And so, you know, the whole second apron restrictions really just kicked in last year. And so now teams are having to deal with it. And so you’re seeing players get stretched, you know, not just waived but stretched. And I think the strategy is changing right in front of our eyes. I think there'll be fewer max players coming up

Erik Torenberg: Interesting. And it used to be this thing where in order to win a championship, you had to have one of the best players in the league. You know, the Pistons were an exception to that. I'm a Knicks fan. I wonder if the Knicks have a, you know, possibility to question that. But if you're playing for a championship, do you think that you have to have, you know, one of the five best players in the league or?

Mark Cuban: Sure helps. It depends on the skillset. So what's changed a lot in just 6 years, 7 years, certainly, you know, the last 10 years, is there used to always be one guy on the court who couldn't shoot. That couldn't get a bucket. They were there for defensive purposes or they protected the rim.

Now they all can score offensively at some level. And so, you know, whether it's the Knicks or anybody else, if you have enough guys who can put pressure on the defense. And that's why, you know, the Knicks made the deals that they did. But if there's one guy that you're playing against that's always the best player on the floor, and you don't have that player. It's going to be tough.

Erik Torenberg: Totally. Well, I think what the Knicks are asking, I guess Boston's asking too, is like, “Can offense championship?” Like if you just stack the…

Mark Cuban: Well, yeah. I mean the Knicks have got some defenders.

Erik Torenberg: But Brunson and KAT are weak defenders.

Mark Cuban: It's gonna be tough. It makes it tough because, once you get to, like even against Indiana, you know, Indiana didn't have—Hali, we tried to get Hali so hard, right. Hali is not a top 5 NBA player, right? He's good. Maybe top 15. Yeah. But he was better right in that moment because you can do matchups over and over and over again across the up to seven games. You can play certain players off the court. So you saw KAT, not even being out there at certain times.

Erik Torenberg: Totally. Yeah no, if the Pacers won the championship, that would've really questioned, you know…

Mark Cuban: And they took them to 7 games. I know. And their best player gets hurt, you know, and we the Mavs, injured and all without Luka beat OKC 3 out of 4 games.

Erik Torenberg: By the way, Dallas is one of the deepest teams right now too.

Mark Cuban: Oh, we're really deep if we can stay healthy.

Erik Torenberg: But yeah, there's so many teams like the Raptors, the Bulls, the Kings, kind of like in perpetual, what I would call mediocrity. And they're expensive, and they're old, and I'm like, you just gotta rebuild. Just admit that you’re rebuilding. And there's a number of teams like that.

Mark Cuban: Yeah. And part of the problem is that there's a salary cap floor. You have to spend up to a certain amount. That's like, because they're trying to keep

Erik Torenberg: Because they’re trying to keep it competitive, I guess.

Mark Cuban: Well, it's just, that's the deal we do with the players. They want more money spent. And so 90% of the salary cap is a lot of money. Right, so it's like $153 million, right? 170 give or take times 9. And so, spending up to $153 million, that's 10 million average. So that means you're overpaying a lot of people. I might have the cap number wrong, but in any event, you're paying a lot of people a lot of money. But if you keep them short contracts Right. You'll get guys wanting to come just to get overpaid. You know, and you almost have to tank and you just…

Erik Torenberg: I have a theory that there's these guys, like, you know, maybe I'm being unfair, but Kuzma, RJ Barrett, Brandon Ingram, who are sort of like built for losing, for getting a lot of money and a lot of points on a losing team. And we're seeing more of that.

Mark Cuban: They're freer on a losing team. And they do it to get paid more money. But teams are smarter than that. But at the same time, there are teams that need scoring. And, you know, if you fall into that category, you're gonna go after those guys if you have the space for it.

But I think what's happening actually now, we're seeing all the salaries kind of revolving around the mid-level exception, which is $14 million, give or take. And so if your guys now realize with the second apron things have changed, so if you get paid more than mid-level, you're a good player. If you get paid at the mid-level, you're a decent player. If you get paid below, you're roster piece, unless you're young and you're on your way up.

Erik Torenberg: For advice, for me as a future owner, what are the biggest ways an owner can influence winning? Is it sort of personnel selection on coaching GM?

Mark Cuban: Always looking for an edge. Yeah, always looking for the edge. You know, personnel selection is hard. It's more art than science. You know, like I told Jalen Brunson when I did his podcast, he was just some chubby guy, right?

Erik Torenberg: I’m sorry we took him from you. Or, you gave him away.

Mark Cuban: I was not happy about that. But, you know, we picked him 33rd. 32, you know, chances to draft him, and they didn't. And we liked just his pedigree of winning at Villanova. But it's his effort. Like when he was with us, his go-to move was either a spin or a move to the left and he'd fall away doing a layup. And he never got to foul because he was always falling away. Now his mid-range jumper is money. But he worked on that. And that's what you don't know when you draft somebody. What mental capacity do they have? You know their athletic ability. But will they work hard to improve their skills? Will they work hard to improve their basketball IQ? If the answer is “yes,” and they have the athletic side of it, you've got a chance, but you don't really know that. Because no one comes in and works out and says, “Okay, I'm an idiot. I suck,” right, and shows you they suck. You know, they've all been trained on how to work out. And they're in college maybe one year.

So, you know, and then if you're going to a college, and they're getting paid now, right, it might be that they're just doing what the coach needs them to do there which may not match what they need to do in the NBA. The hardest part for players that aren't superstars coming right in is you've gotta learn how to fit in. Even with Luka, we didn't know he was gonna be a superstar. We thought he was gonna be really, really good. And we thought he had superstar mentality. But I was terrified because I made a trade. I gave up a pick, you know, to Atlanta. And I remember we played in China the first game, preseason, and I had Shawn Marion that was an ambassador of the NBA. I'm like, “Trix, what do you think? Trix, what do you think? What do you think?” Like “I'm terrified. I'm terrified.” You know, but we know how that turned out.

Erik Torenberg: And by the way, it’s so funny, just the basketball gods, the Lakers made all the wrong moves for so many years, and were in such a tough spot. We talked about the Warriors, they were in a tougher spot, going forward, trading all their picks, being so old, and then Luka falls into their lap just to keep it interesting.

Mark Cuban: I think we’ve done enough of this.

Erik Torenberg: But the Mavs got Coop and, don't trade him, and you'll be good. Zooming out, closing here, how have you thought about how you want to spend time? Like you, for example, could have your own a16z, big venture firm doing all these investments, all this staff.

You seem to run kind of a lean operation, more flexible. How have you thought about the efforts that you wanna get involved with, not wanna get involved with? How thin to spread yourself, so to speak?

Mark Cuban: Well, I mean, I got outta Shark Tank. I got out of the Mavs so I can spend more time with my kids. You know that's number one. Because they're 16, 19, and about to turn 22. They're heading out on their own. So I just wanna spend that time that I can with them. But I think the difference is for me versus everybody. My goal is not to make as much money as I can. Like, I do investing. But honestly, it's just, people email me. I mean, like, I saw a list of who has the most unicorns, I had like 11. And all of them were just from email, somebody emailing me saying, “Check this out, check this out, check this out.” Like Synthesia, you know, I was their first investor. You know, gave them a million dollars and now, so I can get 20, 25% of a company sometimes more and then help them grow. That’s what I like. Taking, you know, longshot tries as opposed to, “Okay, let's raise a fund. Let's get a team together. We'll put in 10 to 25 million.” And then, you know, if they're doing well, we'll put in more and do follow ups. And, you know, we'll have these big scores. I like it better when, you know, Dude Wipes comes on Shark Tank. I gave them 200 grand for 15%. Now they're worth a billion dollars easy. You know, they're a primary brand and they never had to raise another cent.

You know, helping a company never have to raise more money. And making them profitable so that they're insanely profitable. I had this company, Alyssa’s Cookies. The guy, Doug, was living in his car with his family, and he sent me these cookies. They were this big, they were the best cookies I've ever—'cause I always do like the nutrition thing first. What's on the back? No added sugar, lots of fiber, etc. Low net carbs. But when I took it out of the plastic, it all fell apart. So I said, “I'm gonna give you money.” And I got like a third of the company. But I'm gonna guide you through what it takes. So we changed it and I told them to make them into bites instead of cookies and put them in a plastic container. And go to stores and we'll sell them. I'll do a tasting, you know, where I'll go to a grocery store one time and hand out…. Now, no more money in. We'll do 25 million this year. But that's not the important number because we don't spend money on advertising or anything. Makes 12 to 14 million a year. That's absurd. And we just added Whole Foods. We just added Costco. We're adding more and more, you know, outlets. If he can get to a hundred million and make 50 million a year. That stuff's fun.

00:53:43 - Education

Erik Torenberg: So if you could clone yourself a few times, or when you look into the future, are there other businesses like Cost Plus Drugs that you wanna start or incubate or…

Mark Cuban: Yeah, education. If I can do what I want to do with healthcare, which, saying a lot, right? But it would be education just because…

Erik Torenberg: Like AI-first K-12, or what kind of?

Mark Cuban: Yeah, I mean more secondary education, right simply because the cost structure is all fucked up. And it's all built on accreditation. And that's kind of a mafia. And when I see a mafia like that, I just ignore the accreditation and just go do it anyways. And if you get the results or guarantee your results, you’ll get plenty of…

Erik Torenberg: We talked about your healthcare platform, we talked about your messaging. Are there any other policies that if you could, you know, be in charge of the Democratic party or sort of running America that you would be most excited about in terms of its transformative potential?

Mark Cuban: You know what I would say, I would say a focus on entrepreneurship. You know, that, right now, whoever the candidate is, particularly if it's for President, takes credit for everything. And that's ridiculous. They do less to impact the startups and to start the building of businesses than anybody, you know. And so I think what's missing, definitely missing with Biden, maybe less so with Obama, then definitely missing with Trump. Like when you hear Trump talk about tariffs, he never talks about small businesses. It's always, “Hey, I've got this company investing $500 billion, 600 billion.” When you know you've got all these Shark Tank companies that are getting crushed. You know, you increase, you know, from a 2.5% average tariff to a 15% tariff. That's their profit margin, and it's gone. So I would put a focus on recognizing and encouraging and rewarding people starting businesses and using AI, you know, kind of as government as a service, where you simplify all the friction that's involved in starting a business, you know, from the paperwork to get all the requirements, the incorporation, etc, to a bank account, to all of it. Just simplify all of it so entrepreneurs can be entrepreneurs because I think that is what…Joe Biden told this to me of all people. During the Obama administration, they invited us, from Shark Tank to go do a thing for entrepreneurship at the White House. And he sat, and he gave a little speech and what he said was, you know, “I've traveled around the world, and there's no Chinese dream, there's no French dream, there's no UK dream, there's only the American dream. And people come here to live it and experience it. That's what makes us unique.”

00:56:22 - Advice to Silicon Valley

Erik Torenberg: Yeah. And on that note, in closing, what advice do you have for us in Silicon Valley or just tech more broadly who are building with AI on how to make it more popular so that we don't have this big backlash?

Mark Cuban: You have to get with either party and communicate to people the upside.

Erik Torenberg: Because they see the downside. They see, “They, maybe my job is being taken away.”

Mark Cuban: Yeah. And going back to what I said earlier, there will be disintermediation, but I don't think the net number of jobs is going to decrease. I think everything gets reinvented because AI right now is not smart. You know, it's statistical, in a lot of respects. Robotics on the other hand is getting smart. Robotics, if you clean my bedroom, it has to know, you know, to pick up the sheets and all this stuff. But I think the combination of the two will create unique opportunities for jobs.

I think domain expertise is always going to be necessary to train models. But I think, bigger picture, when like we look at a house, everything's designed for the human form. And right now robotics is trying to emulate the human form because that's what we've always done with software. We just emulated what we already did and tried to optimize it. If you can create devices, robots that can look like anything, and you want them to be domestic in some way, you don't build houses the way we do now.

And I think there's going to be a lot of jobs that go in different directions. And I think Silicon Valley has gotta do a much better job demonstrating those jobs, hiring people for those jobs, supporting people that are concerned about their jobs. Even teachers, you know, now when we look at AI tools, you can read papers faster. That's not gonna change anything. You're just doing it the same way. You know, but if a teacher can use AI to customize responses to kids to help them learn. Like you said, AI tutors, but in a classroom, you're gonna keep your job longer. We need more teachers, not fewer teachers. But Silicon Valley doesn't communicate that at all because when you're raising money, it's about what you're gonna change, not what you're gonna sustain. But there's so many unique opportunities.

I mean, I’m old enough, like I would walk in selling PCs and I saw you're outta work, you're outta work, you're outta work. But as it's all turned out, the economy just keeps on growing. And I think the same thing is going to happen here. And I think the other side of it is going to, you know, inner cities. Like I have this thing called the Mark Cuban AI Bootcamp. And all we do is go into, you know, underprivileged schools and encourage kids to sign up for the bootcamp. And it used to be just machine learning, where we taught you that. Now it's, you know, learning how to use models and everything and agents for 15, 16, 17-year-old kids. And it's great, right, because they're going back, and they're coming up with new ideas.

The valley needs to go out there and engage more in middle America so people see totally where the upside is. You know, you're getting this rejuvenation of San Francisco, and all the spending is pointing there. How is anybody else gonna understand that or see that? They don't at all. When in reality, if it's open to them, and that's why I did the bootcamp, more things are going to happen.

People don't realize, like, when the internet first hit, we used to think kids who can do HTML coding were geniuses. And when they moved up to JavaScript, they were even smarter, right? Because they were building all these websites that were getting ready to go public. And AI is kinda the same thing. It's not hard to build agents. And it's getting easier. But knowing an application that you can use to start something and improve something, everybody can do that. Every kid, every mom, every dad. But Silicon Valley is making it sound like, you know, this is the Jetsons and we have no chance.

And I think that has to change.

Erik Torenberg: That's a great note of advice and optimism to end. Mark Cuban, thanks so much for coming on the podcast.

Mark Cuban: It was a lot of fun, thank you. Appreciate it

This transcript has been lightly edited for readability.

Resources:

Find Mark on X: https://x.com/mcuban

Stay Updated:

If you enjoy the show, please follow and leave us a rating and review on Apple Podcasts or Spotify.

Find a16z on X: https://x.com/a16z

Find a16z on LinkedIn: https://www.linkedin.com/company/a16z

Subscribe on your favorite podcast app: https://a16z.simplecast.com/

Follow our host: https://x.com/eriktorenberg

Please note that the content here is for informational purposes only; should NOT be taken as legal, business, tax, or investment advice or be used to evaluate any investment or security; and is not directed at any investors or potential investors in any a16z fund. a16z and its affiliates may maintain investments in the companies discussed. For more details, please see a16z.com/disclosures.

Timecodes

00:00:40 — Introducing the Little Tech Agenda

00:03:23 — Pillars of the Agenda and the small builder vs. trillion-dollar company gap

00:05:00 — Competition, markets, and smart regulation

00:08:59 — “Regulate use, not development” and its misinterpretation

00:10:16 — The AI policy arc since 2023: hearings, fear narratives, and executive actions

00:18:05 — Licensing regimes and open source bans compared to nuclear policy

00:39:39 — China, export controls, and open source in a global race

00:47:21 — Federal vs. state lanes, moratorium fallout, and what’s next

Transcript

00:00:40 — Introducing the Little Tech Agenda

Erik Torenberg 00:00:40

Collin, Matt, welcome to the podcast.

Matt Perault 00:00:42

Thanks so much.

Collin McCune 00:00:43

Thanks for having us.

Erik Torenberg 00:00:44

So there's a lot we want to get into around AI policy, but first I want us to take a step back and reflect a little bit. We had publicly announced the little tech agenda July of last year. There's a lot that's happened since when we first take a step back, Collin, and talk about what is the little tech agenda and how did it come to be at the firm?

Collin McCune 00:01:01

Yeah. I mean, look, a ton of credit to Marc and Ben for having sort of the vision on this. I think when, you know, certainly when I first started here. I arrived. We started advocating on behalf of technology interest, technology policy, and I think what we realized was there have been these big institutional players that have been in DC in the state capitals for a very long time.

Some of them have done a lot of really good work on behalf of the entire tech community. But there wasn't anyone specific who was actually advocating on behalf of what I think, what we call little tech, which I think in my mind, are the startups and entrepreneurs, the, the smaller builders in the space.

And I think beyond that, what we realized was. They're not always a hundred percent aligned with what's going on with the big tech folks. And that's not necessarily always a bad thing or a good thing, but, I think that was the whole impetus of this, you know, how are we going to think about positioning ourselves in DC and the state capitals in terms of our advocacy on these issues, and how do we differentiate in between sort of.

The big tech folks who come with, you know, certain, there are certain degrees of baggage

Erik Torenberg 00:02:10

On the left and the right.

Collin McCune 00:02:11

And the smallest is small from the left and the right. Right. Yeah. And, and the smallest is small. So that, that was really sort of the basic impetus of this.

Matt Perault 00:02:17

For me, it was actually sort of almost a recruiting vehicle.

So I, when, when it hit in July, I was not yet at the firm. I started in November and when, when I first read the agenda. It sort of transformed the way that I looked at the rooms that I would sit in, where there would be policy conversations where all of a sudden you could see essentially an empty seat and, and little tech's not there.

You know, there's a, there would be conversations where people would say, um, and in this proposal we want to add this disclosure requirement, and then we'll have companies do a little bit more and a little bit more. And when you've read the little tech agenda, all of a sudden you start thinking, how is this gonna work for all the people who aren't in the room?

And so for me, the question like thinking about coming into this role in the firm was. Is this a voice, is this a part of the community I want to advocate for and think about? And when you start looking at the policy debate from a perspective of little tech, and you see how many of the conversations don't include a little tech perspective. It comes, from my point of view, it was very compelling, to think about how can I advocate for this part of the internet ecosystem?

Erik Torenberg 00:03:14

And Collin, why don't you outline some of the pillars of the little tech agenda or, or some of the things that we focus the most on? And maybe how it differentiates from sort of big tech more, more broadly.

00:03:23 — Pillars of the Agenda and the small builder vs. trillion-dollar company gap

Collin McCune 00:03:23

Yeah. I mean, just from a firm perspective, right? Obviously we're verticalized, we, you know, we all live and breathe this, and I think that's been very, very competitive for us on the business side. But I also think it's very competitive for us on the policy side too. Right? Obviously Matt leads our AI vertical and that sort of our AI policy lead.

We have a huge crypto effort. We have a major effort around American dynamism, and it's, and this is sort of defense procurement reform, which is something that the United States has needed forever and ever. We have, you know, other colleagues who work on the bio and health theme, and they're fighting on behalf of FDA reform, everything from PBMs. There's a whole, whole vertical there that they're working on. We're working a lot on FinTech related issues. And then, you know, just like classic tech related sort of internet entrepreneurs coming up, what, what does that relate to? There's a lot of tax issues that come along with it, and then of course, obviously there are the venture specific things that we have to deal with.

But look, I think I try and think about this from a basic point of view, which is just like. If you are a small builder, what are the things that should differentiate you between someone who's a trillion dollar company and you have hundreds of thousands of employees, right? If you're five people and you're in a garage.

How are you supposed to be able to comply with the same things that are built for a thousand person compliance teams? Like it's just not the same thing. Right. And and I like there are categories and categories and categories that, you know, Matt and I are dealing with on a regular basis, but that's probably the main pillar, which is five person versus trillion dollar company.

Not the same thing.

00:05:00 — Competition, markets, and smart regulation

Matt Perault 00:05:00

It's made my job actually really hard in certain ways since I started at the firm, because the kinds of partners that you want within our portfolio often don't exist in that a lot of the companies don't have a general counsel. Yeah. They don't have a head of policy.

They don't have a head of communications. And so the kinds of people who typically sit at companies thinking all day about like, what is this state doing in AI policy? What is this a federal agency doing in terms of rulemaking? They're not at startups that are just a couple of people and engineers trying really hard to build products.

Those companies face this incredibly daunting challenge. I mean, it seems so daunting for someone like me, like non-technical and I've never worked at a startup. If they're trying to build models that might compete with Microsoft or OpenAI, or Meta or Google, and that is unbelievably challenging in AI. You have to have data, you have to have compute. There's been a lot written about the cost of AI talent recently. It's incredibly, incredibly daunting. And so the question that Collin and I talk about all the time is. For those companies, what are the regulatory frameworks that would actually work for them as opposed to making that competition even more hard, even more difficult than it already is?

Erik Torenberg 00:06:09

One of the principles I've heard you guys, you know, hammer home is we want a market that's competitive for where startups can compete. We don't want a monopoly. We don't want even oligopolies, you know, a cartel-like system, and that that doesn't mean no regulation. Cause that can, as we've seen, that could be destabilizing too.

But it means smart regulation that, that enables that, that competition in the first place.

Matt Perault 00:06:29

So I think one of the things that was, that, that's been surprising to me to learn about venture is the time horizon that we, that we operate in. So we're our firm. Our funds are 10 year cycles. So, we're not looking to spike an AI market tomorrow, and have a good year or a good six months, or a good two years.

We're looking to create vibrant, healthy ecosystems that result in long run benefits for people and long run financial benefits for our investors and for us. And that means having a regulatory environment that facilitates healthy, good, safe products. It doesn't mean like if people have scammy, problematic experiences with AI products, if they think AI is bad for democracy, if they think it's corroding their communities, that's not gonna be, that's not in our financial incentive.

That's not good for us. And so, that really animates the kind of core component of the agenda, which is not trying to strip all regulation, but instead focusing on regulation that will also, that will actually protect people. And we think that there are ways to do that without making it harder for startups to compete.

Collin McCune 00:07:33

To Matt's good point. I actually, I walk into a lot of lawmaker offices and I, you know, it sounds like I'm pitching my book, but I genuinely say like, our interests are aligned with the United States of America's interest. Because the people that we're funding are on the cutting edge.

They're the people who are gonna build the companies that are gonna drive the jobs, they're gonna drive the national security components that we need, and they're also gonna drive the economy. And like we wanna see them build over a long time horizon. And, and like that is exactly what, how we should be building policy in the United States.

Of course, like half the offices I walk into like, all right, great, that guy, get that guy outta here.

Matt Perault 00:08:13

99.9% of people we talk to think that all we want is no regulation. And yeah, we, despite like writing extensively, like both of us writing, speaking extensively about the importance of good governance for creating the kind of markets that we want to create, and Collin can speak more to it in crypto, I've learned a lot from our crypto practice because the idea there is you really need to separate good actors from bad actors and ensure that you take account for the differences.

And it's true in AI as well. If we don't have safe AI tools. If there is absolutely no governance, it's that's not going to create a long run healthy ecosystem that's gonna be good for us and good for people throughout the country.

Collin McCune 00:08:51

I actually can't think of a single example across the portfolio in which we are arguing for zero regulation.

00:08:59 — “Regulate use, not development” and its misinterpretation

Matt Perault 00:08:59

The core component of our AI policy framework, which was developed before my time, I wish I could take credit and I can't, is focus on regulating harmful use, not on regulating development. And that sentence, regulate use, do not regulate development, somehow is interpreted as do not regulate, and people just omit for some reason, the part that we focus on, on focusing on regulating harmful use.

And that in our view is robust and expansive and leaves lots of room for policy makers to take steps that we think are actually really effective in protecting people. So regulating use means. Regulating when people violate consumer protection law, when they use AI to violate consumer protection law or when they use AI in a way that violates civil rights law at the state and federal level, or violating state or federal criminal law.

So there's an enormous amount of action there for lawmakers to seize on. And um we really want that to be like an active component of the governance agenda that we're proposing. And for some reason, it's all passed over and the focus is just on the, on don't regulate development. I don't exactly understand why that, why that ends up being the case.

Collin McCune 00:09:59

Easy headline.

Erik Torenberg 00:10:01

So there's been a lot that's happened in AI policy and I wanna get to it. But first, perhaps Matt, you can trace the, the evolution a bit over the last few years. I believe there was a time where like, pattern matching with social media regulation a bit. Why don't you trace some of the, the biggest inflection points or, or kind of the debates over the last few years and we'll get to today?

Maybe Collin.

00:10:16 — The AI policy arc since 2023: hearings, fear narratives, and executive actions

Collin McCune 00:10:16

I think we, we have to play a little bit of history. I wanna get to, you know, sort of a point that I think is the really critical point of what we're all facing here. For us, for me, I would say from a policy and government affairs perspective, this conversation started early 2023.

That, that was, that was sort of like the kickoff of the gun. It sort of puttered along and became more and more real over time. But in the fall of 2023, so almost, almost exactly to the day, two years ago. There was a series of Senate hearings in which, you know, some major CEOs from the AI space came and they testified, and I think that the message that folks heard was, one, we need and want to be regulated, which I think maintains that's truth today.

That's obviously, you know what, Matt and I are working on a, on a regular basis. But I think, included in some of that testimony was a lot of speculation about the industry that led to and sort of absolutely jumpstarted this whole huge wave of conversation around the rise of Terminator. You know, go hug your families because we're gonna all be dead in five years.

That spooked Capitol Hill. I mean, they absolutely freaked out about it. And look, rightfully so, you have, you have these really important, powerful people who are building this really important, powerful thing. And they're coming in, they're gonna tell you that, you know, everyone's gonna die in five years.

Right? That's a scary thing for people to hear. And oh, by the way, we want to be regulated. Which, you know, look, that starting gun, I think moved us in hyper speed. Into this conversation around how do we lock this down? How do we regulate it very, very, very quickly. I think that led to the Biden executive order, which we, you know, we have publicly sort of, you know, denounced in certain categories.

That executive order led to a lot of the conversation that I think we're having in the states. A lot of the, you know, sort of bad bills that we've seen come through the states. I think it also, led to a number of federal proposals that we've seen that have not been very well thought through also. And look, you know, I think people are kind of sitting around, they're like, oh, well, you know, was it just like, you know, some testimony from the CEOs that did this? And the answer to that is no. You know, from my, from my point of view and look, you know, they, they deserve a lot of credit. I think the effect of altruist community for 10 years backed by large sums of money were very, very effective at influencing think tanks and nonprofit organizations in DC and the state capitals to sort of push us in a direction where people are very fearful about the technology, and that has, that has shaped, significantly shaped the conversation that we're having throughout DC and the state capitals and candidly, on a global stage.

You know, the EU act and, the EU AI Act, we're, we're public on that. There's a lot of very, very problematic and provisions in there. All of this banner of safetyism came from this ten year head start that these guys have had. So when I always, you know, that that's kind of a bit of the history, but sort of as an aside to this, I always just have to smirk or, you know, smile to try and laugh it off.

But I mean, when people are writing these articles about the fact that the AI industry is, you know, pumping all this money into the system, certainly like there, I'm not suggesting that there's not money in the system. We're obviously active on the political and policy side. We're, you know, we're not hiding that.

But it is dwarfed by the amount of money that is being spent and has been spent over a 10 year window. And candidly, I mean, the reason that Matt and I have jobs is because we are playing catch up. We are here to try and make sure that people understand what is actually going on in this conversation and be a counterforce to this, this group of people and, and this idea, this ideology that has been here for a long period of time.

So that's, I look, you know, that that's kind of the briefer on this.

Matt Perault 00:14:29

I mean, and companies. I think we're ready to consider some policy frameworks that, that I think we're probably really going to be challenging for the AI sector in the long run. Right. And I think that's because I was at Meta, then Facebook starting in 2011 and through 2019, and so after really like 2016, there was aggressive criticism of tech companies.

And the general framing is like, you're not being responsible and regulation needs to catch up. Governance of social media is behind where the products are. And whatever you think about that, that was really the kind of strong view in the ecosystem that like governance has allowed, the lack of governance has allowed problematic things to happen.

And so I think when AI was starting to accelerate and you had certain sort of prevailing political interests, I think that were driving the conversation, companies rushed to the table and I think it was a group of five? Three, five, seven companies who went into the White House and negotiated voluntary commitments.

I mean, we don't even have to make the argument about the importance of representing Little Tech in when you see that there is a set of companies who negotiated an arrangement for what it would look like to build AI at the frontier with all current developers who weren't those companies and all future startups not represented at the table.

I think that is why, like we started to think about the value of having more dedicated support around AI policy, because clearly the views of little tech companies aren't represented in the conversation.

Collin McCune 00:16:02

Well, I mean, let me, let me just add one thing to this. It's Marc and Ben's story. They've told it many times. I was in the meeting as well, you know, and, and like, you know, everything they've said has been a hundred percent true and accurate. But there was a, there was a prevailing view by very, very powerful people of the previous administration that this was going to be only two or three major companies able to compete in the AI landscape and because that was the case. They needed to be basically locked down and, and put in this incredibly restrictive view from a policy and regulatory perspective. And that was gonna be kind of like this. This entity that was kind of like a arm of the government. And I think that that was the most alarming thing that I think we had heard from the administration on top of an incredibly alarming series of events that happened on the crypto side, including sort of wanting to eradicate it off the face of the planet it seemed like. So I think that that all led to kind of the position that we're in now. And certainly like Matt's hiring and the thing, you know, like us building out the team, et cetera.

Matt Perault 00:17:12

That narrative is clearly like a very alarming, maybe the most alarming version of this, but even since I've been in this role, I've heard other versions of it where people will say, "oh, don't worry about this framework. It just applies to three or five companies," or "it just applies to five to seven companies," and I think they mean that.

To provide comfort to us, like, oh, this isn't gonna cover a lot of startups. But the view of the AI market where there are only a small number of companies building at the frontier is not the, that's not the vision for the market that we have. We want it to be competitive and diverse at the frontier. And the policy ideas that we're coming out of the period that Collin's talking about, we're dramatically different from where they are today in a way that I think like some people have even like lost sight of exactly where we were a couple years ago.

There were ideas being proposed by not just the government, but industry to require a license to build frontier AI tools and for it to be regulated like nuclear energy, not just

Collin McCune 00:18:03

Which would be historic for software development.

00:18:05 — Licensing regimes and open source bans compared to nuclear policy

Matt Perault 00:18:05

Yeah. Right. Unprecedented.

And for it to be regulated like nuclear energy with like an international style nuclear like.

Sorry, an international-level nuclear-style regulatory regime to govern it. And we've moved like no matter what you think about the right level of governance, there are not a lot of people now who are saying what we need as a licensing regime where you literally apply for permission from the government to build the tool, but that wasn't that far in the rear view mirror.

Collin McCune 00:18:31

Yeah. And, and look, and we're also talking about bans on open source. Right. I'm like, we're still kicking around that idea at the state level. You know, look for, for us who live and breathe the tech stuff on a daily basis, this, this is, you know, this sounds insane, crazy. But let me, you know, like, just to make it a little bit more real, right?

Like the nuclear policy in the United States has yielded two, three new nuclear power plants in a 50 year period since these organizations have been started and look like you can, some people are pro-nuclear, some people are anti-nuclear. I don't wanna get into that debate. The point though is, is that that was not the intended policy of the United States of America.

That was the effect of putting together this agency and what has come from that. And I think, you know, look. If we do the same thing to AI, had we done the same thing in AI in that period of time, then you don't have the medical advancements, you don't have the breakthroughs, you don't have all of the things that come from this that are incredible, but beyond that.

We lose to China.

Full stop. You lose to China and then our greatest national security threat becomes the one who has the most powerful technology in the world.

Erik Torenberg 00:19:43

Right. And I think, I think the early concern on the open source was that we would be somehow giving it to China, but then we've seen with DeepSeek, et cetera, that they just have it anyways, right?

Collin McCune 00:19:51

Exactly right. Exactly. You know, the idea that we could lock this down, I think, I think, you know, I mean Marc and Ben have talked about this. I mean, I think they've debunked that a number of times.

Erik Torenberg 00:20:00

Just to understand, was, for the previous administration, what was their calculus? Was it that they were true believers in the fears?

Was it that there was some sort of political benefit to having the, the views that they had, especially on the crypto side. I don't understand what's, what is the constituency for anti-crypto stance how do you make sense of sort of the players, or the intentions or motivation, just to understand sort of the, the calculus there.

Collin McCune 00:20:22

You know, I mean, look, I, I think that that's a really, I think that's a really hard one to answer and I'm not sure I can pretend to be completely in their minds. I think there's a couple of different competing forces here. Like one is, you know, what are the constituencies that support sort of that administration, what are the constituencies that support, that side of the aisle?

And I think that. Especially over the last 10 to 15 years, it has been very, very heavy focus on consumer safety, which I think look a very important thing. And we're obviously in alignment on that. I think everyone should be in alignment, have to protect, consumers, have to be able to protect the American public.

But I think that a lot of that conversation has been weaponized. I think that it is, it is a big time moneymaker. I think a lot of these groups either get backing from very, very wealthy special interest. Or they are small dollar fundraising off of quick hits. Like, you know, AI's coming for your jobs, donate $5 and we're going to, you know, and we'll make sure that we take care of this in Washington for you.

And, you know, like pretty easy, you know, it's a pretty easy manipulation tactic, you know, it's used like from a bunch of people. But I think that that's like a very, that, that held very seriously true. Right. And I think the other thing here is that I think, personnel is policy. It's the old saying, personnel is policy.

And I think a lot of the individuals that, were in very senior decision making roles within that White House and that administration came from this sort of consumer protection background where they've seen this, that was constituency. They were put in this position to come after private enterprise.

Like, you know, that was, that was the, that was the goal. Like there's this whole idea out there. I think among some of those folks that, you know, Senator Warren has, has, you know, proposed this many times is, is like, if you're not getting, you know, if you're not going after and getting people on a regular basis in the private sector, then you're not working hard enough.

And I, and like, I, I just, you know, I think that, that, that. Is, is probably like the second thing, and like the third is just, we're at this very weird moment where being a builder and being in private enterprise is, is a bad thing to some policymakers. It's not you, you're not doing good because you're earning a profit and you know, they certainly won't say that, but the activities and the things that they're doing are, are a hundred percent aligned with that, that type of idea.

So I, you know, I, I think that's the basic crux of it.

Matt Perault 00:23:02

I think the things that motivated that approach were done in good faith. And I think, I think it's what you alluded to earlier, which was like I don't share this view, but there are a lot of people who believe that social media is poorly regulated and that because policymakers were asleep at the wheel we woke up at some point, I don't know, sometime in the 2014 to 2018 period and realized that we had technology that we thought was actually not good for our society. And I think that whether or not you think that that's true or not, that I think that was, that has been a widely held view. It's a wide, it's a held view on the right and on the left.

It's a bipartisan view. And so I think when this new technology came on the scene, this was a do-over opportunity for policymakers, right? Like, we can get this right when we didn't get the last thing right. And so I understand that motivation. It makes a lot of sense. I think the thing that we, that, that we strongly feel is the set of policy ideas that came out of that like good faith, belief were not the right policy ideas to either protect consumers or lead to a competitive AI market. Like yeah, some of, many of the politicians who are pushing were pushing concepts that would've really put a stranglehold, I think, on AI startups and would've led to more monopolization of a market that already tends toward monopoly 'cause of the high barriers to entry already.

Those politicians three years before had been talking about how problematic it was that there wasn't more competition in social media. And then all of a sudden they're behind, you know, a licensing regime, which is not. I don't think there's much economic evidence that licensing is pro-Competitive is typically is the opposite, right?

The disagreement is less with the core, feeling like we wanna protect people from harmful uses of this technology and more from the policy concepts that came out of that feeling that we think would've been disruptive in a problematic way to the future of the AI market.

Erik Torenberg 00:24:51

Anecdotally, it seemed from, from afar that some of the concerns early on were almost, you know, to match social media, like around disinformation or even like DEI concerns.

And then, you know, people were trying to sort of, uh, make sure the models were compatible with, compatible with sort of the sort of, you know, um, speech regime at the time. But then it sort of shifted to, oh, wait, you, is this, is there more existential concerns around jobs? Or, or is AI even like nukes in, in the sense of like even people doing harm or AI itself doing harm?

But it, it seemed to. To escalate a bit and, you know, maybe aligned with that testimonial that you alluded to.

Matt Perault 00:25:22

I experienced it as feeling like the goal posts always move and one of the things that I said like that, I said that I started asking people when I was really trying to settle into this regulate use, not development policy positions.

What do we miss? Like if we regulate use primarily using existing law, what are the things that we miss? And I haven't gotten very many clear answers to that. Right? Like, you can't do illegal things in the universe, and you also can't use AI to do illegal things. And typically when people list out the set of things that they're most concerned about with AI, it are, they're typically things that are covered by existing law.

Not, probably, not exclusively right, but primarily. And so that at least seems like a good starting point. Some of the other issues that I think are like understandably ones that we should be concerned about have a range of different considerations associated with them. Like the, like if you're concerned about misinformation or like speech that you think might not be true or might be problematic, there are significant constraints on the government's ability to regulate that.

The First Amendment imposes pretty stringent restrictions and I think for very good reason because you don't want the government to dictate the speech preferences, policies of private speech platforms for the most part. And so, so that, those issues might be concerns, but they're not necessarily areas I think where you want the government to step in and take strong action. And so there, I think there are things that we should probably do as a society to try to address those issues, but, government regulation maybe isn't the primary one. And again, in most of the things that people are most concerned about, like real, real use of the technology for clear cognizable, real world harm, existing law typically covers it.

Collin McCune 00:26:55

I have a theory on this, so I, I think everything that Matt just said is, is spot on. But, but you know, like then, then you're kind of sitting around and you're kind of scratching your head. It's like, okay, well if use covers it and there hasn't been, you know, a, a very incredibly fair rebuttal onto why use is not enough in, in terms of focus on, on the policy and regulatory side.

What's, what's the answer? I think we're, we're experiencing sort of this, I don't know if it's phenomena, but we're experiencing this pattern on the crypto side too, which is, which is we're having a very, very spirited debate on the crypto side of things, on how to regulate sort of these tokens and how do you launch a token in the United States is it a, security is it a commodity.

And this is sort of this age old debate that's, you know, plagued securities, traditional securities laws for years, but also certainly the crypto industry. But what we have found is there are, there are a number of people who have entered this debate who are actually trying to get at the underlying securities laws.

Like they, they want to reform securities laws. They don't wanna reform crypto laws.

Erik Torenberg 00:27:59

Interesting.

Collin McCune 00:27:59

That involve securities. And this is their only venue by which they can enter that conversation because we're not having there, there's no will from the congress or from policy makers to go and overhaul the securities laws right now.

You know, it's just not there. But what is moving is crypto. So people, you know, there are all these people that are now trying to enter this debate and like, oh, well we should re-look at this. I'm like, well, this doesn't have anything to do with it. We shouldn't be entering this conversation yet. They're still pushing, right?

Yeah, and that's kind of muddying the water. I think a very similar thing is actually happening on the AI side, which is, you know. There are a number of members of Congress that feel like, well, we missed it on the 96 Telecom Act. Like, that wasn't, we didn't do good enough around then so we need to re-right the wrongs through the venue of an AI policy conversation. Right. Because if you, if you think about it, right. Assuming that use doesn't go far enough for someone. Right, and this is the same conversation that we're having in California right now or in Colorado right now. If uses does not go far enough. Okay. Well then it would be really, really simple if you could have a privacy conversation around this.

If you could have an online content moderation conversation, an algorithmic bias conversation around that, you could do all of that. Wedge it through AI and then assuming AI is actually going to be the thing that we all think it's gonna be. Now you've put basically a regulatory funnel on the other side.

Like you've put a mesh screen where everything has to run through AI and therefore it runs through this regulatory proposal you put together.

Matt Perault 00:29:31

The thing that I've really been wrestling with in the last few weeks is whether those kinds of regimes are actually helpful in addressing the harm that they purport to want address.

And Colorado is a really good example. So there are all these bills that have been introduced at the state level. Colorado's the only one that's passed so far that set up this, this regime where you basically have to decide are you doing a high risk use of of AI or low risk use of AI and this would be for startups.

That don't have a general counsel, don't have a head of policy. Can't hire an outside law firm to figure it out. High risk, low risk. And then if you're high risk, you have to do a bunch of stuff. Usually impact assessments. Sometimes audit your technology to try to anticipate is there going to be bias in your model in some form, which maybe an impact assessment helps you figure that out a little bit.

But it's probably not going to eliminate bias entirely. It certainly isn't going to like end racism in our society. There was a Colorado is now their, their governor, their attorney general have, have put pressure on the legislature to roll back this law because they think it's gonna be problematic for AI in Colorado.

And so there was just a special session there to consider various different alternatives. One of the alternatives that was introduced proposed codifying that the use of AI to violate Colorado's anti-discrimination, anti-discrimination statute is illegal. That's consistent with the regulate harmful use framing that we've talked about, and it's instead of having this like amorphous process where maybe you address bias in some form, maybe you don't, this goes straight at it.

It's not a bank shot. It goes straight at it. Where if someone uses AI in a way that violates anti-discrimination law, that would be. That could be prosecuted, the attorney general could enforce. And I don't, I still don't understand why that approach is not, is somehow less compelling than this complex administrative paperwork approach.

I think it's kind of the reason that Collin's describing, which is like, people want another, a different bite at the apple of bias, I suppose, but it's not clear to me that that, that it's actually the best way to effectuate the outcomes that you want, as opposed to just criminalizing or creating civil penalties for the harm that you can see clearly.

Collin McCune 00:31:32

It's also, I, I mean in, in policymaking and bill writing, it, it, it's really, really easy to come up with bad ideas. Yeah, it's easy, right? Because they're not well thought through. The first thing comes to your head, someone publishes a paper on something. Here we go. It takes real hard work to get something that actually works and then it's even harder to actually go through a political and policy negotiation with a diverse set of stakeholders and actually land the plane on something.

Matt Perault 00:31:57

I think, I think that's part of the reason that people think that we are anti governance because when we've, I mean, Collin, again, he lived this history. I'm coming in late to it, but like as we were ramping up our policy apparatus, these were the ideas in the ecosystem licensing, nuclear style regulation, like flops, threshold based disclosures, really complicated transparency regimes, impact assessments, audits, which are a bunch of ideas that we think are not going to help protect people and are gonna make it really hard for low resource startups. And so we've been trying to say, no, no, no, don't do that.

And so that sounds like deregulate, but for whatever reason, it's been hard so far to shift toward like, here's another set of ideas that we think would be compelling in actually protecting people and creating stronger AI markets.

Erik Torenberg 00:32:43

Right now we don't see, you know, terrorists or criminals being aided, you know, 1000 x with AI in, in performing terrorism or crime.

Like when I ask people like, what are you truly scared about? Like, gimme a concrete scenario. People, you know that they'll be like, oh, what about like bioterrorism or something? Or what about, you know, cybersecurity, you know, theft something. We seem very far away from that. Is there any amount of development at, you know, in the next few years, any amount of breakthroughs where you, where you might say, oh, you know maybe use isn't enough or, or do we think that that will always be a...

Matt Perault 00:33:15

I think it's conceivable. I mean, and, and I think we've been open about that. Like we, we, we think existing law is a good place to start. It's probably not where we end. So Martin Casado, one of our general partners, wrote a great piece on marginal risk in AI, basically saying like, when there's incremental additional risk that we should look for policy to address that risk. And so the situation you're describing, I think might be that, I think what you're getting at is a really important question about just potential significant harms that we don't yet contemplate. We get asked often about our regulate use, not regulate development framework.

Are you just saying that we should address issues after they occur? And I understand why that's a concern. Like there might be future harms and wouldn't it be nice if we could prevent them in advance? But that is how our legal system is designed. And typically when you talk to people about ways that you could try to address potential criminal activity or other legal violations ex-ante before they occur. That's really scary to people like, Erik, what if we just learned a lot of information about you and then predicted the likelihood that you might do something unlawful in the future. And if we think it's exceeded a certain threshold, then we're gonna go and try and take action against you before you've done it so that we can prevent future crime that you're laughing cause it's laughable. We, we don't want a kind of ex-ante surveillance both because it feels invasive, but also because it often is ineffective. Like you might, it might, we might run some test that shows that maybe you're likely to be predisposed to some kind of criminal activity, but we don't know until you've done it that you are going to, that, that you've done it.

And so I think that kind of approach, again, I think it's motivated by a really valid concern. And a valid desire to prevent harm. What if we could prevent harm before it's occurred? The challenge is the regulatory framework, I think probably won't do that. It probably won't have the effect of preventing harm.

And there are all these costs associated with it, mainly from our perspective inhibiting startup activity.

Erik Torenberg 00:35:12

Marc once told me on a podcast, he told me the joke, which is, uh, man, man goes to the government. Uh, you know, I go to the government because I have this big problem, now I get a lot of regulation. Now I have two problems.

Okay. Let's talk about the, the state of AI policy today. There's a lot that's happened in the last few months with the, the moratorium, the, the Action Plan. What are some of the things that we're excited about right now? What are some of the things we're, we're less excited about right now? Why don't we give a breakdown of where we're at right now?

Matt Perault 00:35:40

So I think given what Collins described about where things were a couple years ago, it's great to see the federal government certainly the executive branch, but not just the executive branch. I think this is in Congress across both aisles being supportive of frameworks that we think are much better for little tech.

So trying to identify areas where regulatory burden outweighs value and where we can right size regulation to make it easier for AI startups. As Collin said, support for open source. We were in a really different place on that a couple years ago. Now it seems like there's much more consensus. And again, actually it was across the end of the last administration and the current administration around the value of open source for competition and innovation.

The, the National AI Action Plan also had great stuff in it about thinking through the balance between the federal government and state governments, which is something that we've done a lot of thinking about. There's an important role for each but we think the federal government should really lead regulation of development of AI states should police harmful conduct within their borders and I think there's stuff in the Action Plan that would try to ensure those respective roles. There's also a lot of stuff in the Action Plan. It wasn't really talked about much. It wasn't sort of the headline grabbing stuff that I thought was incredibly compelling in terms of, again, trying to to create a future for AI that just works better for more people.

And a really good example is the stuff on worker retraining that focused on different programs that could help workers if they're displaced as a result of AI, as well as monitoring AI markets and labor markets to make sure that we understand when there are significant labor disruptions. So I think it sort of gets at a point that you were alluding to a couple minutes ago about like, what happens when there's something really disruptive in the future.

Can you predict with certainty that there won't be this crazy disruptive thing and no, we can't, there, there might be significant labor disruption. Others at the firm have talked extensively about how typically there's always, there are worries about labor disruptions when there's new technology introduced.

Typically, there are increases in productivity that end up being good for labor overall. We think that's the direction of travel, but you never know. We can't predict it with certainty. And so I think it's a really strong step to try to just monitor labor markets to see what the disruption might look like so that we're set up to take strong policy action in the future.

Collin McCune 00:37:53

Can I, can I just say one thing about the AI Action Plan?

Matt Perault 00:37:56

Sure.

Collin McCune 00:37:56

And I, I don't wanna juxtapose this to what we saw under the Biden administration, which is, incredible amount of activity under the Biden administration. Incredible amount of activity under the Trump administration. But, you know, look, I kind of view these executive orders and these plans that come out from administration are very, very important.

And some of them have true policy. They direct the agencies to do things, to come out with rewards and then take under rulemakings and things like that. But from an AI Action Plan perspective. For me, it was so significant because I think it turned the conversation on its head before it was, we have to, we have to only focus on safety with a splash of innovation. And now it is. We understand how important this is from a national security perspective. We understand how important this is from an economic perspective. We need to make sure that we win. While people, while keeping people safe. Right. And that dynamic and that shift of rhetoric is incredibly important because what that does is it signals to the rest of the world, it signals to other governments that this is the position of the United States and will be the position for the next three and a half years, and this is the position of the United States to the Congress. So when the Congress is looking at potentially taking up pieces of legislation or taking actions, or even committee hearings, which, you know, for the, the broad base of what we're talking about are fairly insignificant, all of that is sort of kept in mind.

So now the conversation has shifted significantly and that that is really, really important.

Erik Torenberg 00:39:30

Speaking of, of winning Collin, I'm curious for, for our thoughts, on AI policy vis-a-vis China, whether it's export controls or any of the, you know, issues we care about.

00:39:39 — China, export controls, and open source in a global race

Collin McCune 00:39:39

Yeah, I mean, well, I mean, look, first and foremost, we've talked about it already. I mean, we have to win, right? And, and I think, I think that that is, that is at, that is at the main thrust of a lot of what we're doing here and a lot of the way that we think about this from a firm perspective. You know, I think first is making sure that the, the founders and the builders can build appropriately with appropriate safeguards and an appropriate regulatory structure.

The second is, how do we win and make sure that America is the place where AI is, is probably the most functional and foundational vis-a-vis China. You know, I think that, there has been a long conversation, the Diffusion Rule that came out from the Biden administration specifically on export controls.

Many, I think, panned that proposal. I think that that was, a lot of people suggested it was probably too restrictive. It wasn't the right way to think about things. I think, you know, we have spent most of our time, Matt, leading this effort has spent most of his time, our time specifically focused on how are we regulating the underlying models and how are we regulating hopefully the use of these models.

Versus specifically sort of on the export control piece. What I will say though is very concerning, sort of some of the proposals that came out from the Biden administration, some of the proposals that we've seen at the state level and some of the proposals that we've seen at, in, at the congressional level, federal standpoint that dealt with specifically export controls on models themselves. And we're still kind of having this conversation. There's, there's a, there is a policy set that has been kicked around for a while. It's called the outbound investment policy, which is basically how much US money from the private sector is flowing into Chinese companies.

And very noble, laudable, you know, super supportive of that concept. You know, we are a very sort of primary America, America first sort of organization here. We're investing primarily in American companies and American founders. So, you know, we, we are, we're very supportive of it. But when you, you sort of edge into the idea that we might inadvertently ban US open source models from being able to be exported across the country. Like by definition of open source, there is no, there are no walls around these types of things. So that's one of the areas that we've been very, very focused on and I think obviously very important to make sure that we don't have these very powerful technologies, US main technologies in, in the hands of our Chinese counterparts and the PLA and CCP using this against us.

But I also think that we need to make sure that we're not extending too far and limiting the power of open source technologies to be able to kind of be the platform around the world. You know, the final point that I'd make here is we do all ultimately and fundamentally have a decision to make, as you know, the US, which is, do we want people using US products across the world?

Which helps for a whole bunch of different reasons, but certainly on soft power from a national security perspective, or do we want people to use Chinese products? The more that we lock down, obviously American products, the more Chinese, the Chinese will enter those markets and sort of take a land grab in that space.

Erik Torenberg 00:43:04

Why don't you get into more what happened with the moratorium and, and the fallout that ensued.

Collin McCune 00:43:07

I think this one's is a bit complicated. There was a perception about the moratorium when it came out that it would've prohibited all state law from existing for a 10 year window. Obviously, that's a long period of time.

I'm not sure we would necessarily completely agree with that policy stance. That from our point of view is a misinterpretation for a whole bunch of different reasons of actually what the language said. But you know, sometimes in DC a lot of times in DC perception is reality. And that that kind of, that kind of took hold.

But I, I also think that there, there are also, you know, strong competing forces like we've discussed right from the I, I think the doomer crowd or the safety crowd that were very, very anti, that had had used all of their tentacles that they've spread out over the last decade to try and move in and try and kill this.

I think they also were successful in leveraging some other industries to try and come in and also move forward to try and kill this thing and look. You know, by virtue of the vehicle, the underlying procedural vehicle, this reconciliation package that it was moving in, it was a partisan exercise. It was gonna be Republicans on Democrats, and that was that, right?

And there was nothing, even a prominent AI policy that was gonna be dropped in a reconciliation package that was ever going to drag Democrat votes over it because it was such a big sort of Christmas tree style thing that had all kinds of, all kinds of tax reform positions, et cetera. And if you are in one of those situations, the margins on the votes become very, very, very small.

So all it took was, you know, one or two Republican senators hitching their wagon to some of these ideas that were out there to tank this thing, right? And look. I think that's gonna be a situation that you're gonna fight in any sort of political policy, legislative outcome, or any sort of, any, any sort of issue that you're gonna be running within the Congress.

Right. But I think more so than anything. Um, and we heard this repeatedly from a whole bunch of different people. And this is what we've also experienced. The industry was just not organized well enough, right? And that's not just the industry. It's also the people who care about this thing that aren't actually industry stakeholders.

The stakeholders who were pro some level of moratorium or some level of preemption were just not organized. And I think that that, that was a, you know, both a eye-opening moment, but also an important moment because I think what we have done in the proceeding, you know, three, four months since this thing has gone down is we've taken a long, hard look at what we need to do collectively from a coalition to be able to be in a better position next time we're there. And so what does that look like, right? I mean, first and foremost, it comes with writing, doing podcasts, talking about these things, talking about the details of what's actually in these proposals and what it actually means for states and, and the federal government to make sure that we're fighting through the FUD that's coming through, because it's always gonna be there. There's misrepresent misrepresentation all over the, all over the field. The second piece is let's all get on the same page, which I think we've, we've worked very hard to do and, and where we can find alignment, we, I think we've found that alignment between big, medium, and little.

And then I think the third and probably the most important is. What are we doing on sort of the political advocacy side to make sure that we have the appropriate tools to be able to push forward in a way that ensures that America continues to lead and that we don't lose out on this race to China. And that's, you know, part of the reason that we have recently announced our donation to Leading the Future PAC, which will have, you know, several different entities underneath it, which I think is, is designed to sort of be that political center of gravity in the space.

And that will fight at the federal level and the state and local level. So we're, we're happy to be a part of it and I, we expect, you know, there will be others that join this sort of common cause fight on the AI side.

00:47:21 — Federal vs. state lanes, moratorium fallout, and what’s next

Erik Torenberg 00:47:12

If we could wave a wand, what would we like to be done at the state level, what we like to, versus the federal level versus how should we think about that, that interplay that compared to where we're at now?

Matt Perault 00:47:21

So I, I think there the helpful answer here comes from the constitution. Constitution actually lays out a role for the federal government and a role for state governments. Federal government takes the lead in interstate commerce, so governing a national AI market and governing AI development, we think is primarily Congress's role.

Sometimes when, when people say that, I think the, what other people hear for some reason is states should do nothing, and we have been, we've tried very hard to be very deliberate in not saying that and making clear that states have an incredibly important role to play in policing harmful conduct within their jurisdictions.

So criminal law is a perfect example. There is some criminal law at the federal level, but the bulk of criminal laws at the state level, like when you think about routine crimes, if you are going to prosecute someone, uh, prosecute a perpetrator, the, it's likely that that would occur under state law. And so to the extent we want to take account of local activity that, um, that would, where there's criminal conduct involved and we wanna make sure that the laws are robust enough to protect people from that activity, that's gonna be primarily state law.

Oddly enough, I mean, as Collin is describing, like we, this isn't the delineation that we've started out with. There are a lot of state laws that have sort of taken the approach of some, sometimes explicitly, Congress hasn't acted. So we have a responsibility to act, and that's true to some extent.

Like you can act within, states can act within their constitutional lane. Some of what states have done. Have gone outside that lane. And so we actually just this week released a post on potential dormant commerce clause concerns associated with state laws. And the basic idea there is that there's a constitutional test that says that states cannot excessively burden out of state commerce if, when it, when that greatly exceeds the in-state local benefits. And so courts actually weigh that there's a balancing test. Are the harms cost to out-of-state activity do, do those significantly outweigh the benefits on the local side? And we think that, at least for some of the proposals that have been introduced, it's likely that they won't, that the benefits are somewhat diminished relative to what the proponents think they are and that the costs are significant, like the cost of a developer in Washington state for complying with a law that's in California, or a law that's in New York is gonna be significant. And so our hope I think is not that the dormant commerce clause ends up serving as a function that makes it hard for states to enact laws, but actually just get serves as a guidepost for states around the kinds of laws that they might actually introduce.

And I think it pushes in the direction that's consistent with our agenda, which is to, to take an active role in legislating and enforcing laws that are focused on harmful use.

Erik Torenberg 00:50:07

Looking in the next six months to a year, what, what are the issues that we're most focused on or that we're thinking about are going to, you know, be playing a role in the conversation?

Collin McCune 00:50:14

I think it's first and foremost some level of federal preemption, and I wanna be very specific about this, again, to Matt's point. We're not talking about preempting all state law. We're talking about making sure that we have a federal framework specifically for this model regulation and, and hopefully how the models can be used.

Right. I think that's gonna be so, so critical because we can't, just like any other technology, no technology can live under a 50 state patchwork. And, and, and that's, that's been the biggest issue that we've been fighting over the last year and a half or so. So I think, I think that. I think that there are some other sort of policy sets that I think will be handled beyond that that I think can kick into sort of workforce training.

I think there's some literacy things that should be coming up. Obviously there's a huge robust conversation around data centers and energy that I think it will be really, really important. But above all, I think most of our time and energy will be focused on trying to have some level of federal standard here.

To try and drive the dividing line between the federal and state government, which I think Matt has already done a ton of great work on.

Matt Perault 00:51:29

I think this is just a super exciting policy moment for AI. There's the, the last couple years where I think there are a bunch of ideas that have been proposed and for the reasons that we've discussed, we think those ideas fall short, both in terms of protecting consumers and in terms of ensuring that there's a robust startup ecosystem. Most of those laws, I think, have actually not succeeded in passing. So like the, there were a number of laws introduced at the state level in the, in, in this past year's legislative sessions that we thought had of strong likelihood of, of passing and I think to date, none of them have passed. Collin has also been building out the expertise and skillset and capacity on his team. We just hired Kevin McKinley to lead our work in state policy and he, I think, will help us to take a real affirmative position in the legislative sessions ahead on what might actually be AI policy that's good for startups.

So instead of being in the position of saying no, 'cause we're sort of starting late and kind of with one hand behind our back. I think we're in a position to really actually try to articulate and advance a proactive agenda in AI that's compelling. I think Collin hit the main parts of it. Ensuring proper roles for the federal and state governments. Focusing on regulating harmful use, not development, and there are specific things that you can do there in terms of increasing capacity and enforcement agencies. Making clear that AI is not a defense to claims brought on under existing criminal or civil law and, and technical training, I think for government officials to make sure that they can identify and prosecute cases where AI is used in a harmful way.

And then all this infrastructure and talent stuff that, that Collin is describing, worker retraining, AI literacy. We've also given some thought to the idea that has been articulated by a number of lawmakers and was in the National AI Action Plan of creating a central resource housed in the federal government.

And you could also do it in state governments as well, that lower some of the barriers to entry for startups, you know, compute costs and, and data access. And we think that's really compelling in terms of ensuring that startups can compete. And that idea, like many of these is bipartisan, it's been supported by the current administration.

It was supported by leading Democrats over the last couple years. So that's the kind of thing that we are hoping that when we have the room and position to really advocate for an affirmative agenda that will get some traction in policy circles.

Collin McCune 00:53:46

We are not always in a hundred percent alignment with other people in the industry.

You know, and, and I, I, I think, I think that that's, you know, big, medium, little, you know, across the board there's other sort of like consumer advocacy groups that obviously feel differently about these things. I think for the most part. The industry is generally aligned on some level of a federal standard here and understanding that the thing again, that won't work is a 50 state patchwork.

And I think that that's super, super important because I think for the first time, you actually have this sort of alignment there. And if you have that sort of alignment, that's kind of momentum that you can to actually push things over the finish line and get something done.

And I, and I think, look, also, the Trump administration to their credit, has also been incredibly supportive of this idea too.

Matt Perault 00:54:31

There's a like. That's an incredibly important point. One criticism usually raised in sort of an implicit criticism sort of way is. "Hey, you're the little guys, but often you align with the big guys so aren't you just saying, aren't you just in favor of a deregulatory agenda that works for big tech?" And one of the things that I think is really extraordinary about the Little Tech Agenda is it's really nonpartisan and it's doesn't take a position on big little. It basically says, here's the agenda.

And when you agree with us, we'll support you. And when you disagree with us, we'll oppose you. And that's not party line. It's not big little. And so I think what we saw over a, the, the, the, the phase that Collin was referring to kind of initially in the recent set of AI policy was a phase of divergence between big and little.

Licensing regime, bigs were sort of pushing it, little was concerned about it. Then there's, then there was a period of convergence and I think actually if you look at like the National AI Action Plan comments across a range of different providers, as Collin's saying like a lot of them, they had some core similarities.

So, so lots of large companies have advocated for federal preemption. We don't oppose that just because big companies are advocating for it. We think that that's good for startups. I think it's possible. I'm curious. I mean this is really, you know, Collin really understands this in a way that I don't like, how the political chips will fall.

I think it's possible we're in a period of some divergence, and one thing that we hear repeatedly, which is sort of funny, is people will bring us stuff and they'll say, industry agrees with this, so we expect you to agree. You can't, industry already agreed, you can't disagree. And we say the big parts of the industry have agreed, but we, we sometimes we agree with them, but sometimes we have different views.

And so when we disagree, it's not because we're trying to like blow up a policy process or make it diff difficult for lawmakers who are trying to move something forward. It's because we're looking at it. We're looking at it through this particular lens. And I think. I hope it's not the case, but I think there might be more fracturing in the months ahead.

Collin McCune 00:56:25

Yeah, I agree with you on that. And by people, he means lawmakers, just to be specific.

Erik Torenberg 00:56:30

That's a great place to to, to wrap. Collin, Matt, thanks so much for coming to the podcast.

This transcript has been lightly edited for readability.

Resources

Read the Little Tech Agenda: https://a16z.com/the-little-tech-agenda/

Read ‘Regulate AI Use, Not AI Development: https://a16z.com/regulate-ai-use-not-ai-development/

Read Martin’s article ‘Base AI Policy on Evidence, Not Existential Angst: https://a16z.com/base-ai-policy-on-evidence-not-existential-angst/

Read ‘Setting the Agenda for Global AI Leadership’: https://a16z.com/setting-the-agenda-for-global-ai-leadership-assessing-the-roles-of-congress-and-the-states/

Read ‘The Commerce Clause in the Age of AI”: https://a16z.com/the-commerce-clause-in-the-age-of-ai-guardrails-and-opportunities-for-state-legislatures/

Find Matt on X: https://x.com/MattPerault

Find Collin on X: https://x.com/Collin_McCune

Stay Updated:

If you enjoy the show, please follow and leave us a rating and review on Apple Podcasts or Spotify.

Find a16z on Twitter: https://x.com/a16z

Find a16z on LinkedIn: https://www.linkedin.com/company/a16z

Subscribe on your favorite podcast app: https://a16z.simplecast.com/

Follow our host: https://x.com/eriktorenberg

Please note that the content here is for informational purposes only; should NOT be taken as legal, business, tax, or investment advice or be used to evaluate any investment or security; and is not directed at any investors or potential investors in any a16z fund. a16z and its affiliates may maintain investments in the companies discussed. For more details, please see a16z.com/disclosures.

a16z General Partner is joined by Martin Casado (General Partner, a16z) and (General Partner, ) to unpack the debate that lit up venture Twitter/X: should founders and VCs chase consensus, or run from it?

They explore what “consensus” really means in practice, how market efficiency shapes venture outcomes, why most companies fail from indigestion, not starvation, and the risks founders face when they’re too far outside consensus.

Timecodes:

00:00:30 - Consensus vs non-consensus investing

00:06:38 - Measuring market efficiency

00:11:04 - How consensus affects founders

00:15:22 - Is venture getting smarter?

00:23:19 - Defining “consensus” at seed stage

00:32:46 - Bigger outcomes

00:43:06 - How to frame consensus

00:49:02 - How to think about incentive alignment

00:53:06 - Has multi-stage won seed investing?

Transcript

00:00:30 - Consensus vs non-consensus investing

Erik Torenberg 00:00:30

So Martin it looks like you've helped spark a little bit of an existential crisis on venture Twitter, and I thought we'd all come here to talk about it.

Martin Casado 00:00:39

Great. Super. Let’s do it. I’m excited to be here.

Erik Torenberg 00:00:42

Why don’t we recap, Martin, from your perspective, what were you saying in that tweet?

What were you trying to say in that tweet? And then we can get into the great back and forth that you and Leo had and get into the conversation.

Martin Casado 00:00:54

So, let me paraphrase the tweet. The paraphrased version of the tweet is, it's dangerous to do non-consensus investing. Like that's a dangerous idea. The impetus of the tweet, which, by the way, wasn't well thought out, which I think a lot of the viral tweets happen to be not well thought out, is, you know, listen, I've been an investor for 10 years, I've done almost 200 investments, either as, you know, running the fund or being directly involved.

And it seems being blinkered to how VCs view companies is actually quite dangerous because you're so dependent on follow-on capital. And actually it reminds me a lot of being an academic. You know, I used to write a lot of papers, and like, you do all of this great research, but when you wrote the paper, if you didn't actually think about how the program committee would view it, like it wouldn't get accepted, right? It felt very similar to that.

And so that was the origins, but I wanna be very clear. I did not say, and I would never say consensus investing is a good idea. I'm just saying not being aware of consensus is a bad idea. And I think the underlying, the last thing I'll say on this, I think that my underlying belief is early markets are actually pretty darn efficient, a lot more efficient than people realize.

And so if you're alone in your view, you may just be missing something.

Erik Torenberg 00:02:20

Leo, we're stoked to have you join us as a friend, fellow venture nerd. What was your reaction?

Leo Polovets 00:02:28

Yeah, I mean, I actually agree with a lot of what Martin just said, which is, eventually you have to get to consensus, whether it's when you're investing or later because otherwise, you know, if you're dependent on capital markets, it's very hard to keep the company alive if nobody wants to fund it. I would say, like for me, and maybe we invest like a tick earlier, more like towards pre-seed and seed, but for me a lot of my best investments have been more on the non-consensus side.

Not in terms of, I had some crazy good insight, and nobody else had it, and like I'm just brilliant, but more like, you know, these companies often struggled in the early days because before there's proof points, it's not obvious that it'll be a good idea. And then once they get good, the valuation skyrocket so fast that like, you could still get good multiples, but they're just much lower than at early stages.

Erik Torenberg 00:03:17

Yeah. And then there's sort of broader commentary on looking at a list of big winners over the last, you know, 15, 20 years and saying, “Hey, what was consensus, you know, which were consensus, which were non-consensus. And then Martin, what were your reactions to that sort of broader commentary?

Martin Casado 00:03:36

Well, listen, I mean, again, it wasn't meant to be a technical tweet where like the wording was exact. On the face of it, it’s almost an ill-defined statement because we don't know what consensus means, right? And so then everybody picks apart the consensus. But here's my reaction to the list of like the Airbnbs and this and that, which is, I think we need to be very careful not to conflate a company having a hard round with market consensus, right? Like if you look at the list Keith Rabois put out, which is great, and I love Keith, I mean these are like MIT founders, known spaces. Like I'll bet if you took like the median value of their raises over the lifecycle of the company, I'll bet they're way above market.

Many of the companies were YC companies, and so I just think it's so easy to come up with these anecdotal, “Oh, this one company had a tough raise,” when that's definitely not within the spirit of what I was trying to say, which is, markets are actually quite efficient. If the market is efficient, and it's a good company, the price is going to be high.

And if you don't recognize that, then you're probably beating yourself as opposed to the market, right? And so it really comes down to, you shouldn't be looking for good deals with respect to other investors. You should be looking for good companies, and price shouldn't sway you from that. I mean, that's really at the heart of this, and so I just don't think that list, unless you actually run the numbers, which we haven't done, I just don't think it demonstrates that the idea that consensus is important is wrong at all.

Erik Torenberg 00:05:22

There are a few quibbles I had with some of the names on that list, like some people put Anduril, and it certainly was a controversial investment. But you know, Palmer Luckey, you know…

Martin Casado 00:05:32

I mean, second time founder, billion-dollar exit, Trey, who's phenomenal.

You know, Elon, this is in the shadow of Elon, who shows that you can already create these defense tech companies. I mean, if that's our definition of “non-consensus,” it just shows how insular we are as a community. I mean, it's almost an indictment of us that we even make this list.

Erik Torenberg 00:05:52

And wasn't the seed round at like a hundred or something? Like there was a very expensive early round.

Martin Casado 00:05:56

Every round was super expensive.

Leo Polovets 00:05:59

I'm not sure an ex-unicorn founder would ever be non-consensus really.

Erik Torenberg 00:06:04

Yeah. It is interesting because there's also sort of, you know, there are rounds that are maybe, you know, non-consensus at 10 million or something, but then become, you know, super hot rounds at 50 or 100 and then become $10 billion companies or $100 billion companies.

And even if you invested at that, you know, consensus round you 10x’d or 100x’d. And so it's sort in the face of, “Hey, if it's a hot deal, that must mean it’s not good.” Peter Thiel once had a line, which was like, “The faster and higher the up round, the more you should invest because it's like, you know, working.”

00:06:38 - Measuring market efficiency

Martin Casado 00:06:38

Yeah. I would love to do a correlation analysis. Actually, Leo and I had, I thought, a very interesting discussion on trying to figure out how you'd actually measure this, how you'd actually throw some data at it. We actually have an analyst working on it now, like the data isn't ready yet.

So I have a new one actually. I want to test this with you, Leo, on a good thing to test. So I’ll bet the best prediction, the best correlate of a high up round outside of the business is the fact that the previous round was hot.

Leo Polovets 00:07:08

I think that's probably true.

Martin Casado 00:07:11

And if that's the case, it would suggest that the market is actually pretty efficient because it's almost inductive that like the previous round knew that the next round was going to be hot.

Leo Polovets 00:07:21

So, I do agree with that. I think the question for me is like, “Where is there more opportunity,” right?

Because if there is, you know, like the 5 hot companies keep having great rounds, and then there's like 10,000 not hot companies, but a 100 of them will become hot over time, even though the odds of becoming hot are low, most of the hot companies end up coming from the not hot batch, right?

Martin Casado 00:07:45

Right, so the question comes down to is it easier to spot the company nobody sees or get into the company that's obviously good? And maybe even further than that, which is, to what extent do even high price rounds under-price hot deals? Because if I'm right, if the view is correct, that hot deals are hot because they're good companies, not, like actually the market is very efficient, and that drives the most of returns, then I think the next obvious question is, well, if that's the case, then the market isn't that efficient because it's underpriced the company, right? If the majority of returns are in high-priced rounds, and the market has underpriced it, but I think risk-adjusted, that's not necessarily true.

Which is, it could be still priced right because there's still chances it goes to zero. So I guess my sense is until we run the numbers, we're not going to quite know the answer. But I think a lot of these theories prove out pretty anecdotally, and I think maybe that's the problem is there's kind of an anecdote for every theory.

Leo Polovets 00:09:08

Yeah. I think the basket analysis is probably the most interesting one, right? Of like, not how did this one company do, but how did this, you know, portfolio of companies that raised like really quick follow-on or had like 10 term sheets at this series A, like how did those end up doing over time?

Martin Casado 00:09:23

There are even cases in my portfolio where a super hot company, from an investor standpoint, so, many term sheets, the business didn't work out, you know, at the level that you would kind of expect, but the outcome was still really good. And so in some level, even independent of the productive asset, like human opinion about it matters.

So there's almost two ways you can slice this conversation. One of them is like, the asset is what's productive and produces the value, right, and the market will determine if that's valuable or not, right? So that's kind of this productive asset view, and that's kind of the one that I hold, which I think that actually investors are very smart. I think that they know which companies are good, and then they pay for those. That's kind of my view. But that's the productive asset view. But there's another view which is, independent of whether the company is good or not, there are things that people think are good, and so you're almost like playing to like the human perception of the company, independent of the underlying business.

And I would say, again, anecdotally, until we run the numbers we won’t know, that also seems to be a bit true.

Leo Polovets 00:10:36

So I've been in Venture for like 12, 13 years now. I've definitely seen this in sectors where like sectors have fallen in and out of favor, right? We had like e-commerce was hot and then it was dead, and then like Dollar Shave Club got acquired, and it was hot again.

And it's like e-commerce didn't, I think the fundamentals didn't change that much year to year, but like the valuations and the like appetite for investing and maybe starting companies changed a lot year to year. And so that to me is sort of an indicator, like it's not just the fundamentals. There are all these other like forces as you mentioned.

00:11:04 - How consensus affects founders

Erik Torenberg 00:11:04

One other part to your tweet, Martin, that I think was underappreciated was sort of the risk to founders of being seen as non-consensus. Because founders need to raise money, and they need to raise follow-on funding, you know, within 18 to 24 months, you know, sometimes even sooner.

And so if everyone is passing on you or people are bragging about how, you know, other investors don't want to do your deal, that’s not going to be super helpful to you in your next round.

Martin Casado 00:11:36

I actually think the most interesting aspect of the tweet was like the sociological study that followed of like how different people interpreted it, right?

Like the tweet itself was pretty banal, right? It's just kind of a non-statement. It's almost tautological. But like different constituencies viewed it very differently. So like, I would say relatively inexperienced investors kind of used it as an opportunity to be like, “Oh, Andreessen Horowitz consensus invests,” which anybody that knows anything about our investments knows that’s just totally not true. I mean like even my own portfolio, many of the top deals I've done, nobody else was in the deal, you know, etc, right? So like this is a statement about consensus investing. So that was one cohort. There was another cohort like Leo and Keith who have a lot of data and they have a lot of really interesting good things to say, and that ended up being great discussion.

I think there's still a lot more to do there, but most of the founders, and I got a bazillion DMs who’re like, “You're totally right.” So the founders clearly feel this tension that it’s dangerous to be non-consensus because, you know, they have to cater to VCs, and they know it, and they see the pattern match responses.

They deal with this all of the time. And so from a founder perspective, it's like, you almost have to be non-consensus to have alpha in the actual product market, but you have to look consensus when you’re raising. And I think that's actually probably right.

Leo Polovets 00:13:08

I think this is probably one area where I differ a bit. I think there's benefits to being non-consensus because from the company side, I think when the money is hard to raise, you tend to be more frugal with it. And then also, if the next round is less certain, like I think there's less of a sort of like, “it could crumble at any moment” aspect, right?

Because I think when it is hot, and you're raising subsequent rounds very quickly, unless the assumption of things will go perfectly, if anything slows down, it's like, “Hey, now everything, like now you can't raise any more capital all of a sudden.” And I think if you're in the mentality of, you know, growing quickly and spending, I think that's pretty hard.

On the flip side, if you're non-consensus, it tends to be like, you tend to be more cash efficient, tend to be more frugal out of necessity. I think the other side is, it depends on the form of like consensus, but sometimes there's also like much softer diligence. Like I think the worst form of consensus I've seen is like, “Oh, Sequoia, Andreessen did this round. Like, let me just do 2x markup in two weeks because I want to be in the same company.” And then like, there's no diligence there, right? It's just like, oh, this is hot, let me do it. But I think then like maybe you're overlooking like, is it actually a good business? Like, you know, Sequoia and Andreessen and we all make good investments and bad investments and, you know, so it's like maybe this is one of the bad ones and you're just marking it up because you want to be in the hot deal and like that ends up not being good for anyone.

Martin Casado 00:14:25

I think this is a tremendously important and good point. I tend to believe now that most companies fail from indigestion, not starvation, which is they just raise too much money too easily. They don't listen to the actual market, which is, you know, the customer base. And as a result they just have a bunch of bad practices and end up running out of money.

And I think that there's a lot to that. I actually think in 2021, if you look at, if you just did a study of that cohort, the companies that had these billion dollar Bs. If you remember that time, it was totally crazy. I'll bet that's probably one of the biggest wipeouts of capital.

So I definitely think like consensus investing is definitely very dangerous, and only leaning into this for a founder is definitely dangerous. But I also think the flip side is true, which is if you're totally blinkered to it, I think your life is pretty tough.

00:15:22 - Is venture getting smarter?

Erik Torenberg 00:15:22

And so there's a broad question as to like, you know, of the companies that do win, you know, how many of them are sort of competitive rounds versus not competitive rounds.

And sort of what is the duration between them being non-competitive rounds and then becoming not non-competitive, and what percentage are really able to sort of… And one question I have is like, “Is the market getting more efficient over time?” A lot more investors, you know, we should be getting smarter as an asset class on how to evaluate these companies, a lot more capital.

Are we getting better? And if so, what does that mean?

Martin Casado 00:16:02

Well, I'd love to hear Leo's view on this.

Leo Polovets 00:16:05

It’s something I've been thinking about for a while. My take would be that for non-consensus companies, it's getting more efficient because the more investors there are, the more likely you are to find at least one or two that like what you're doing.

I think for the consensus companies, it’s starting to get more inefficient, right? Which is like when you have 10 term sheets, you get, you know, 5x the market, like what maybe the fair value should be. And then like, it's great for the founder and maybe again, it’s a little bit more of a house of cards if things go south at all.

But it's also like, it's not necessarily great for investors, right? Because you might have to pay 2, 3, 4x over like the actual intrinsic value of a company, or like the likely future value of a company in order to get in.

Martin Casado 00:16:48

But that would be actually efficient, right?

The price is actually approaching the return profile. From a market standpoint, that'd be efficient. I mean, it sucks from an investor standpoint because prices go up.

Leo Polovets 00:16:59

Yeah, that's what I'm saying, right? Like for founders it's getting hyper-efficient. Or maybe there’s such an imbalance for like, the really hot companies that, you know, maybe your price gets bid up way past where it should be.

And similarly, for non-consensus companies, it's the opposite, right? Where like, there's not enough investors, so your price is lower than it should be, perhaps, right? But I think for me, like those two are kind of opposite ends of the spectrum.

Martin Casado 00:17:25

I totally agree. This is a great question. So I think we can all acknowledge that there's a failure mode where the consensus gets bubbly. And then companies raise too much capital and then there's a bunch of wipeouts, right? So that has always happened. That will always happen. So that’s just part of the market.

We, I think, can also all agree that there's parts of the market where there's probably unnecessary pessimism. So for example, right now during this AI craze, like, you know in my area of infra, a lot of the traditional companies that, you know, two years ago would be great, like can't even raise right now, just because they're not in the sweet spot.

And so I think that will always be an aspect of the market too. But in general, for the mean investment, I do feel like the market over time has gotten a lot more efficient. Meaning, you know, we can deploy more dollars with more regularity, and the price is converging on what will ultimately be a fair price.

You know, this is acknowledging both of these failure modes on either side of this.

I mean, we're seeing one right now. I mean, this is the reality, there's AI companies that clearly are raising, you know, speculative money where nobody even really understands the business model, and there's great companies that can’t get investments. So we're seeing this right now, but I will still say the reality is, OpenAI has grown tremendously, and Anthropic has grown tremendously, and Cursor has grown tremendously. And so like there is some underlying market signals to fuel the chaos.

Leo Polovets 00:19:14

Yeah, I think part of it's like if you ever look at vintage year data for venture funds, it's probably a good way to see how consensus and non-consensus do over time. Because when you look at like the dot-com bubble years, like, I think the median fund was terrible, and I think it's just like, “Hey, everyone overpaid, and then the companies weren't worth that.”

And so like, you know, even though everything was hot, like it didn't do well, and then a lot of the funds didn't do well. If you look at the Airbnb, you know, Uber like 2010-ish era, it's kind of the opposite area where like, I think the top quartile funds like crush it, and it's because the market was pessimistic.

And so if you were willing to invest and like you had a different opinion, you did really well. And now it's probably kind of somewhere in the middle.

Martin Casado 00:19:54

Maybe I'll just go through like kind of my own startup just as kind of a single anecdote to frame the conversation a little bit.

Right, so, you know, I did my PhD at Stanford. I was a classic, you know, take the research, do a startup. You know, we had so many term sheets before we had any idea of what we were doing. And it was like the hottest thing ever, and it was great. And so we did a seed fundraise, actually Andy Rachleff, you know, Benchmark, Andy Rachleff joined my board and, you know, we raised at the time, which would've been a super, super high price kind of seed round, which is 10 million post. This is in 2007.

Then the market tanked in 2008, and we still didn't know what we were doing. And you know, it was just a bunch of researchers. And so we couldn't raise any money at all. I mean, Sequoia very famously, you know, gave us a black eye and, you know, we couldn't raise. And then, you know, as we started to come out of the recession, Andreessen Horowitz, NEA, Lightspeed, a few got very interested. And then we had a pretty hot round again. With Andreessen, it was actually over the market price, even though the business wasn't quite working, but it was starting to see signs of life. Then we had an incredibly hot round because the business started working, and then when we actually sold the company, I mean it, you know, it returned a fund, you know, it was one of the highest acquisitions on multiples of revenue at the time in enterprise software.

And so you kind of ask the question, “Was the initial flurry of interest warranted or not” because it turns out like we were probably a month from going bankrupt and we actually didn't know what we were doing. And the company definitely wasn't working. And then actually what we had pitched at that time didn't make any sense.

We were going to change, you know, switch hardware, which didn't make any sense. And so there's one view that's like, the market was over-exuberant. You were lucky. There's another view that says actually the initial conditions were there to do it. I just feel like if you run the data, it just seems like the companies that have good outcomes did have sufficient interest along the way because there are enough signals to do it.

Leo Polovets 00:22:13

I think, at least on my side, for a lot of pre-seeds and seeds I’ve done… I went back, I think over my top like 10 investments, maybe 6 or 7 or 8, took a month to raise a seed round. And a lot of times, like a lot of passes, like they were all down to the wire, but then they ended up doing better over time.

I think that transition from like non-consensus to consensus ended up being really important because if you never transition, it's really hard. And if you're always, you know, if you're always consensus, that's great for you. But like, one thing I noticed that was interesting is a lot of the companies that struggled, obviously some of them just go to zero, right?

Because they struggled because the business isn't that great and people recognize it. But the ones that did well, a lot of times the gap between like the C and A or the A and B was literally like 20x or 50x, right? And so I think part of it's like, as an investor you can still get good returns at like the series A or B in those companies, but it's just, I think it's so different to invest at the seed where there's like a 1000x versus like the A at a billion, where now maybe there's still like a 10x or 20x, which is very different.

00:23:19 - Defining “consensus” at seed stage

Martin Casado 00:23:19

So I've got a question for you, Leo, because I think that you play a bit of a different game than we do, which is, so if you have a seed, which is, let's call it non-consensus, and again, we're using this very vague definition of a consensus, but like, you know, they're having a tough time raising, you’re the only person putting money in.

Do you have a theory on how it will be consensus, or is your belief that the underlying productive asset is going to do very well, and that by definition is consensus? Do you see the question? So the question is, is it, this is just true belief in the underlying business? I mean the ultimate sign of success is just the business is really working.

So are you like “For the next raise, the business will definitely be working?” Or do you have some other theory on what will attract the investors?

Leo Polovets 00:24:07

I would say it's often the latter. I would say, and especially true these days because I'm investing more in deep tech companies, and so at seed it's very rare to see like, “Oh, there's an asset that's gonna be working here at the series A because usually the asset is still going be like being developed at the series A or maybe series B.” I think what I'm looking for is like, there's maybe not enough here for somebody to write a $5 or $10 or $20 million check, but the company has milestones that I think if they hit them, then it would become, you know, sort of consensus enough to merit a check of that size. And then I'm basically trying to evaluate like, okay, the company has these milestones. Do I think they could hit them or not? And also if they hit them, are they compelling enough? But I think that's sort of the big, you know, investment wager.

Martin Casado 00:24:49

Yeah. So in this case you do think about like what the follow-on thing is going to want to see. You have reached a conclusion for the current round that is non-consensus.

Leo Polovets 00:25:03

And I would say like the consensus piece is part of it in that I definitely meet companies where they're like, “We're raising 3 right now, it'll help us do these milestones, and then we think we can raise 10.” And then there's other ones where, you know, it's like, “We're raising three now. We're gonna hit these milestones, and then we wanna raise like a $50-100 million series A.” And that's actually a much harder bet, right?

Because you have to assume they're going to be consensus by the time they raise the next round, and it's going to be like a top 5% series A, and that's a hard bet to take. For the companies where the capital needs are more modest, or they have like a more traunched roadmap planned, I think it's a little bit easier to, you know, to predict like, “Hey, would these milestones be enough to raise 10?”

Like a lot of times, I don't know if it'll be enough to raise 100, like probably not, but 10 feels like pretty feasible if you do the things you think you're gonna do with this 3.

Martin Casado 00:25:49

Has your kind of view on this shifted in the last… like do you find this AI wave to be different than previous waves or fairly similar?

Leo Polovets 00:26:00

I'm probably a bad person to ask actually. I haven't invested much in AI because of the deep tech angle, so I see maybe like 10, 15% of my companies are pure AI. Others obviously use it in some way, but that's not the product.

Martin Casado 00:26:13

Well how about deep tech then? Because I think that's also, you know, like pretty different than what we were all investing in 5 years ago.

Leo Polovets 00:26:19

So maybe on the AI side, and I’ll touch deep tech next. I think AI is interesting to me because on the one hand, I've never seen faster growth, right? Like people talked about the like triple, triple, double, double, double thing for a while of getting from a million ARR to 100 in 5 years. And that seems so antiquated now, right?

Like most companies are doing it in like one or two years.

I think on the flip side, the endurance or like how long those companies endure, last, and grow feels like much more of a question mark. Because in the triple, triple, double, double, double ARR, like if you hit 100 million ARR, and there was no one close to you, you probably just keep growing.

And now it feels like you could hit 100 and then you drop to 50 because someone else came out with a better product. Yeah. So, you know, I think there's like, the growth is amazing, and then the moats are weaker. And so I think there's a counterbalance there, and I'm not sure how I'd evaluate it because I haven't invested that much of that stuff.

Martin Casado 00:27:08

I agree. Yeah.

Leo Polovets 00:27:09

On the deep tech side, I definitely see areas with a lot of hype from time to time. For example, we invested in defense a lot, 3, 4 years ago. And then we kept looking, we basically paused for a year and a half or 2 because after the Ukraine and Israel thing, you know, prices just went up like 2, 3, 4 times, but the company fundamentals didn't change. And then it started being an opportunity cost of like, should I invest in this defense company at 40 when there's this really great, you know, energy company at 15? And so I think defense was kind of like that. I think bio has had a lot of ups and downs.

I think in robotics, like humanoids are probably one of the most hyped areas where the valuations just get crazy before there's any, any revenue. I feel like I kinda lost the thread in the original question, but…

Martin Casado 00:27:55

No, this is great. I mean, I was honestly just wondering how you thought about this current wave, and you did a great survey of the set of the waves and I actually agree.

Leo Polovets 00:28:02

I would say like for the consensus areas, like humanoids, like we end up not explicitly, but implicitly avoiding them because once you have a few companies that have raised like hundreds of millions, whether they end up being great outcomes or not, I think it's pretty hard for someone to start something new with like, you know, near zero resources and team.

Martin Casado 00:28:20

Yeah. You know, it's interesting when you do the humanoid, so I think there's all sorts of types of investing, and they're all pretty valid. One type of investing is humanoids are clearly interesting. Big companies are clearly interested in it, so why don't you back a bunch of good teams? And worst case they get acquired, and I think that's totally legitimate, but that's not how I think at all. Like for me, like the company has to make sense as a standalone business at scale. So things like humanoids are tough for that just because the unit economics right now are just so unknown. Like competing with a human body is a very, very hard thing to do.

And then of course you can be like, “Okay, well, you know, we'll put it where human beings can't go, like a car factory.” But then all of a sudden now, you know, you're building a manufacturing company, you know, so you verticalize heavily, and the company has to look at kind of whatever sector that the robot is going into, and it's more constrained and I don't understand the competitive set and, you know, yada yada yada.

So I just feel like, from my standpoint, the idea that this is very buzzy and hot, you know, in the industry for big companies, and it may have an M&A, I don't know how to invest that way. I just don't know how to handicap that. And so the way that I tend to view these things, I mean like for AI, for better or for worse, like you have great unit economics.

I mean, you know, everybody knows kind of like, and we always talk about the OpenAIs and the Anthropics, but take like ElevenLabs, for example, or Midjourney. I mean these are just famously model companies where the unit economics are great. They're grown, you know, very quickly. And so I understand that.

But you know, I think there's kind of been this weird, and this happens a lot where people take the example of these model companies, and they apply it to totally different spaces where you don't have the proof points, you don't have the economic case, and they kind of apply it. And that's one thing I don't know how to do.

So certainly I don't believe, you know, we should all just follow the common consensus around areas to invest in. But I do think that like, there's going to be a pool of capital and it's going to want companies to look a certain way. And if you don't consider that when you're investing, I think life will be a lot more difficult.

Leo Polovets 00:30:34

Yeah, I agree. Sort of an aside here on the humanoid stuff. I think what I've seen over the last, like 10, 15 years is if the market is big enough, it really distorts like VC investing because, you know, it used to be that. You look at a market, you're like, “Oh, it's a 2 billion a year market. If there's a 1% chance they could capture it, they'll be worth this much. So let me justify a seed price.” If the market's like $5 trillion of human labor or something, like any price makes sense, right? But then, but then I think that really just starts with like “How much value is there in this?”

Martin Casado 00:31:05

The most boneheaded partner meetings were like, “Well, yes, it is cold fusion, but this is the largest market ever. So on the off chance it works…” I'm like, this isn't engineering, man. This is like laws of physics. I'm not sure that like, you know, a good, you know, software founder is going to bend the laws of physics. But yeah I think I totally agree. I also feel like, I don't want to harp on this too much, but like unit economics is so important.

I mean, like, what is the story for autonomous vehicles, right? The story for autonomous vehicles is that even after the industry has put $100 billion in it, 100 billion, the unit economics are still, you know, let's call it on par with Uber. Let's just call it that, right? And so does that make sense for venture investment?

It’s really hard to build a standalone business with those types of economics. I mean, Google can do it, sure. And Tesla can do it, sure. But can startup X do it? No. And so you're either playing for, this is a great company that got acquired, which a lot of that happened, and people made good money.

But that’s, again, that’s not saying that, you know, the startup… Or you're building picks and shovels. Like Applied Intuition, where you're like, you're building software for this market. But I do think that a lot of investment dollars do follow these spaces where there really is no thesis on the ultimate unit economics.

And I think you're exactly right. I just think that there's this kind of Market TAM sloppiness that says, “Well, if the market's infinite, then the expected payout is high.”

Leo Polovets 00:32:41

Also infinite.

Martin Casado 00:32:44

That’s right, also infinite, exactly right.

00:32:46 - Bigger outcomes

Erik Torenberg 00:32:46

When I look at my portfolio, I see both… Some of the winners, Pave and Scale, were non-consensus, non-competitive, you know, unproven, but very talented founders. And then on the more consensus, competitive Jack Altman and Qasar were…

Martin Casado 00:33:04

Wait, how is Scale non-consensus?

Erik Torenberg 00:33:07

At seed. You know, Alexandr Wang was 18.

Martin Casado 00:33:13

It's a total known space. He's phenomenal.

The A was done by Volpi, who's amazing. I mean, I just feel like this is a very narrow definition of “non-consensus.”

Erik Torenberg 00:33:25

Sure. For nearly most of the rounds, it was competitive. So yeah, I can agree.

Martin Casado 00:33:33

I mean Dan Levine was… I mean, come on, these are like the best investors in the world.

Erik Torenberg 00:33:35

I brought the example to say that Qasar’s round was almost an order of magnitude more expensive. And I think what people have been late to really internalize and what a16z was super early to internalize, was just the outcomes are an order of magnitude bigger, maybe two orders of magnitude bigger.

And so you can get seed-like returns at, you know, order of magnitude or even two orders of magnitude more expensive. I mean, remember, YouTube, Instagram were considered, you know, very expensive acquisitions at, you know, just a few billion dollars. And you know, in a few years, we're gonna have more trillion dollar companies. And so, once we truly internalize sort of the outcome expansion on the order of magnitude, I think it makes sense to Leo's earlier point, that then it would beg the argument of like, okay, but can you have, you know, 1000x like returns at not just what we used to consider seed-like pricing, but maybe at series A or maybe even series B?

Martin Casado 00:34:29

Well this is a very interesting question because you actually do run into fund mechanics as an actual, you know, price modulator in this discussion. So you're exactly right. So again, I'll go back to my company. So my company was acquired for $1.2 billion. We had, let's call it, you know, less than 10 million ARR, right? So does that make any sense? No. And then a lot of people are like, this is totally crazy. This makes no sense. Except for when I left, you know, three and a half years later, like the run rate was, you know, $600 million within VMware, who acquired the company.

And then right now it's, you know, let's call it 2 billion, right? It was actually, at one point in time, it was, I think it was 40% of the growth of VMware, like the business unit that I ran, that was part of the acquisition. So clearly it made sense to VMware. So as a result, you should say all the check sizes should be high for the winners because the outcome was so good.

And this actually returned a lot of money to a lot of investors. The problem with that is I just think that that would mean fund sizes would be too large and you'd have to unlock different pools of capital, which by the way did start to happen during the SoftBank and the Tiger and the Coatue era.

So you could argue that all of their theses were correct, right? Like SoftBank was actually right and Tiger was right, and it was actually a macro issue that caused all the pullback, and that's gonna come back again. I mean, that's, I think, a very legitimate thesis. But I really feel the reason that prices don't continue to go up is more just access to LP capital.

So, Leo, let me just try to, to make this a bit more concrete, which is, I think what Erik said is correct, which is the outcomes are so big, it suggests the prices are too low that we actually pay. You know, the fact that we get the returns we do suggest the prices are too low.

So the question is, why are the prices too low? And I think the answer is like, we just don't have the dollars to place all of those bets. And a number of people have actually questioned exactly this. Very famously SoftBank questioned this. Tiger questioned this, and so they raised these, you know, Insight questioned this, they raise these huge funds, and they deploy a lot of capital.

And those experiments had very mixed success, but it's not obvious to me that the reason they had mixed success is because the prices were too high. I mean, there's a lot of reasons why those could not have worked, including kind of macro cycles. And also the fact that none of them were Silicon Valley insiders, none of them were, you know, traditional early stage investors, etc.

So there's a very reasonable question, which is, you know, maybe someone should just go run the Tiger strategy again, but as a Silicon Valley insider.

Erik Torenberg 00:37:10

There's the failure cases just to some degree. You know, I mean, Thrive raised bigger funds, Founders Fund raised bigger funds. We raise bigger funds. You know, the winners have also, you know, multi-stage, have raised bigger funds.

Martin Casado 00:37:24

It just could be that this is just the market being efficient. Like actually the reason that more money is going into this and the funds are getting larger is because the opportunity set is larger and this is just a market working its way out.

But Leo, you're very quiet. This is actually a pretty controversial statement, so I wanna make sure that like…

Leo Polovets 00:37:41

I'm not sure what you mean by “We should be paying more.” Do you mean that, like, you think the current prices are still like well below where they should be and I guess if so, like…

Martin Casado 00:37:56

I'm riffing off of Erik’s statement, which I thought was right.

Which is venture capital has been a top returning asset class. And you can look at individual investments. If you just take the top 10 percentile of funds, you know, they return so much money. So there is an argument that even with these high prices, they're still underpriced

Erik Torenberg 00:38:18

And to put it differently, Leo, it's like, a seed fund may say, “Oh, I'm not gonna invest in something at 50 post or 100 posts because I don't think there's a 1000x, you know, potential. I don't think Anthropic is gonna be a $100 billion company, or you know, OpenAI is gonna be a $100 billion company” or whatever it is.

But it turns out it is.

Leo Polovets 00:38:31

I think it would be contrarian now to say OpenAI is going to be a $100 billion company.

Erik Torenberg 00:38:41

Right, exactly. I mean a few years ago. And so, it doesn't seem like we've sort of truly internalized that this is the norm, that there's gonna continuously be a $100 billion outcomes.

Martin Casado 00:38:59

Or that the market just continues to grow and therefore it necessitates larger fund sizes. I mean, I would say that probably the venture market was what, a hundredth the size, 20 years ago.

Leo Polovets 00:39:10

Probably something like that. It's kind of wild to think about.

Erik Torenberg 00:39:13

Yeah. And we did think a few years ago that there'd be a great contraction in the asset class. That 2021 was a blip and that, you know, it would sort of right size back to where it used to be.

And it doesn't seem to be the case that it's going to 2010 levels. I’m not sure if you guys have the data on you, but when, when I talk to our team, when I talk to Thrive, it seems that people think, no, more capital is just gonna keep entering.

Leo Polovets 00:39:39

I think that's just because companies stay private longer too, right?

But I think the actual number of $100 billion plus companies in the last 20 years is pretty small. Like, I don't know the exact number, but I bet it's like 10 or 15 or maybe 20 or something. So it's like you're really betting you can get like the one every year or two that gets there.

Let's say you're doing a series A at like a billion post or something, right? And you want a 100x, even ignoring dilution.

Erik Torenberg 00:40:09

You'd have to bet that there's more of them, more of them are going to happen, and that there are also more ways of getting liquidity from them.

Martin Casado 00:40:19

But that also kind of suggests purely by the numbers that the most important thing is, if you can, the most important thing is being in one of those independent of price.

I mean, that's the high order bit.

Leo Polovets 00:40:34

I mean, I generally agree, right? Like if you're in like the best company of the year, I don't think ownership matters that much. I don't think like the price matters that much. If it's gonna be the best company, it's like 10 years forward.

I guess to your earlier point where, you know, if venture funds had more money, they like do higher valuations. I mean, it sounds like then you could do the higher valuation today too, though, right? You could just be like, “Hey, if we just want to get in this one, we'll pay twice the price and get half the ownership” or something. Right?

Erik Torenberg 00:41:02

But you also need a diversified portfolio. You need enough companies to get…

Martin Casado 00:41:05

You need the fund size to run that strategy. This is why I think a lot of this comes back to fund size. I mean, even in the Andreessen portfolio, I was just thinking off the top of my head, we have 3 companies that are—4 companies at the $100 billion mark, right?

I mean, there's Stripe, Databricks, Coinbase, OpenAI, and so they're not that rare.

Leo Polovets 00:41:26

You guys have awesome coverage. I guess the question is like, how many more could you name though from the last like, you know, 15 years? My guess is 10, 15, probably not like a hundred, right?

Martin Casado 00:41:36

Yeah. I mean 20 billion plus, there's a lot. And that used to be so rare. And enterprise software, it used to be an adage that nobody ever broke, you know, 20 billion or 10 billion, right. And Palo Alto networks was at 15, and we were like, “This is crazy.” Now there's so many of them that have, and so maybe with a hundred billion, you're right.

But in the world that I live in, the amount of like decacorns is probably an order of magnitude more than what it was 10 years ago. And on the face of it, that would argue for an order of magnitude higher fund size, if you wanna play the strategy of being in the winner. I mean, there's clearly multiple strategies. For me, the key question, I don't know the answer, I want to run the numbers, is if you take a dollar of earnings, like a dollar of earnings for a venture capitalist, did that come from a company that raised at high prices or not?

I would guess the answer is “yes,” just because the winners are so outsized.

Leo Polovets 00:42:36

I mean, I'll say there's like multiple ways to play it, right? Which is if the outcomes are 10x bigger, you can have a 10x bigger fund and basically run the same playbook, keep the same ownership, and like a big outcome still returns the same amount of the fund.

You can also do more investments at, you know, a fraction of the ownership, and then, each investment maybe moves the needle less, but you have a higher chance of hitting like, you know, the Stripe of the year, the Uber of the year. So I think there's definitely different models that could work here.

Martin Casado 00:43:04

Yeah, that's a good point.

00:43:06 - How to frame consensus

Erik Torenberg 00:43:06

I want to make a few related points here. So one is, I remember someone quote tweeted Martin’s tweet and said, “This is a sign that the asset class is dead” or something. The idea of a more efficient market.

And I think what that really means is more that that individual's firm is… If an individual firm can't compete and win deals in an efficient market, they're going to lose. And so, my second point, which is I think a lot of venture capitalist identity is tied in being non-consensus in being able to see things that others can't see because it's hard to win against all these other much bigger, much more well-funded players. And for that reason, I less want to use the term “consensus,” “non-consensus” because it's so core to people's identity, and more want to use the term like either it's a hot round or it's not a hot round, you know, it was a competitive round or not competitive.

And I think another way of framing that, it's not perfect, is, is the company working or is the company not working at the point of investment? And let me add some nuance to it, which is, if something is working, then it's like, you know, what is the price, and what is sort of the, you know, potential return multiple and how does that work with your threshold, etc.

There's some things that are competitive and not working, but have an incredible founder, whatever. It's early enough that people believe the vision and so you're still paying that price based on what you think. And then there's lots of things that are not working, or not obviously working. And we’ve chosen to do less, I believe, consumer things that are pre traction. So it's basically, it's like, do you want to invest in things that have traction or no traction. And there's failure modes with both. And not every hype thing, not every competitive thing has, has traction of course, but it's just another way of framing this. I'm curious, feel free to quibble with my framing.

Leo Polovets 00:45:03

I think I saw the same quote tweet. I'm probably somewhere in between. Like I don't think venture is dead. I think it gets a lot more fun if it's purely consensus. The reason is, I think in a purely consensus world, like it all just comes down to the cost of capital.

And so if my LPs want 5x and yours want 2x, you could pay two and a half times higher prices, and the company's not better. It’s just like, “Oh, like your cost of capital is lower, so you're going to win all the time.” But also it’s like we all see the same value. Everyone sees the same value of just like, who wants the smallest return that can still stay in business.

And that just feels less exciting to me.

Erik Torenberg 00:45:42

That’s exactly right.

Martin Casado 00:45:45

I'll get a little bit philosophical on this, but like, the thing that's always bugged me about PE investing and public market investing is it just doesn't care about productivity really.

I mean, it does to some degree, but I just like, you know, if you're in a large public company, like I was, you realize that the public markets really care about predictability over innovation. And so innovation has stifled so much, and in fact it kind of causes large companies to protect themselves through kind of incumbency and monopolistic practices and everything else, just because they're not allowed to be aggressive on growth.

So I feel like it's almost this negative force on progress and innovation, and I don't want to be too dramatic about it, but I just feel like… I'll bet if you draw a dollar at random, that gets invested, you know, 90 cents of that dollar goes into like keeping incumbents alive and/or, you know, predictability, and not to growth.

And I'm a huge believer in creative destruction, man. I'm like, “Fuck, man, get him out of the way. Let's invest in growth.” And so I love the idea of venture as an asset class getting more efficient. And I love the idea of more money going into it. Because the entire thesis is growth. You never invest on, I don't, I mean, I'm sure you don't Leo, I never invest on downside loss. I don't care. You only invest on upside. And so to me, more dollars going into ventures is only a positive for humanity. And again, I don't mean to sound too grandiose, but I do feel it's just a net positive.

Leo Polovets 00:47:30

Well, so maybe on that front, like, I think that's a really interesting perspective. More from a company perspective than investor perspective, I feel like a lot of the most disruptive products were maybe non-consensus at the time.

Martin Casado 00:47:42

Totally.

Leo Polovets 00:47:43

Where you start with, you know, like no buttons on the iPhone or you got like Uber, instead of taxi, it's a stranger driving. And those are the ones where I think if you were like, “I'm gonna build a taxi company, but it's like 20% more efficient,” like probably could be a big business, but not quite the same level of disruption and growth as like, you know, you take a big bet and very high chance you're wrong, but if you're right, like you're gonna be, you know, in a really good position.

Martin Casado 00:48:07

Yeah. And this is so critical. I'm glad you brought it out. I really believe the best companies themselves are non-consensus to customers. I just think that the investing market is different than that. Like they kind of understand that. And therefore, a comment on investors being consensus is very different than a product being consensus. Does that make sense? Like. Investor sentiment, I think, is actually much smarter than people think. Like the adage is, VCs are dumb. Like they just, you know, chase trends, and all of that is true.

But the reality is, as a group, we have identified a cohort of companies that are quite disruptive and invested in them and priced them. And the companies themselves tend to be actually quite non-consensus to the actual consumer or to the market.

00:49:02 - How to think about incentive alignment

Erik Torenberg 00:49:02

I do wanna build, Martin, on your point because I think it's so interesting just to comment on how not everyone's incentives are totally aligned here. Especially between sort of the, what's good for the individual and what's good for the ecosystem. And so in the sense that, yeah, you know, if you're an individual VC, you don't want more capital. Or if you're a founder, you don't want more founders in your space. And so people were saying, competition is bad. You don't want competition. But competition is what fuels incredible product. It's like the Darwinian process. Like this is how, you know, we get, you know, bigger startup outcomes, a startup ecosystem, having more value, incredible products for customers and users.

Martin Casado 00:49:41

This is how we solve cancer, man. More money goes into VC, and we invest in companies. And then as opposed to investing in dying companies' ability to retain their place. 100%. Like all the finance needs to change.

Erik Torenberg 00:49:52

I think VCs are trying to straddle sort of, you know, LP incentives, founder incentives, their own incentives. And there is some overlap, and there's magic there, but it's also just worth acknowledging that not every individual person is aligned, and that's okay.

I also do still very much believe in the barbell, that there will be, you know, these big, massive funds that continue to win, and invest, and compound value. And also these smaller, focused, concentrated, experts, the boutiques who absolutely crush it. And we all work together.

Martin Casado 00:50:25

So Leo, we’re gonna run the numbers. I was trying to get it done by now, but there's a lot to do. The numbers are fuzzy. I just want to walk through what we’re going to be looking at, and then maybe we will schedule another podcast once the numbers are out to actually discuss it.

So one of the numbers we're going to look at is, if you cohort companies into winners and not winners, call it, looking to whether, on average, for that company, the rounds were priced above or below median for other companies at a similar stage, right? So this will say whether, you know, is it relatively high priced for winners or not?

And then the other one, which is even more difficult to determine is, given actual returns, are the bulk of the returns from companies that were, on average, high priced or not? And I think these two numbers will give us a sense to whether the market is actually pretty smart about the value and the price. You should not look for price arbitrage if you're looking for returns. Does that sound fair?

Leo Polovets 00:51:41

Yeah, I think that sounds fair. I definitely agree with the not looking for price arbitrage piece because I'll say, for me personally, my best investments have been ones on average that took a while to raise their seed round.

A lot of people didn't get or didn't like it. But on the flip side, some of the biggest misses are also the ones where it's like, “Oh, we liked everything except the price.” And like we thought it should be a 10, and some big fund gave them a term sheet at 20, and we passed, and then now it's a $10 billion company.

So maybe that was like, not a good pass.

Martin Casado 00:52:10

Yeah. You know, you know, Leo, honestly, as we go through this conversation, it does strike me that I think a lot of this is honestly, we have a bit different perspectives. Like I have to deploy a lot more money than you do, right? Like I’m a series A investor who needs to basically cap out 30 to 40 million in order to have a significant position.

And so I may have to be a bit more concerned about this than you do at the early stage. And I’m sure stage does color this conversation quite a bit. Because everything you're saying is totally sensible to me, so I don't think there's any disagreement.

Leo Polovets 00:52:42

It was definitely something I was thinking about, which is, I think if every check you write has to be at least 100 million, I think it's actually very hard to do non-consensus.

Martin Casado 00:52:50

Yeah. Right.

Leo Polovets 00:52:51

Because there's not a lot of companies that hit a stage where you'd invest 100 million, but it's still not clear if it's a good company or not.

And I think the earlier you go, if it's like $30 million checks, 10, 5, 1, I think you get more and more of a category where like you have the option and you could do the one assuming you have access to the consensus opportunities.

00:53:06 - Has multi-stage won seed investing?

Erik Torenberg 00:53:06

Leo, I'm curious what, and you know, you guys have absolutely crushed it at seed with, you know, Robinhood and Flexport, etc. But I'm curious what you think of Ramtin’s sort of thesis that multi-stage has won seed more or less in the last like 10 years. That when you look at a lot of the big winners, they were done from multi-stage firms, you know, at seed.

I'm, one if you agree with that sort of, you know, reading of history and then two, if you think that’s, well, definitionally you probably don't, think if it’s likely going forward.

Leo Polovets 00:53:37

Can I join you guys? I actually thought this was an interview. Sorry, what's the second part of the question?

Erik Torenberg 00:53:44

Did multi-stage win seed or more than seed firms win seed? Obviously there's, you know, First Round, Susa, you know, like lots of great seed firms. But when you look at the aggregate of winners. Do they have a multi-stage with seed, or not?

Ramtin’s argument is they had a multi-stage seed, and that's why he co-invests with multi-stage as his whole strategy. And then just, you know, the past isn’t, the future necessarily. What do we think about the future?

Leo Polovets 00:54:10

I haven't rigorously analyzed like the $10 billion, $50 billion outcomes. Over the course of Susa, I think we've invested in like 10 or 12 unicorns roughly, maybe like a third of those or quarter of those had a series A investor at seed. I’m not really counting, like sometimes it was like, “Oh, this series A investor did a 50k check in the YC round or something.” I mean like actually like took half the round or more.

So most of them still were seed only or were like, seed funds dominated the early round and then they went to multi-stage very quickly after that. But, so in my experience, I think there's a subset of seed where, I don’t know if I'd say multi-stage funds one, but they have like a very strong advantage.

Where if it is a founder that previously built a business that exited for 100 million, and they're like in the space that they know super well, that's gonna get done at like 40 instead of 20 or 80 instead of 20, post. And chances are it's gonna be a multi-stage and not like a boutique seed firm.

So I think for that segment, like multi-stage hasn't won, but I think it's probably like the predominant, like the majority of the time, they have a big leg up. I think for the other ones where it's less obvious, it tends to be much more seed fund dominated. Yeah.

Resources:

Find Leo on X: https://x.com/lpolovets

Find Martin on X: https://x.com/martin_casado

Stay Updated:

Find a16z on Twitter: https://x.com/a16z

Find a16z on LinkedIn: https://www.linkedin.com/company/a16z

Subscribe on your favorite podcast app: https://a16z.simplecast.com/

Follow our host: https://x.com/eriktorenberg

Please note that the content here is for informational purposes only; should NOT be taken as legal, business, tax, or investment advice or be used to evaluate any investment or security; and is not directed at any investors or potential investors in any a16z fund. a16z and its affiliates may maintain investments in the companies discussed. For more details, please see a16z.com/disclosures.

In this episode, Erik Torenberg is joined by Chris Best, cofounder and CEO of Substack, along with a16z general partners Katherine Boyle and Andrew Chen.

We trace the origin story of Substack and its cultural impact, including how it reinvented the business model for independent media. We also explore the evolution of blogging, the rebundling of media, and what the future holds as attention becomes the scarcest resource.

Timecodes:

00:00:51 - Substack’s impact

00:03:15 - Substack’s approach to free speech

00:11:19 - Substack’s vision

00:15:36 - How Substack builds product

00:19:39 - Algorithms aren’t all bad

00:22:01 - Would Substack launch an ad network?

00:29:10 - How Substack is changing media

00:33:01 - How will media consumption habits change?

00:38:09 - Disrupting mainstream media

00:45:15 - Why Substack raised $100 million

Resources:

Find Chris on Substack: https://cb.substack.com/

Find Chris on X: https://x.com/cjgbest

Find Andrew Chen on Substack: https://andrewchen.substack.com/

Find Andrew on X: https://x.com/andrewchen

Find Katherine on X: https://x.com/KTmBoyle

Find Katherine on Substack: https://boyle.substack.com/

Transcript:

00:00:51 - Substack’s impact

Erik Torenberg 00:00:51

Katherine, we've been talking for years about how much we love Substack even before we were formally affiliated with the company. Why don't you go first and talk about what you find so remarkable or striking about Substack's impact?

Katherine Boyle 00:01:03

Yeah. I think the impact is truly understated.

And I think we've moved so fast as a country and as an internet and as a world in the last few years that we've sort of memory-holed what it was like in 2020, 2021, but particularly for media, how crazy the 2020 moment was for anyone in the thought leadership space, anyone in the media space. So let me just go back to the summer of 2020.

James Bennet, who was the editor of the op-ed page at the New York Times, was forced to resign for publishing a sitting senator. An op-ed by a sitting senator who is still in office. The craziness that was around writing anything that was seen as heretical or asking questions or something that was seen as unorthodox in 2020.

There were mass firings, you know, Twitter deplatformed a sitting president, you know, Facebook as well, right? It was an extraordinary time and I would say a fearful time where very many people were afraid to say what they were thinking. You know, there was always rumors of people having unfettered conversations, like how dangerous it was that the people were having these conversations behind the backs of journalists. And there was one platform that stood up and said, "Hey, we are protecting free speech." And that was Substack.

And I think people forget that because it's just seen as, "Oh, of course we're in this new time. You know, Elon bought Twitter in November 2022. The Overton window has swung open. People can say what they thought," and I think people have forgotten that only a few years ago, we were in desperate times where people were losing their livelihoods.

No one was willing to say that freedom of speech was under attack. But the one platform, the infrastructure that was there to support those people, it was Chris, it was Substack, and they never wavered. And so I think that is the cultural impact. We would not be where we are today without Substack.

So I get very emotional. I'm like a superfan of Substack. I was on Substack in 2021. I'm very proud of that. But it's like, it's one of these things where I think we need to remember that we could have been living in a totally different time, in a totally different culture, had people not stood up and had the courage to say that freedom of speech matters.

00:03:15 - Substack’s approach to free speech

Erik Torenberg 00:03:15

And this was years before Elon had bought X and it was just kind of the first bastion of free speech. Chris, why don't you talk about sort of, when was the moment or what was the evolution for how you guys decided, "Hey, we're gonna take this stance, even if it's going to upset some of our most important writers, even if you're going to upset some employees, some investors, the ecosystem."

Talk about what that was like for you.

Chris Best 00:03:37

I've always seen the free speech thing as sort of an important pillar, but not the main pillar of what Substack is actually setting out to do. You know, the way that we think of Substack is making a new economic engine for culture and the idea, and it's not a partisan idea, it's not a political idea directly, it's just this idea that, you know, great things are made by independent voices who can do the work they believe in, make money, have editorial freedom, have a direct connection with their audience. Basically, you know, the backdrop of starting Substack was just this idea that, hey, you know, the internet came along and smashed a lot of the existing business models for culture. And what came in the wake of that was these massive internet-scale networks that are phenomenal businesses and that connected everybody like never before and had a lot of amazing positive attributes but, in my estimation, in our estimation, were kind of driving us crazy.

And the core of Substack is this idea of independence. This idea that the individual, left to do the thing they believe, say the thing they believe, make the thing they believe supported by an audience that's there for them is this like crucial ingredient in a healthy culture, in a free society. And freedom of the press, freedom of speech is one necessary precondition for that. I kind of think that in the long arc of history, that's not hopefully that controversial of an idea. I think it's a very American idea.

But at the time, just because the world was as it was, it was kind of out of vogue, shall we say, quite severely. I think in 2020, the thing that surprised me, the people that felt the brunt of that were not conservatives, were not Republicans, necessarily.

It was the people in the liberal media, and, in my telling, I would say, selectively the best and most interesting people that were getting just thrown from the ramparts. And the fact that this thing that we were creating, this new economic engine for culture that gives you this independence happened to be there at a time where a bunch of the most interesting writers in the world were getting summarily tossed from their longtime institutions.

That lined up really well for us from a business perspective. It was spicy from a cultural perspective, but that's the gig.

Andrew Chen 00:06:11

And maybe just to quickly add, I was going to say that it's amazing to see on the other side of the coin with just the blogging ecosystem, how much of that's changed. We've gone through LiveJournal and Xanga and Blogger, and we had Google Reader, RIP, and then you had basically a phase, Chris, when I met your co-founder Hamish initially and the company was three people.

This was also an era where the blogging ecosystem was dying, and you had the open, WordPress-powered blogging ecosystem, but there was no economic model. You ended up with a lot of spam, a lot of people hacking these poorly-maintained PHP websites.

I think this was also a really important moment to actually save blogging and writing on the internet, to actually create a model that, for a long time people just thought, "Oh, well, I'm just going to plug XYZ Amazon book and get affiliate fees, or I'm going to put Google AdSense all over my blog.

That was the only way to create this in any sort of economic thing. And for all of us that are in tech, it was cool to see that, you know, you had Ben Thompson from Stratechery really show that oh, there's maybe an alternate model. But it was almost always like a curiosity and something that was annoying to actually build. You know, you'd have to set up your blog. You'd have to set up your payments. You'd have to do all these other things. And so, I think also a really important moment for Substack to emerge from the internet media side to actually clean that all up and actually make it easy to actually put together something that became the big economic engine.

Chris Best 00:08:13

In the early days, people would often say to me in an accusatory tone, "Substack is just blogging with a business model." And I was like, "You know, that sounds pretty good. Like, if that's all it was, that would be pretty cool." And it turns out it's more. It's podcasting, it's a whole network. But I don't know, that seems good.

Erik Torenberg 00:08:28

It really reaches the dream, achieves the dream of reaching your audience in the sense of, if you have a hundred thousand Twitter followers, but you can't really engage them, and you're dependent on the platform, that's not as thrilling as owning your own email audience.

And I think what I love about what you guys did is you took the risk that, "Hey, we're going to give people their emails, and they can choose to leave if they want to, as opposed to being trapped to the platform. But we're just going to build such a compelling offering that writers are going to wanna stay."

And it's amazing, years later, to see a large majority, if not all, of the biggest writers stay on the platform.

Chris Best 00:09:10

There's only one thing that's better than people staying on the platform, which is when people leave the platform, take advantage of the export features, and then subsequently return to open arms and come back.

We call them boomerangs, and we love to see that too. I think the right to exit is really important. People thought that was very dumb. They said, "Well, if you just let your customers leave, won't they just leave?" And I think, in the short run, that might be true, but, in the long run, that created the right structure for us. It meant that we have to and still have to build a network that has enough value that even though you can leave, you don't want to. And even if you do leave, you might choose to come back. And I think that has caused us to keep the right thing at the forefront of our minds.

But I would say I think there's something even more important about the direct connection, which is, it's not just that I can leave, it's that, in my mind, what a subscription is is the option to give someone the right to reach out and tap you on the shoulder. It's to say, "You don't have to send me an email all the time if you don't want to, but if you want to send me an email, if you want to send me a push notification, if you want to show up at the top of my inbox, I kind of give you that right." And something that that lets you do as a writer or as a creator is to take creative risk.

Something that I hear a lot from YouTubers, people who are very good at YouTube, people who have massive followings, who are very successful, is, "I have this idea for a thing that I could make, and I know that it would be great, and I know there's an audience out there who would like it, but I can't make it because if I made it the way that I want to make it, no one's going to see it because it doesn't please the algorithm." And so the direct connection, in addition to being this way you can bring your audience with you, is a way to give humans the power to override the algorithm and say, "Hey, I've got this trust relationship with my audience. I want to exercise it and go out on a limb and say, 'Hey, I want to call in that favor and have you pay attention to this thing that I'm saying is good.'" And sometimes it might be bad and you might unsubscribe, but sometimes it might be great, and it might be something great that could not have existed if the only way to reach everyone was to please the algorithm every single time.

00:11:19 - Substack’s vision

Erik Torenberg 00:11:19

So in the beginning it was a blogging platform with a business model, as we just said. And the vision has gotten bigger into more of a network, more of a platform across formats. Expand on what is the big vision for Substack? And I'm also curious how that's evolved, if the vision in 2018, 2019, 2020 is the same vision as it is now.

So tell us the vision, then we can trace the evolution of it.

Chris Best 00:11:44

I would say that we started from the very beginning with I think a very ambitious, some might say derangedly ambitious vision. Again, the backdrop was kind of, we think that the internet has massively reshaped the economic incentives for media.

The origin of the company, I'll just briefly tell this because it's germane here, was I was taking some time off in my last startup, and I had always wanted to be a writer. I'd always been an avid reader. I've thought that what you read matters. What you read, the media you consume is not just a way you spend your life. It changes who you are. It changes who you are as an individual, it changes how you see the world, and it changes cultures and societies. And so great writing, great media, great culture in general is this inherently valuable thing. And my first instinct was, "I should make some of that. Like I could write an essay. How hard would that be? I know how to program. I know how to type." And I started writing what was supposed to be this essay or this blog post detailing my frustrations with the media economy on the internet. So this is where it started. I'm like, "Waaa, waaa, waaa, look at all these great things the internet has done, but it's also kind of created these, you know, mimetic evolutionary landscapes that are driving us nuts. You know, this is going nowhere good. Look at how the culture is shifting. Waaa, waaa, waaa." And I sent this essay to my friend Hamish, who's actually a writer, and he let me down very gently.

He said, you know, "It's 2017, and your essay is about, you know, maybe the newspaper businesses are in trouble, and maybe Facebook is not an unalloyed good. Dude, we know, everybody knows that. Or everybody who's in my industry knows that. But the better question is, 'Let's say that all of those things you're complaining about are true. What could you do about it? How could that be different?'" And we started arguing about that. And so we had this sort of, I think maybe this is an a16z-relevant thing, this sort of techno-optimist idea that it's like, look, you're not going to turn back the clock. If there's new powerful technologies that are changing how everything works, and those things come with trade-offs, and there's upsides and downsides, and there's contingencies, there's historical contingencies where the world could tip in one of many ways. The right way to address that is not to lament it or to wish, "Hey, we should go back." It's, "Hey, we should put these things to use in service of people. We should imagine what the best version of this future is as these new networks take off, as these new technologies take off. And we should work proactively to help usher in the better freer more exciting version of that future." You know, heady stuff. So we had this big idea, this big sort of grandiose thing, and then we just had the kernel of the way to start.

And the way to start was we could make it dead simple to start a paid email newsletter. And that was a thing that there was probably 20 people in the world that really, really wanted it, but they really wanted it like it was going to be the best thing ever for them. And it was the kernel; it was like the smallest possible instantiation of that much bigger idea where you were going to create this new economic engine that lets any independent voice make the things they believe in, make real money doing it. It's a way around the cold start problem because you could have an individual person, like the very first Substack newsletter made total sense.

So we started with the grandiose version of Substack firmly fixed in our minds. Even then, I think, we looked at YouTube as something that was maybe the closest version to this thing that already existed.

00:15:36 - How Substack builds product

Erik Torenberg 00:15:36

Talk more about how you decided to launch Notes or go from, "Okay, we've got this sort of business engine where we've got all these writers making a lot of money. Where do we go from here?"

Chris Best 00:15:47

I'll tell one step before that because it went into my thinking, but very early on, in the very early days of Substack, we were like, "Okay, the thing that's going to be really different about Substack is it's all going to be paid because that's the thing that aligns the incentives. That's the thing that makes this thing different. And so in order to be very pure to our vision, we are not gonna allow anybody to have a free Substack or to send emails to free people." And that evaporated with our first customer because he was like, "Oh, okay, then I'll just use MailChimp for the free version and then I'll funnel the people here."

And he created this stitched-together thing and it was like, "Oh, this is really dumb because if you want to be successful, if you wanna make a successful paid Substack, you have to have a free Substack. And in order to make that experience good and have the conversions actually work, we should just support that. And it's not an abrogation of our vision. In order for the thing to work, you have to provide this other thing." And then the thing that we realized not too long after that was the same was actually true about Twitter and about the social networks, which was, you know, in 2018, 2019, if you wanted to have a successful Substack, you had to also have a successful Twitter, or a successful Facebook, or a successful LinkedIn. You had to have some top of funnel place, you know, the same way that the legacy media was struggling, and you had to have Facebook traffic, or you had to have Google traffic, or you had to have something.

There was always these other networks that were the source of your business. So even if you were this independent writer, independent creator, you were downstream of these other platforms, and that had both a philosophical consequence, which is we're trying to make this place that has these different incentives, but you're still at the whim of the thunderdome, right?

You still have to play the Twitter game, or you still have to play the Facebook game. And it had this very practical problem of those networks don't give a shit about you as a creator that makes money. And, you know, Mark Zuckerberg can decide in a fit of pique that people are annoying him about politics.

So he is going to like turn off politics. And if you are a politics creator that depends on Facebook for your livelihood, that's an existential event. And it's not even, because it's like they're trying to do that. It's just like, hey, these networks twist and turn, and they don't really have any intrinsic interest in helping you build your audience and make the thing you believe.

And so we had this idea that in the long run, the only way we were going to really make that work for people is to build one of these networks ourselves that was built on different laws of physics. And so we were going to build, you know, a network, a destination, a place that you could go and experience the internet and have all of those great things that you get from social networks, but with a different business model and with a different incentive structure.

It's not going to like replace them, but it'll live alongside them, and it'll be the one place on the internet where it actually does want you to succeed. It actually does want you to go and find something interesting and long-form to read or long-form to watch. It does want you to find and fall in love with something enough that you might choose to pay for it.

And that's going to create a very different feel from everywhere else that just wants to keep you glued to the screen. So we had sort of this theoretical idea of why we had to do this thing, but we also knew that it was going to be quite difficult. Like it's very hard to start a new internet-scale network. And it took years.

00:19:39 - Algorithms aren’t all bad

Andrew Chen 00:19:39

By the way, Chris, to your earlier point on this, on the algorithm, it's so interesting to watch actually all of the major platforms move towards the algorithmic For You world because in that world then actually the creator's relationship with their audience is even further away, right?

Maybe it doesn't matter. And this all originally started with, "Oh, well, you know, we have this problem of any social app where you need people to follow enough folks so that they get enough feed content, and, well, one way to solve that is even if you're not following somebody, maybe we'll just kind of suggest things." And it turns out then the algorithms are so good that maybe that should be their entire feed is just suggested content.

But then what does it mean as a creator to even build a following on one of these platforms if, even if you have a hundred thousand followers or whatever, maybe they'll see none of your content because the algorithm is caring less and less about the follower graph these days.

Chris Best 00:20:40

Definitely, and there's two tacks you could take with that. And the one that I think a lot of people, their first reaction is to say, "Oh, well, algorithms are bad, right? Like the algorithm is," whatever, "It's severing our ties. It's putting us into bubbles. It's exposing us too much to people outside our bubbles." But like, whatever the thing is.

Okay. So trade-offs with algorithms. Therefore algorithms are bad. I think a more productive take is: algorithms are powerful, and they're a tool that people use, and they serve the ends that we tell them. And if we tell them better ends, they'll help us get better results. And so this is something that we talked about a lot at Substack because I think people had this—there was a lot of our users who felt like at the time, they were like, "The good thing about Substack is there isn't an algorithm, and I just connect directly. And that's the thing that's actually good." And I think the take that we have is, there's something that's much better than that, which is what if there was an algorithm that actually served you and that was actually trying to help you find the things that you deeply valued and actually had a, you know, like the nerd term for this is an objective function.

If the objective function, in other words, the secret hidden master that the algorithm is serving is actually your own interest rather than, you know, trying to sell you more ads.

00:22:01 - Would Substack launch an ad network?

Erik Torenberg 00:22:01

You have a very sophisticated writer base and then, by extension, a very sophisticated reader base, very high-value audiences.

And now especially with video, and people aren't used to paying for video in the same way they're used to paying for writing, partially because of Substack's innovation there. Will you also launch an ad network at some point, or do you think that risks sort of the golden goose in some way?

Or how, how do you think about that?

Chris Best 00:22:25

I kind of take the same thing we talked about with an algorithm, the same thing about building a network. I'm going to say the same thing when we talk about AI, which I assume we will do. But I see, you know, sponsorships, advertising is a powerful force. The thing that would break Substack is if we just looked at the same way that the legacy social media things built advertising and said, "Oh, we're just going to copy that. Like, that's gonna work." Because if we did that, the thing we would be doing is importing the incentive structure and the business model that puts the platform at odds with the people on the platform.

On the other hand, there are a ton of Substackers today. Some of them are like, in my opinion, some of the very best Substackers, who are selling sponsorships. And I think there's a version of unlocking more economic opportunity, more economic upside that is aligned with the idea of independence, the idea of having differentiated value and quality.

And so we're very interested in doing that. But my belief is we have to take sort of a first-principles approach and not just, you know, stuff ads in a thing but ask the question like, "What would the good version of this be?" And help build that.

Erik Torenberg 00:23:49

Yeah. I think the bear case for ads has been sort of, "dumb-it-down content" or sort of "clickbait for the masses."

The bull case has been sort of allows niche writers to monetize without charging their audience a ton. Or it doesn't succumb to audience capture in the same way that a subscription could. Basically there are pros and cons with both business models, and you guys have to figure out how to integrate it in a way that works for the reader and the writer.

Chris Best 00:24:22

And I think the same is true of all of this magical AI technology that's coming online. I mean we're building a live product that basically feels like doing a FaceTime call and then magically turns into a highly produced podcast and a YouTube video and a series of short-form clips and a transcript, and pretty soon it's going to be in whatever language you want.

And we're going to live in a world where, you know, one thing you could have is you could have a bunch of like AI slop that kind of keeps dumb people clicking. The other thing you could have is you could have a future where there's way more creative leverage and where the people who are making this independent stuff, who have the independent voice can do way more, can make something much better, can realize their vision much more fully.

And so in all these things, I think you look at the technology, not as good or bad, you look at it as a powerful means to an end, and if you pick the right ends, then applying the technology is very exciting.

Katherine Boyle 00:25:21

This is something I think you were so early to understand, that is sort of common knowledge now or becoming more common knowledge, but wasn't five years ago, which is that everyone can be a creator, and we don't have enough content.

I think there's this horrible meme, like podcasts are over, we have too much content, there's too much online, and it's like, actually it's the opposite. If you look at any of the For You feeds, I mean, most of it is now AI slop, which says that there's just a dearth of extraordinary content.

And what I always thought was so brilliant about what you understood about professional writers, and having been a professional writer, it was almost like you were inside my psyche, the hardest part about writing is writing. Like it's really, really hard to get started writing if you're like a true writer and you have writer's block. And so everything you can do to make the production of that writing easier, everything you can do to sort of create the flywheel where your readers are expecting something, you're artificially creating deadlines, if you can create something very quickly that turns into a host of different products, that then gives you the positive feedback loop that you need to keep doing it.

Like there was something about, from the very beginning you really understood sort of the artist's way or the writer's drama of just how difficult it is to be a creator. And that exists within everyone, right? None of us are, our day jobs are not writing, right?

But all of us are writers, all of us are creators on this pod. And so there's something about, if you can make people's lives much easier and make the creation loop easier, people who have day jobs will then do it and create magical, great content to rival the kind of terrible content that now is being produced by these like meme farms.

You had a very early insight, and you're seeing, sort of, AI pushes this direction where it's going to be this hybrid of really creative people using AI to make beautiful products that otherwise it would be like the barrier for entry is way too high to do that.

Chris Best 00:27:15

I started a whole company to procrastinate from finishing an essay. So I definitely know that end of it. The thing you're describing and the way I would've put it at the time and I would still put it is: we've entered a world where attention is the scarce resource.

And that's actually not new with AI. I date this to kind of the social media, the internet revolution where it used to be, like when I was a kid, you could get bored. You could be sitting around, and you'd be like, "Dang, I wish I had something to pay attention to right now, and if you could give me something free to distract me, that would be a really good deal."

And that was the situation where the original social media network giants rose up was it's like, there's this land grab for attention. Everybody has so much attention to give and not enough things to distract them, and we have won that war. We have won the war on boredom. Right? Nobody has the problem of, I have five minutes, and I don't have anything to do to kill that time, but the amount of attention I have is not infinite. And so now I live in a world where there's no scarcity of content, but there's a huge scarcity of good content. There's a huge scarcity of things that are worth paying attention to, and this was the fundamental insight of Substack is, you know, as somebody who has one life to live, if I could spend a little bit of money to get better culture, better ideas, more interesting use of my time, things that help me become more the person I want or aspire to be, that's actually a phenomenal deal, and it would be insane of me not to be willing to spend money or spend a bit of effort to find that better thing. And people are starting, the culture is starting to catch up now, I think, to this reality that's been true for a decade. That, you know, you're spending your life, when you choose what media to consume.

00:29:10 - How Substack is changing media

Erik Torenberg 00:29:10

I think another huge contribution that you guys have made is around price discovery, where it turns out that the true value for, let's say, Noah Smith isn't 80k writing at Bloomberg. It's a million dollars or whatever it is that he makes now writing on his own. If only it had existed when Katherine was a reporter at the Washington Post, maybe she wouldn't have had to—

Katherine Boyle 00:29:33

Wouldn't have had to suffer through this venture career.

Chris Best 00:29:34

There's two people on this, maybe three, all of you are actually people that we've tried to recruit to be Substackers that wound up at a16z instead.

Erik Torenberg 00:29:43

And so it is just fascinating to see you guys align, kind of, value capture and value creation in a way that wasn't aligned beforehand. And we're starting to see not just people go independent, but also sort of the rebundling happen where people like Bari, where Katherine is on the board of The Free Press, um, build Substack-first, you know, media companies and other people as well. Talk a little bit about sort of the unbundling and rebundling and kind of the future of how you see media companies being built.

Chris Best 00:30:16

This actually reminds me of one of the first things Marc Andreessen ever said to me when we were talking about Substack.

He said, "You're gonna do to media what the venture capital industry did to software companies or to tech," which was, there used to be this time where if you were somebody who knew how to build great software, the way that you could do that would be to go get a job from somebody in a suit that would tell you what to do and pay you a salary.

And the hidden reality of that situation was the people who actually could make the things were creating so much value that they were massively getting underpaid and under-recognized compared to what they were doing. And like less obviously but even more interestingly, once you could free them up from that structure, and you actually put them in charge, put the people who are actually making the thing, make them the boss, it massively increased variance in this very positive way. Didn't always work, right? Not every software programmer is going to be a great founder. But the best founders who actually build the thing are so much better, and the results are so much more interesting and extreme and wonderful than the world where they just got bossed around by whoever was the software company middle management, that the net effect of kind of like pulling the talent out and unleashing it and putting the lunatics in charge of the asylum in tech was this renaissance, basically. And I think the same thing is possible in the cultural industries. I think that the people who are actually making the stuff are the heroes. They're putting themselves on the line.

If we're going to have a renaissance and a new flourishing of culture, those are the people that are going to make it and the people that are investing in them and, you know, investing their time and their money and participating. And the ambition that I have and we have at Substack is to basically just like build what they need, build the tools they need, build the network they need to have a fighting chance to win. And I think we're on the way.

Erik Torenberg 00:32:35

Yeah, it's interesting. And even in VC there are sort of, you know, solo capitalists as sort of like the Noah Smith example, but then there are also people who, you know, go on and build, you know, bigger platforms, much bigger than their individual selves.

Chris Best 00:32:48

I think of them as ambitious media founders, right? We have a whole team at Substack who's dedicated to the principle that if you're an ambitious media founder, we want Substack to be the best possible place to realize the biggest version of your ambition.

00:33:01 - How will media consumption habits change?

Erik Torenberg 00:33:01

Let's talk a little bit about the future of media in a sense of, you know, there's only 24 hours in a day. There's a portion of that people spend, you know, engaging in content, and it all competes with each other. Looking out a few years, do you see the amount that people spend on just that overall content in general increasing? I guess I'm curious, if video is obviously going to increase, are people going to be reading less, or is just more of everything, or how do you view consumption habits changing over time?

Chris Best 00:33:35

I wrote this piece called The Two Futures of Media where I kind of argue—I think inevitably when you ask these questions, you get into sort of like weird philosophical questions. Like what is the purpose of media, and what are we doing here? And I think that one of the purposes of media is to entertain, to have some effect.

The extreme way to say this is people who use media like a drug, right? I'm going to sit there, I'm going to scroll this thing, I'm going to watch this thing, it's going to have some effect on me in the moment that's going to create a pleasant feeling, and that's one of the things that I want from it.

And I think that that side of media is going to get supercharged. We have very sophisticated AI goonbots now. Is that a good thing? I don't know, but it's happening, and we're going to have that across everything, every short-form video, everything that could be like this, you know, it's almost approaching wireheading, the science fiction idea of, like, you plug a wire into your brain, and it stimulates the pleasure centers.

I think that future is, we're well into it. It's only accelerating. The stronger the technology gets, the stronger that thing becomes, and the stronger it's a hazard for people, quite frankly, because there's a mode of consuming media and culture that is like drug addiction, where it is compelling in the moment, where it is something you want, it is something even you'd be willing to pay for, at least spend your time on, but it pulls against your long-term interest. Remember, the media you consume is not just how you spend your time, it's who you become. And so it degrades you. And so it makes your tastes get more base, it makes you want more of the dumb thing. It sort of pulls you in. That's already happening.

It's going to continue to happen. That's a big part of the future. I think that thing is baked in right now. But that's not the only purpose of media. Right, the other purpose of media is culture. The other purpose of media is to like live in a society and become the kind of person you want to become and to figure out how to live and act back on the world, like the intersubjective, multiplayer game of building with other people.

And that is something that people really, really want as well. And I think that the same technologies that are making the first thing much more compelling can make that second thing much more compelling as well. And the thing that I think we can do at Substack is to create a version of that thing that is also fun and is also good and is also empowering. And you don't have to kind of be like—I don't want you to have to be like a monk to use Substack. You're like, "Well, I could scroll TikTok, or I could go to the library and flip through some microfiche." And it's like, yeah, you could do that, but nobody's actually going to do that.

And so if we can take kind of like the good and interesting and culture-laden future of media and make it really good and make it really compelling and have people make money from it when they make something truly great and have people realize that, you know—I aspire that the Substack app could be a place where you look back at the time you spend on it and think, "Damn, I'm glad I did that. That made me a better person. That made me more interesting." And I think that that is possible and that if we—When there are these massive changes, when the world changes, when technology reshapes everything, I think the fact that there's going to be change can become inevitable. But which version of the change happens, which future you go to is contingent. Right, people often ask like, "Is the future determined," or "Is great man theory true," you know, "How does history happen?"

And I think it's just both, right? There are these inexorable changes that are going to happen no matter what, but then in the moments of change, which future emerges is contingent on the choices people make and the accidents of history and individual decisions. And so I think the thing that is possible for us to do is to build a version of that second future of media where people are reading things that make them smarter, or listening to conversations that plug them into the world, in general, acting back on the culture and participating and engaging in ways that they value, and that that creates a ton of economic value and creates—This is why it's an economic engine for culture—creates like a whole world that is intensely valuable and great.

Is that going to be the world that everybody goes to? No. Some people are going to sit on the AI goonbot, but I think we can make a real difference by making that second future as good as possible.

00:38:09 - Disrupting mainstream media

Erik Torenberg 00:38:09

Building upon your culture point, I've started to see some academics, also on Substack.

We've been talking a lot about disrupting media. I'm also curious if you think much about sort of academia or books or kind of these adjacent industries? Or is that a distraction or you don't think about it super deliberately?

Chris Best 00:38:30

I'm a total crank on the subject of academia, so it might be fun.

This is sort of like ill-considered on my part, but I think a lot of science is totally broken. And I think that the scientific project is incredibly important and one of the most valuable things in history but that the practice of science and the current situation in academia and especially in academic publishing is like pretty far from good and even to the point of like, I think maybe peer review is a huge mistake and doesn't actually work, and, you know, we've got this thing that's supposed to make everything good and there's like this massive crisis of huge bodies of fake science that nobody believes because it's all LARPing. And I'm very interested in the idea of like, what if you apply some of these same principles? What if you give people an alternative? One way you could do science, if you want to, is to go on the internet and publish it. I think that's actually pretty radical and pretty interesting, and I see some early shoots of people doing that. It's a topic that I am excited about and think that there's more that we could do, but hasn't been kind of like central to our efforts so far.

Andrew Chen 00:39:51

I was just going to say something about books, right, because I think the process of why people decide to write books today is in itself a really interesting decision. Because like, first, you have to spend, you know, it's a multi-year project to actually, you know, write a book.

And I worked with HarperCollins to do the Cold Start Problem. I think it's like been three or four years ago, but it takes like three years or something like that to actually write a book. And then many of you guys know that if you literally just get enough pre-orders that you can get 10,000 units sold, that's like a bestseller.

I mean, it's like people are not reading books right now, which is insane. There's literally, I think, like one book printer left in the US. And so if Michelle Obama decides to write a book around the same time as you, they're like, "Oh, the printer is booked for the next, you know, X months." It's all down to one set of printers, which is itself insane. And so Chris, when you compare that to the amount of work involved in writing a book versus being able to click the publish button and have that go to, you know, a hundred thousand people's inboxes each day, it's a completely different thing. Now, it is fascinating that there's certainly a cultural prestige in the fact that books have been around forever, but I have to imagine that it just changes over time. I imagine that, you know, it's like when people were playwrights, and film gets created, they're like, "Oh, wow, people love film, but it's not as prestigious as, you know, plays."

And then TV is the same. It's like, "Oh, people are watching a lot of TV. It's not as prestigious as being in a film." And then we're going to go down the same thing with like, YouTube stars and streamers or whatever. I think a lot of this stuff is obviously very much lagging.

And the ability to just reach, you know, hundreds of thousands of people with something that you write over a cup of coffee is like itself just so powerful, when you really think about it from an ROI basis of like writing a book or a Substack. And of course they're not mutually exclusive. If I were to have redone my whole thing again, I probably would've like, written it all on Substack and then taken it and put it into a book, as opposed to thinking about it like, "Oh, I'm going to lock myself into hotel rooms during my vacations and try to crank out all these pages and then kind of do it all as one big thing."

Katherine Boyle 00:42:19

I agree with that. I think there's this like cyclical moral panic that happens, and it certainly happened in media, where it's like "People are writing on the internet without an editor. Oh my God. No one's editing the writing on the internet. What are we gonna do," right? Like that was like the media's version of this.

It's the same thing happening with books. "People are reading, but it's not in a book. Like they're reading things on the internet. The book process is—" It's incredible to me, and it happens all the time. It's always like legacy industries realizing that the internet actually is a thing, that it becomes easier to produce the same thing you were going to produce in a book format, or the same thing you were going to write for a print paper can be put on the internet, and it's the same content.

And so I think there's always these like moral panics that we're somehow getting dumber or people aren't reading enough, and that's a huge problem. And I just don't think people are looking at it holistically, that people are reading, they're reading in different ways. Yes, you could say something, there is a huge problem if young people, you know, grow up never having read an actual physical book that was written before the current times. That's a different discussion. But the moral panics about the actual medium I think are something that are very cyclical, have been happening since the birth of the internet.

And it hasn't necessarily affected—it's affected the freedom of what we can actually say and the freedom of what we can get our hands on, but it hasn't necessarily affected our ability to read. And certainly, I would argue it actually hasn't affected our ability to make arguments either, which I know would be controversial in some domains.

But I think it's more the moral panic of industries realizing that things are changing, and they have to adapt.

Erik Torenberg 00:43:54

Yeah. And there's an interesting question about what is the source of intellectual culture these days? Is it more streamers, is it more Twitter anons?

Is it more professors, more journalists? I think, you know, Alex Danco wrote this really interesting case for why it's long-form writing. And one of the reasons he said was, it's not that everyone reads the long form writing, it's that an important group of people reads it and then translates it or transmits it in kind of a different format, and then the masses read, or sort of engage with that content.

And I think just having a more sophisticated understanding of the supply chain gives us a greater appreciation for long-form writing as a source.

Andrew Chen 00:44:37

Yeah. And Erik, to your point there, it's like, what that means is everything that you read in printed-out pieces of paper in traditional press is like delayed by a huge amount compared to the actual discourse that's happening on the internet.

And long form of course is like you're actually able to generate really original thoughts. And then of course all the meme wars are where the real discussion happens in real time. So you kind of have both flows like generating cultural knowledge over time.

00:45:15 - Why Substack raised $100 million

Erik Torenberg 00:45:15

So Chris, we're here partially to celebrate your big round. 100 million, is that right?

Chris Best 00:45:21

100 million. Yeah.

Erik Torenberg 00:45:22

So talk about, why raise a hundred million? You're already crushing it as a business. There's already a lot of cash. What are the big plans here?

Chris Best 00:45:30

So I think the big story of this to me is we've had this long-term ambition for, what are the pieces of Substack? I literally put this meme in my pitch deck, which was the handshake meme. And one hand is a model that supports independence. And the other hand is an internet-scale network.

And to me this is the core of what Substack is always meant to be is hey, this model that supports independence, but also this, like this place, this part of the internet that's a first-class destination that has this like thriving scene that feeds it. And I think after years and years and years of working to kind of make that into a reality, we have that fledgling network alive now, and it's growing. And I see the next phase of Substack as feeding that machine and helping it grow and throw off all of this value and like economic value for the creators and cultural value for the world. And it's kind of going to mean rebuilding the company to match that scale and ambition.

And this fundraise was really just a way to unlock that kind of transformation. And so we are sort of like in an exciting period of reimagining, you know, the product, the company, and what this thing can become now.

Erik Torenberg 00:46:49

Well, that's a great place to wrap. Chris, thanks so much for coming on the podcast.

Chris Best 00:46:52

Thanks.

Stay Updated:

If you enjoy the show, please follow and leave us a rating and review on Apple Podcasts or Spotify.

Find a16z on Twitter: https://x.com/a16z

Find a16z on LinkedIn: https://www.linkedin.com/company/a16z

Subscribe on your favorite podcast app: https://a16z.simplecast.com/

Follow our host: https://x.com/eriktorenberg

Please note that the content here is for informational purposes only; should NOT be taken as legal, business, tax, or investment advice or be used to evaluate any investment or security; and is not directed at any investors or potential investors in any a16z fund. a16z and its affiliates may maintain investments in the companies discussed. For more details, please see a16z.com/disclosures.