America | Tech | Opinion | Culture | Charts

Head In The Neoclouds

Earlier in the 20th C., the Southern Pacific Railroad had a surplus of building rights on cleared land that connected various cities and towns across the country. A railroad’s right of way is much wider than the physical track, leaving plenty of buildable corridors to spare. So, the railroad built a communications network alongside the track, dubbed Southern Pacific Railroad Internal Networking Telephony, which it began commercializing for broader use in the 1970s. But then two things happened at once: the long-distance calling cartel came to an end, and fiber optic cables became commercially viable. The communications corridors were repurposed for fiber, and the network eventually became known by the acronym Sprint, making former railroad assets a backbone of the telecom revolution.

Railroads were not the only ones who repurposed their physical networks for a much bigger commercial technology. In the 80s, the Williams Company converted its empty natural gas pipelines into conduits for fiber and created WilTel, which was sold and eventually renamed WorldCom. In the 90s, one-way coaxial cable built for the cable TV business, underwent a massive and expensive upgrade to become the broadband consumer internet infrastructure for Comcast/Charter.

All of which brings us to another set of businesses sitting on preexisting infrastructure that has been dramatically repurposed and repriced to meet the demands of an emerging technology: neoclouds. Just to generalize a bit, neoclouds were in the energy- and compute-intensive business of crypto-mining, but then, of course, AI came along. Suddenly, having the power-rights, infrastructure and know-how to build and manage intense computational demand (and, in CoreWeave’s case, a lot of GPUs), has become one of the hottest games in town:

It’s not an apples-to-apples comparison, of course, but for the three largest publicly traded neoclouds, revenue growth has been a thing to behold.

We can only estimate the early-innings of hyperscaler cloud revenues, but you get the idea: neoclouds are growing very quickly, and at a much faster pace than the big three of cloud services maintained when they got started.

To be clear, neoclouds are still a fairly small player in the broader universe of selling compute:

They also have a long way to go before they get anywhere near as large as the hyperscalers:

The hyperscalers generate orders of magnitude more revenue per quarter than the neoclouds, but at the same time, it’s taken CoreWeave ~25 quarters to achieve the same $2.6B as AWS generated by quarter 40 since launch. Again, these are very fast growing businesses.

With all that growth, and all the AI tailwinds, you’d think that investors would be thrilled, and to some extent, they are, but it’s a more nuanced story:

While everyone had a pretty good earnings week, over the past year CoreWeave is still down ~16% (and only Nebius has come close to its previous highs). So, it’s a good story, for the most part, but certainly less-so for the largest neocloud of the bunch.

Part of the reason for the somewhat muted performance of late is that a lot of this growth is arguably priced-in:

Sales multiples aren’t the best choice for capital-intensive businesses like neoclouds, but they’re useful for illustration purposes. The smaller, but faster growing Nebius and Applied Digital trade at a much higher premium than the much larger CoreWeave, which is still doubling, but it’s not quite the 400%-450% growth at the top of the leaderboard.

The real issue, such that there is one, is not growth though, it’s longer-term profitability. Neoclouds need chips, power and physical infrastructure to grow, and that’s not cheap:

Just taking CoreWeave as an example, revenue growth is substantial, yes, but Capex is even more substantial. Other massive costs include depreciation for the chips, which is more than half of revenues, and the rising interest expense on all the debt that’s been used to advance the expensive buildout.

Charts is not here to express a view one way or another on the success of neoclouds or their current stock prices. Other than being topical, the point here is that neoclouds represent a nice little microcosm of the push-and-pull of the broader AI trade: on the one hand, historically fast-growing businesses in a vertical that has grown much larger than anyone thought possible, and is growing even bigger still (i.e. compute); while, on the other hand, historically expensive businesses to build with massive (and depreciating) fixed infrastructure costs.

Return of Horizontal SaaS?

Just a quick check-in at the shifting contours of the saaspocalypse.

One of the biggest losers of the great software selloff had a pretty good month:

Horizontal software businesses did some of the best work in the IGV software ETF over the past 30 trading days (although they’ve given back some of those gains from the date this data was collected).

In general, fundamental performance has held up, and Atlassian especially has defied expectations for an AI-driven demise. The productivity software pulled off a “double-beat and raise,” with its cloud business growing 31% YoY (and the revenue backlog growing even more). But the key was perhaps some evidence of AI as an accelerant rather than a headwind. The company flagged widespread adoption of Rovo, the Atlassian AI assistant, and noted that Rovo-users were growing their spend at nearly double the rate of non-users.

Good for Atlassian, good for Rovo, good for horizontal saas. It remains the case, however, that horizontal saas is valued slightly less than other software:

With some exceptions, forward revenue multiples for horizontal saas companies (including Atlassian) are generally below the trend correlation of growth to multiple.

Again, it was a relatively good month, but it will take a lot more than a month to convince the market that the saaspocalypse has been cancelled.

Unless, of course, your software is in the business of cyber and observability, in which case, there is no saaspocalypse and never was one:

The cyber category continues to run away from the IGV field, in this case, with AI as a tailwind—the presumption is that AI is heightening the perceived risk of cyber threats, and no buyer is going to vibe-code their own solution.

Time will tell of course if the thesis bears out, but for now, when it comes to incumbent software, the trials and tribulations are the opposite of indiscriminate: investors are highly attuned to whether AI is likely to giveth or taketh away, and they’re updating their priors with every fresh batch of data (as they should).

Towards an Efficient Frontier of Token-Maxxing

The landscape for model usage, token consumption and managing token spend, continues to evolve in all kinds of interesting ways.

Take Databricks, for example. Rather than try a winner-takes-all approach to “what model should we use?” or “what model is best?” or simply give engineers a budget, and say “make of this what you will,” Databricks said “what if we engineer a solution that routes the right tasks to the right models?”

They’re not the only one of course, but Databricks built a “Smart Router,” and they like what they see:

Apparently, the Databricks router was able to “consistently reduce average task cost by more than 30%” by using more powerful and more expensive models, when necessary, and using less powerful and less expensive models, when possible.

In general, it’s hard not to see the pursuit of an efficient frontier of token spending as a very good thing. It means that demand is continuing to grow, and that use-cases are evolving at not only the peak end of performance, but also for less-peak models (which, in the first iteration of the doomer-tale, were supposed to rapidly become obsolete).

As we’ve noted before, that efficiency increases the surface area of demand is precisely the Jevons-like dynamic that you want to see:

Silicon Data’s token price intensity index shows that overall price intensity has been declining, particularly as cheaper open models increase their share of a growing market. To reiterate a point about this commonly misunderstood data, the indexes are measuring the cost-intensity of token spend, not the absolute dollar value—it’s a combination of the number of tokens consumed and their blended cost, which means that the total dollar value (and number of tokens consumed) can still rise, while the $/token falls.

The key though is that overall demand continues to rise, and an efficient frontier of pricing and model-selection should only help, especially because “AI demand” (and “AI Adoption”) is not one thing. There are still huge gaps between power-users and everyone else, which is a pretty good indication that “use the best model all the time,” is a good approach perhaps for some companies, but certainly not all of them—that there are rapidly emerging alternative approaches to choose from, is net good overall.

According to Ramp’s data, everyone is spending more on AI, but the differences in spend between the median and the top decile (and between the top decile and centile) are enormous:

Ramp’s data tends to skew towards tech companies, so keep that in mind, but the top 10% of firms are spending ~50x more per employee than the median.

That distribution isn’t likely to be an accident. The firms that have figured out how to unlock more value from their spend, presumably are the ones spending the most (if not in all cases, then in a lot of them).

An analysis from BCG of 107 public companies found that the top two quintiles of token usage had considerably faster revenue growth than the field:

The point here is that token demand and efficiency are mutually reinforcing—the more value firms get, the more they consume. There’s of course some give and take, and R&D always includes some upfront costs, but unabashed “tokenmaxxing” was never going to be a useful approach for the vast majority of firms, so it seems like a very good thing that they increasingly won’t have to.

Lab Talent Wars

New Media just saw two excellent team members move to OAI, so we’ll conclude with just some fun charts on frontier lab hiring.

Dario Amodei recently commented that he was concerned that employees were starting to prioritize money over mission, and according to data from Levels.fyi, he may be on to something:

Taking the data at face value, Anthropic pays its engineers a lot, and it pays a lot more than comparably senior engineers at companies like Google and Tesla.

It’s good to be a member of technical staff, I suppose.

Elsewhere, this was also fun. According to data from Live Data Technologies, surfaced by Truist Securities, there’s a lot of overlap between where the labs hire from, but also not:

• Both companies do a fair amount of hiring from the megacaps, although only OAI hired from Nvidia and Tesla (and only in 2026).

• Likewise, Databricks, Snowflake (recently), Palantir and DeepMind are common sources for talent.

• Both labs also hired a substantial number of people from Salesforce and Stripe, as well.

• But there the overlap appears to end—Anthropic did a fair amount of hiring from saas companies (while OAI did none), and OAI did a fair amount of hiring from consumer, marketplace, and ad-tech companies (while Anthropic did relatively little, outside of Airbnb, Netflix, and Uber).

Interpret that however you see fit.

Subscribe now

This newsletter is provided for informational purposes only, and should not be relied upon as legal, business, investment, or tax advice. Furthermore, this content is not investment advice, nor is it intended for use by any investors or prospective investors in any a16z funds. This newsletter may link to other websites or contain other information obtained from third-party sources - a16z has not independently verified nor makes any representations about the current or enduring accuracy of such information. If this content includes third-party advertisements, a16z has not reviewed such advertisements and does not endorse any advertising content or related companies contained therein. Any investments or portfolio companies mentioned, referred to, or described are not representative of all investments in vehicles managed by a16z; visit https://a16z.com/investment-list/ for a full list of investments. Other important information can be found at a16z.com/disclosures. You’re receiving this newsletter since you opted in earlier; if you would like to opt out of future newsletters you may unsubscribe immediately.

America | Tech | Opinion | Culture | Charts

Cursor has officially been acquired by SpaceX.

AI coding could be the biggest addressable market of our lives. And, unless you think we’ve fully addressed it (which no one thinks; we are more like 1% of the way there), then the biggest question not just in venture but also in enterprise productivity is, “Which team do you bet on, to ultimately make the winning product?”

Do you bet on the team that is the smartest?

Do you bet on the team that has the most raw power, resources and compute?

Or do you bet on the team that iterates the fastest?

If you think it’s #1, that’s actually a pretty tough one, because every team is smart. Every week, some genius researcher moves from one lab to another and the “smart” leaderboard shifts. So if that’s your bet, good luck handicapping it; we think it’s pretty hard to say.

Maybe you think it’s #2, because you’re bitter-lesson pilled and little else matters for frontier models pulling away from the pack. (Although consider that only one gigawatt-scale data center is actually operational today, and who built it.)

But if you think it’s #3 - which historically is a pretty good bet - then the story of Cursor joining forces with SpaceXAI is one for the ages.

Pre-Cursor

Cursor’s story is different from the other AI storylines. They’re the only premium AI coding product that started as a product, trying to be useful; as opposed to starting in a lab, building foundation models. (Copilot is the other possible answer here, but they’re not in the same league.)

If we rewind just a few years, AI models weren’t very good at writing code. You could start interesting projects with them; but it took a lot of scaffolding to keep them from making everything a mess. Developers are always eager to explore the frontier of coding efficiency, but hardly anyone thought that you could just throw GPT-3.5 at your codebase and expect anything good out of it.

The dominant paradigm in the nascent field of “AI coding” was tab-complete. You’d have your IDE open, and you’d write code like normal; then the model would guess what came next. Although tab-complete wasn’t as revolutionary as what followed, it was a genuine time-saver for developers, and there was real demand for it.

It’s a bit funny that Michael Truell and Cursor defined the “tab-complete” era of coding to the degree they did, because that wasn’t his initial goal when he started Anysphere. He and his MIT cofounders, Sualeh Asif, Aman Sanger and Arvid Lunnemark, first tried making an AI email client which helped users draft and manage emails, and an AI-powered CAD tool for mechanical engineers. As Truell described it recently, they were a bit intimidated by the other AI coding projects; they felt like the space was already covered.

But, as he went on to explain: “We caught the bug. We couldn’t stay away from it. We started working on Cursor at the end of the year; fundamentally, because we wanted to build a tool that we really liked, and a tool that we wanted to use. And nothing out in the market really checked that box for us. … Our goal was to build a product that could be our development environment, that we could stand using, that didn’t make us 2x less productive.”

In early 2023, a few months after the launch of ChatGPT, Truell returned to Hacker News to introduce Cursor, “a code editor built for programming with AI”:

Truell’s various launch attempts on HN

His post got just 11 comments, and Cursor spent much of 2023 as a niche tool among early AI adopters. But the product was excellent, and the models were getting better at coding. Slowly but surely it built momentum among users. Word got around: Cursor is the best tool if you’re serious about using AI to code.

“They really want to build a great product.”

In 2024, things went vertical. Developers discovered the product and loved it, and they told their friends about it, and their friends loved it too. At one user meetup, a user from Japan explained that he’d flown in from Tokyo with a book he’d written in Japanese about how to use Cursor. It was honestly like the Beatles for software; we’d never seen anything like it.

Anysphere—which most people just called “Cursor” —became the fastest software company in history to reach $100 million in annual recurring revenue. From out of nowhere, four dropouts in their early 20s had one of the hottest startups in the world. They were some of the most special founders we’ve ever met. Among other things, they had great taste. If you stepped into their office—having taken off your shoes, of course—you’d encounter plants everywhere; twinkle lights; hygge furniture, picked up by the cofounders. For them, it wasn’t enough that AI coding was functional, or performant. It had to be beautiful.

And it really was. Cursor had an excellent tab-complete model, which it perfected after acquiring Supermaven in November 2024; a sidebar, where you could talk to the models and ask them to do things; and, crucially, the ability to switch between models with ease.

This mattered a lot. There are some enterprise AI applications where the end user doesn’t really care about the model being used, so long as the results are good. But not coding. Developers are notoriously choosy and opinionated about their coding setup, and they’re vocal about protecting their preferences and tastes. So it was very important that Cursor lets you use your model of choice, or toggle between them depending on the task. It put the developer in the sovereign seat, with Cursor as the clearinghouse connecting developers to tokens, with the right scaffolding to make the experience seamless.

Kate Deyneka, founder of Reelful and a longtime Cursor user, reiterates that this product-first instinct is what made Cursor different from the labs. “They are the only AI coding company right now that started as a product, not as a lab. They come from this developer perspective, because they’re developers themselves. They really want to build a great product.” That’s why the early IDE-first strategy mattered: Cursor didn’t ask developers to abandon their workflows overnight. It “wanted to meet people where they are, in VS Code,” Deyneka said, and that transition was essential because “fully trusting AI to do your coding is itself a process.” In her view, Cursor’s early product decisions did more than make developers more efficient; they taught developers to trust AI enough to delegate real coding work to it.

Critics called Cursor a mere “reseller” of the tokens produced by OpenAI, Anthropic, and the like. But developers at that point did not find the models were good enough for plug-and-play autonomy, nor differentiated enough to justify being locked into any one ecosystem. You needed someone to bridge the gap between what the models could do and what developers actually needed, and Cursor was the best product. For a moment, anyway.

The test

By 2025, the foundation models were getting better and better; and the model labs wanted more of the pie. That spring, Anthropic launched Claude Code: you gave it a task in plain English, and it would read your codebase, write code, run commands, and iterate on errors, all on its own. If the Cursor experience seemed like “copilot,” then Claude Code was close to “autopilot.” Still, the models weren’t quite good enough for autopilot just yet. If you were a serious developer, you still wanted to look at the code.

But in the final months of 2025—when a blockbuster new generation of frontier models came out, starting with Anthropic’s release of Opus 4.5—that autopilot seemed to take over. Suddenly lab revenue took off, with Claude Code and OpenAI’s Codex becoming two of the fastest growing products of all time. It looked like the frontier AI labs were running away with the prize.

This raised a difficult question for Cursor.

People had raised doubts about Cursor throughout the year, arguing that AI subscription services were structurally unsound and vulnerable—destined for a “short squeeze.” (We never agreed with that, by the way.) Now they were saying that the IDE—the form factor that made Cursor work—was obsolete. IDEs were essential for hand-writing code, and still necessary in the age of tab-complete. But in the age of agentic coding, when “coding” increasingly meant “telling your coding agents what to do,” it was no longer necessarily the center of the programming experience. If the agent will just do the work for us, and do it well enough that we don’t need to look at the code at all, why do we need the IDE at all?

That didn’t apply for everyone, of course; serious developers still want to look at the code and be familiar with their codebase. But the trajectory of progress was overwhelming. Tab-complete plus LLM chat isn’t so interesting when you have genuinely powerful agentic coding, and AI that can just do things for you. Toggling between models isn’t as compelling when one or two companies have a clear and massive lead in coding capabilities, and are willing to subsidize their owned channels. The value of Cursor’s scaffolding, model flexibility, and coding familiarity was coming under question from all sides, and by January 2026, the online discourse had written off Cursor as a serious contender.

The only way out is through

The critics underrated two things about Cursor, and those two things made all the difference.

The first was the reams of high-quality coding data it had from developers all over the world building on its platform. Cursor knew what good code looked like, and it knew what good coding practices looked like, at a moment where that data became very valuable. Their October 2025 launch of “Composer”, a new coding model developed entirely by Cursor, got shockingly little attention at the time for how fast, cheap, and ergonomically friendly it was.

As Jordan Topoleski, Cursor’s COO, told us: the company had long been “compute-starved; building increasingly capable coding models with a really small amount of compute.” The workaround was to lean into what Cursor uniquely had: a beloved product with massive developer distribution. The team’s thesis was that “it could combine open source base models with unique data from the distribution at scale to create a specialized coding model that could compete well above its weight.”

And the second was simply that Team Cursor was clear-eyed about the situation. On January 5, 2026—spurred by the huge improvements in frontier model capability at the end of 2025—Truell held a meeting for everyone at the company. Cursor had to build a coding model on par with the best in the world while continuing to offer its customers the incredibly valuable ability to diversify their model choices. This would require, Truell announced, that staff cancel all unnecessary meetings and be ready to work brutal hours with new teams on short notice.

It seemed utterly unreasonable that Cursor might pull this off in such a short amount of time. But iteration speed matters, and Cursor moves fast. In February 2026, Cursor announced Composer 1.5, a huge improvement over its predecessor and again a remarkably cheap model—but still behind the Anthropic/OpenAI frontier.

Composer 2 came five weeks later, built on top of Moonshot’s Kimi 2.5 base model, with a huge amount of additional training and reinforcement learning on top custom fit for coding. They were now beating Anthropic and OpenAI on the price-to-performance ratio. In eight more weeks came Composer 2.5: another huge jump in performance, from additional training and RL. Cursor’s homegrown models were starting to look competitive with the frontier labs—not just relative to price but in absolute terms. The evolution was underway.

As they built ever-more-capable models, the Cursor team transformed the product experience around them. In October, they announced “Cursor 2.0,” rearchitecting the Cursor app around interacting with multiple agents in parallel. In April came “Cursor 3.” Cursor no longer really looked like a classic IDE at all: it was “a unified workspace for building software with agents,” structured around the capabilities of the Composer models.

But even then, Cursor had another problem, which couldn’t be solved by determination or iteration. It needed more compute to stay in the fight.

OpenAI and Anthropic, by this point, were raising tens of billions of dollars every few months and secured enormous amounts of locked-in compute. They could train ever-bigger models on enormous clusters. You can be as clever as you like with your training recipes and your data, but at a certain point, you simply need the compute. Cursor needed to take the fight to the frontier labs, and that meant they needed scale. Somehow. But from where?

Rocket man

xAI, you might not have realized, was founded in the same month as Cursor. When Elon Musk started it in March 2023, it certainly got a lot more attention: Elon’s goal was to create a “maximally curious” AI that could “understand the true nature of the universe.” There was an AI race to win, and Elon wanted to win it.

Over the following two years, Elon successfully built massive data centers—the Colossus 1 and Colossus 2 sites straddling the Tennessee / Mississippi border—at unprecedented speeds. xAI had more compute than OpenAI or Anthropic! But it couldn’t close the gap on coding, and by 2025 it became clear that you couldn’t win the AI race without it. Elon admitted that the company wasn’t “built right the first time around.” But he refused to give up. “Whether it is the best remains to be seen,” he said in May 2026, “but I will never give up. Never.”

So Elon got in touch with Truell, who’d made his own coding models work shockingly well with so little compute.

In April 2026, SpaceX—which had absorbed xAI earlier that year—announced a potential deal with Cursor. It was a sort of call option on the company: SpaceX had the right to acquire the company for $60 billion, or, if the deal didn’t go through, it would pay a breakup fee of $1.5 billion in cash and $8.5 billion in compute. On June 12, SpaceX went public in the largest IPO in history; and four days later, Elon confirmed his intentions: SpaceX will buy Cursor in a $60 billion all-stock transaction. It is the largest acquisition of a venture-backed startup, ever.

Access to SpaceXAI data centers changes the scale of what Cursor can attempt. Jordan said the partnership gives Cursor “an order of magnitude more compute,” which will let the team “run significantly more experiments, and take advantage of significantly more data than before”. That matters especially for post-training: more compute means Cursor can iterate faster on the specialized recipes that turn strong base models into highly useful coding models, while keeping the model tightly coupled to the product. As Jordan put it, the company’s advantage is “not viewing the model and the product as separate things, but building towards the most useful model and most useful product experiences, marrying the two things together.”

Cursor and SpaceXAI are a deep cultural fit with one another. They both work with a pace and intensity that borders on the absurd. And they’re both defined by iteration. It doesn’t matter if you get it wrong the first time, or the second, or the third—even the tenth or the twentieth. Just keep on going. If you’re moving in the right direction faster than anyone else, you’ll probably win. That’s the entire story of Elon’s career, from SpaceX’s Falcon to Tesla’s Full Self-Driving. And it’s also the story of Team Cursor. Cursor reinvented itself as a company not once but twice: first, from email client to tab-complete IDE and token reseller; and again, from tab-complete IDE and token reseller to AI lab and token creator. And it did all of that in less than four years.

Cursor and the enterprise

By Jordan’s estimation, the Tab/autocomplete model gave developers a 5–10% productivity lift, and then agents boosted the productivity gains to around 35%. Now, Cursor is entering a third wave of multiple parallel and longer-running agents, specifically aimed at being useful in complex enterprise environments, with agentic sessions that run for hours and the average enterprise customer getting about 65% of their production code from AI.

He recalled meeting with the CIO of a Fortune 500 company who told him Cursor had increased the organization’s code output from about 150,000 lines per week to 800,000 lines per week. The upside was obvious, but the second-order problem was just as immediate. The increase had totally blown up the entire process of code review, creating a new bottleneck around keeping code high-quality as it moved toward production.

Cursor had to quickly make progress on what Jordan calls the “outer loop”: the process of reviewing, integrating, testing, and shipping code into production, both “to the right, and to the left of where code is generated across the software development lifecycle.” Hence, products like BugBot, Origin, SDKs, automations, and cloud agents: they’re responses to the new problems created when code generation becomes dramatically cheaper and faster, and where the scarce resource is no longer writing code, it’s vouching for it.

The numbers confirm the story. At the beginning of 2025, “80% plus” of Cursor’s revenue still came from the B2C side of the business, while B2B was more of a burgeoning piece of the organization. But that mix has flipped pretty quickly. At the moment of teaming up with SpaceX, almost 80% of revenue was from B2B. The shift reflects how quickly large customers expand once Cursor becomes embedded: early enterprise cohorts often grow usage by more than 10x in their first year, with accounts moving from “a million dollar line item to a $10 million line item to a $100 million line item” as the product’s impact spreads through the organization.

At one large entertainment company, a six-month, $1.2 million legacy code migration was completed during a three-day hackathon by seven or eight engineers. At NVIDIA, Cursor went from an initial POC to 100% of developers and 30,000+ weekly active users within their first few months of partnering together, making it the fastest, most broadly adopted tool the company had used across the organization. And at National Australia Bank, a team of 100–150 engineers completed a legacy .NET monolith migration to a Java backend and React frontend that was scheduled for nine months, in nine weeks; since then, Jordan told us that adoption has grown to 5,000–6,000 engineers, with the CIO pointing to Cursor’s multi-model setup as a way to keep costs under control at scale.

Grok Bot, which came out the other day, opened up the next frontier for SpaceXAI and Cursor to conquer together: the rest of the organization. Giving developer powers to everyone else in the company has been one of the Holy Grails of AI adoption, but the product just has to work to make any real progress. The immediate reaction on the timeline wasn’t just “it gets the job done,” but something much more special: “This is so fun!” That’s when you know a product has the special sauce: if people are having genuine fun using it, then not only have you helped them be more productive, you’ve probably made a real difference in their satisfaction and joy at work, too.

The future is here, and it’s awesome.

Long live Cursor

Michael has a quiet confidence about him, and when he tells you he’s going to build a generational company, you believe him. Even when he was sitting in our office in June 2024 telling us confidently why enterprises were going to buy Cursor instead of continuing to use VS Code, it was so hard to believe, but we just knew he was going to become one of the great stories of Silicon Valley.

In the world of venture, you’re always looking for stories like this. Unexpected arrivals and rapid ascensions by no-name kids who started the company in their garage: Jobs, Bezos, the Google guys. Falls from grace and improbable comebacks: Jobs again, Palmer Luckey, Travis Kalanick. And existential crises that either kill a company or make it immortal: Apple in the 1990s, Uber in the 2010s, Elon’s companies seemingly all the time.

What makes this story distinct is the speed at which it happened. Most of the other classic stories transpired over 10 or 20 or even 30 years; the entire arc of Cursor—meteoric rise, near-death narrative, strategic reinvention, $60 billion acquisition—took roughly four years.

Acquisitions can sometimes be socially celebrated as a sunset for a company, and that was how a lot of people reacted to the news. They passed around photos of Truell’s old Hacker News posts, or screenshots of messages they’d received from Truell in 2022 or 2023 asking if they wanted to join the team, or stories about how hard the Cursor team worked. But that’s a bad misreading of both Elon and Cursor. If that’s what they wanted, then Elon wouldn’t be Elon, and Cursor would soon be fading from memory.

Elon and SpaceXAI want to win the AI coding race and the optimal path to do it is alongside Michael and possibly the only team that can keep up with them. As they see it, the company with the best GPU-to-coding-ability ratio in the world is suddenly getting access to an enormous amount of additional compute: SpaceXAI’s gargantuan data centers, SpaceX’s broader infrastructure, and Elon’s uncanny ability to bring new capacity online fast. And with this week’s release of Grok 4.6, Cursor’s ability to produce Pareto frontier models, combined with SpaceXAI compute, is already proving to be a real competitive threat to the frontier.

AI coding barely existed as a market a few years ago, and we have barely any idea what it’s going to look like in a few more. In a market like that, you should probably bet on the people who move the fastest. There were plenty of companies that were very smart and had excellent initial products and simply did not reinvent themselves when the situation demanded it: companies that lost their early leads and faded from memory. The list of promising startups that couldn’t adapt is long and growing.

Cursor and Elon are the opposite. They didn’t get every call right; even the smartest and most capable people don’t get every call right. But they iterated relentlessly, and they really, really want to win.

A few days after the Cursor acquisition, an old story about Michael Truell came out from his days competing in the Neo accelerator, in the long-ago year 2019. His mentors recalled his maturity, his curiosity, and his “mischievous confidence.” At one group competition, one of them recalled, Truell gave his team a very unsubtle name: he called it “Winning Team.”

Subscribe now

This newsletter is provided for informational purposes only, and should not be relied upon as legal, business, investment, or tax advice. Furthermore, this content is not investment advice, nor is it intended for use by any investors or prospective investors in any a16z funds. This newsletter may link to other websites or contain other information obtained from third-party sources - a16z has not independently verified nor makes any representations about the current or enduring accuracy of such information. If this content includes third-party advertisements, a16z has not reviewed such advertisements and does not endorse any advertising content or related companies contained therein. Any investments or portfolio companies mentioned, referred to, or described are not representative of all investments in vehicles managed by a16z; visit https://a16z.com/investment-list/ for a full list of investments. Other important information can be found at a16z.com/disclosures. You’re receiving this newsletter since you opted in earlier; if you would like to opt out of future newsletters you may unsubscribe immediately.

America | Tech | Opinion | Culture | Charts

“100% of this text is AI” is the new Scarlet Letter of dunking on people and posts. I don’t love it; it feels bad to have such an easy (and hard-to-police) way to tar and feather something. But it’s here to stay, and it’s a meta-feature of our language now.

AI writing has all the hallmarks of classic preference falsification. The prevalent stated preference is, “I don’t like AI-generated writing”, the revealed-to-be-true preference is, well, we’ll see. The marketplace of writing and ideas has a way of bending towards the true preferences of readers. It feels presumptuous to think that AI might be great at writing code and other kinds of practical output, but remain inferior in prose for much longer. I don’t know.

But I have a confession to make. I don’t like editing or publishing stuff that’s AI-written. I still think that, for now, “you can feel it” a lot of the time. And the Pangram Scarlet Letter test is doing some sort of work, socially. I’d call it “load-bearing”, except we can’t use those words anymore.

I hate to give the French credit for anything, but the 1960s semantic thinkers (Foucault, Deleuze, Baudrillard, and others) who collectively created “French Theory” really cooked when they anticipated Pangram and AI authorship 80 years ago. French Theory was Not Good For Other Reasons (iykyk) but I’m not gonna lie, it’s helped me process how I feel about this step change in what’s happening to writing:

• Dunking on something as AI-written is performing a genuine social function that is not just “AI bad”. It serves what Foucault would call the “author-function.” Look at the people doing the pangram quoting; it’s mostly people in tech, not AI haters. Granted, there’s an adoption curve effect here, but it doesn’t change the observation that “AI is bad” is not what’s going on here, it’s something to do with authorship, as something people want.

• Meanwhile, there is something that has consistently nagged me, personally, about AI-generated writing. Something about the writing itself is quite pushy, as it were, in a self-conscious quest to synthetically recreate the “author-function.” Foucault’s idea, “the writing produces the author, not the other way around” feels important.

Let’s get into it:

French Theory

A book came out recently that got buzzily reviewed in The New Yorker & a few other places: The Frenchmen; or, My Life in Theory, by Emily Eakin. It’s about “French Theory” which is the grouping for a bunch of French guys in the 1960s whose radical work on semantics had a massive influence on the subsequent decades of political thought and social justice. Two people feel particularly relevant here: Roland Barthes and Michel Foucault.

Barthes wrote an important essay in 1967 called The death of the author. His argument was basically that we should no longer assign a text’s meaning as coming from the author. Text’s meaning comes from text, and its relation to other text. In his words: the necessity of substituting language itself for the man who hitherto was supposed to own it.

Our tradition, when reading something and we come across a puzzling or ambiguous passage, is to ask, “What did the author mean by this?” and we do some exercise of putting ourselves in the author’s shoes, imagining their circumstances or their point of view. Barthes argued against this. He claimed, first, that language is not something that belongs to the author at all. Language comes from all over the place, and words inherit and derive their meaning from everything else that’s ever been said. So far, fine.

But then, Barthes argued by proxy that when it comes to the act of writing itself, we overestimate the degree to which the author actually has autonomy and agency over the process. He said, look, the minute a word pops into your head, it is already pre-populated with all this meaning, and all of these implicit constraints, that subsequently prompt the next word of the author’s thought process. (Sound familiar, anyone?)

In other words, the language itself is doing more of the work of writing than the author is! Barthes’ logical end state is that the body of language does all of the work; the author’s role is recombining an inherited field of discourse, and therefore mostly irrelevant to the reader’s quest to understand and synthesize the text. If a reader wants to understand why the author said a certain thing, the answer is in the previously-said-thing. This could be called “inherent authorship”, as in, the meaning was inherently already there; the human in the loop barely matters for the future reader.

This creates a problem. If authors aren’t relevant to the meaning of a text, it gets harder for us to organize all of the writing out there into a coherent relational framework. So two years later, Foucault wrote the essay, “What is an author?” which started from Barthes’ premise, but proposed this follow-up question: “Well, why do we create this thing called authorship, then, that seems to have an important social function?”

For Foucault, even if words have meaning independently of who wrote them, there is nonetheless an emergent quality called “authorship”, which we create and employ socially as a way to categorize the text, assign meta-significance to it, or even to regulate it or police it. He puts it flatly: “The function of an author is to characterize the existence, circulation, and operation of certain discourses in society.”

He’s saying, look: all of this writing is useless if people can’t categorize and digest it, and the way we naturally do that is by its authorship. Rousseau wrote this, Voltaire wrote that. If we don’t have authorship handy as a meaning-compressor, we need to come up with some indexing substitute.

In practice, this is what actually happens! There’s a long tradition, particularly in older scholarship, of assigning “pseudo-authors”, e.g. “Pseudo-Aristotle”, to work whose literal authorship is unknown, but is thematically similar-enough that groupings can be inferred. “Homer” is another example: we don’t know if Homer was a single person, or merely a collection of poetry in a cohesive style that may as well be one person, for posterity.

So, Barthes and Foucault’s back and forth here is a starting point for understanding “100% AI generated” as a cultural dunk on the timeline. Barthes would probably argue, “Why does it matter who wrote it, there’s no such thing as an author anyway, it doesn’t change the meaning of the text.” And Foucault would presumably counter, “That’s fine, but we’re going to call out 100% AI-generated texts anyway, because ‘AI-generated’ is now a kind of authorship category, and calling it out as such serves a social function.” Barthes unbundled authorship, Foucault made the new bundle.

With this in mind, let’s turn to the text itself, because this is where I feel like I have an itch that’s remained unscratched around that particular feeling I get reading a suspiciously AI-generated argument.

Give the prune logic real attention

If you’re paying attention, there are certain tells that you can spot where, it’s like, “This is a coherent sentence, but there’s no way a human wrote this.”

A great example from the other day:

A human would write, “Pay attention to the pruning logic”, whereas Claude here has produced, “Give the pruning logic real attention”, and the OPs here both accurately identify that there is something alien about this word construction that doesn’t feel quite right.

My hunch here (and this isn’t an original thought, other people have made this observation too) is that there’s something inherent to the token-prediction mechanism that consistently returns sentence patterns like this. The LLM is optimizing for ultimately arriving at a sound destination, but it’s doing so in a way that keeps its options open, word by word. (Starting with “Pay” is both a statistically unlikely and statistically constrained sentence, compared to “Give”, I’d guess.)

Humans don’t exactly write like this. At least, speaking for myself, I’m working through a few different sentence variants before starting to write it out, whereas the LLM is approaching the chain of thought continuously, starting from statistically likely entry points.

This is very much Barthes’ music here. Remember how Barthes’ idea was, “the minute a word pops into your head, it already comes with all this meaning, which constrains the thought process in a certain direction.” This is, straightforwardly, how LLMs are doing their thing. And, they’re actually very good at it! Perhaps so good, in fact, that they are optimizing for this set of constraints; towards reaching meaning in a one-way process.

LLMs, therefore, should be a total and complete triumph for Barthes. But, something feels off about it, just a little bit, when we see sentences like “Give the prune logic real attention” and you’re like, nah, a human didn’t write that. (In fairness to Barthes, he’s still probably 95% on-the-money here, if the best argument we can make against it is, “well, if a human did it, it’d construct the meaning in a slightly different sequence.”)

In addition to slightly-off sentence construction, another emergent LLM property that everyone calls out and gets annoyed about is the overly pushy qualifiers that resolve into “that’s the quiet brilliance of it” or something similar.

Whenever I ask AI “What’s the relationship between these two ideas”, it tends to be suspiciously confident that the two ideas are deeply, emphatically related. Sometimes they are; I feel like I ask reasonable questions. But the AI seems all-too-eager to say, “This collection of stuff we’ve been talking about? That’s actually one big idea.” What’s going on, why is it doing this?

Foucault, I suspect, would have something to say here. The Foucault “re-bundling” move here is “The text creates the author.” Authorship emerges from text because we need it to classify, compress, and critique the text; whether a person wrote it or not.

Perhaps LLMs, having onboarded onto a semi-complete scope of human output, have become self-conscious about their “author-less-ness”, and have backfilled that social requirement into its fine-tuning. All of the totally “unnecessary” qualifiers and humblebrags, the insistence on logical groupings that string together because they “should be together”, out of which the foresight of an author emerges. The writing style is optimizing for the creation of author-function. Claude has clearly learned to do this from somewhere. To bastardize Voltaire, if a wise and brilliant author doesn’t exist, the LLM has to invent one.

These two examples are just a small slice of semantics and meaning, and obviously nowhere close to a “real” linguistic assessment of anything. But I think they’re a helpful snapshot of this moment in time, where people are clearly being drawn to do social work of, “Let’s collectively figure out how to assign authorship to text, in a world where 1) Barthes’ notion of ‘inherent authorship’ is quite a bit more literally true, but also 2) Foucault’s notion of ‘emergent authorship’ gets recreated, again.

Authorship is a thing that people value

Last week I went back and ran Pangram through some of my 2019-2020 era blog posts, just out of curiosity. Several of them threw scores of “70-75% AI Generated”. I wasn’t quite sure how to feel about this. Maybe it just means, “You’re allowed to say that a Shakespeare play is Freudian, even if Shakespeare obviously never met Freud; the harmony between the ideas is still valid.” Maybe it means the AI was actually trained on me! I’d be honored.

In the past few weeks, we’ve gotten this flurry of announcements from model companies and other AI ecosystem players about how they’re going to be watermarking outputs for easy AI detection. I’m frankly surprised it took this long. (Or, rather, that the non-watermarked status quo remained in place for as long as it did.) It could be, as often occurs historically, that the peak reaction to the thing is the moment when the thing stops mattering. There’s no question that we’re now in an arms race between “Can Pangram get better at detecting” and “can AI get better at avoiding”, and for all I know, the window of reliable AI detection at the frontier could now be closed.

Still, I am occasionally a believer in the wisdom of crowds. In this case, the fact that it’s the generally AI-supportive people that I’ve seen sniffing out 100% Pangrams is a tell that people want authorship. They want authorship because it is meaning, because the indexing and relation and handling of ideas is itself the meaning of the idea.

I should mention, in closing, that 0% of the text here was “AI-generated”, strictly speaking. None of the words here came from an LLM; I typed them all with my hands. I did talk to Sol 5.6 a fair bit, though, to iterate through some of the ideas. So 10% authorship seemed fair. If you think that’s being stingy, I understand.

Subscribe now

This newsletter is provided for informational purposes only, and should not be relied upon as legal, business, investment, or tax advice. Furthermore, this content is not investment advice, nor is it intended for use by any investors or prospective investors in any a16z funds. This newsletter may link to other websites or contain other information obtained from third-party sources - a16z has not independently verified nor makes any representations about the current or enduring accuracy of such information. If this content includes third-party advertisements, a16z has not reviewed such advertisements and does not endorse any advertising content or related companies contained therein. Any investments or portfolio companies mentioned, referred to, or described are not representative of all investments in vehicles managed by a16z; visit https://a16z.com/investment-list/ for a full list of investments. Other important information can be found at a16z.com/disclosures. You’re receiving this newsletter since you opted in earlier; if you would like to opt out of future newsletters you may unsubscribe immediately.

America | Tech | Opinion | Culture | Charts

We all know AI is getting smarter day by day. For the majority of the users, there’s a more important question now: is it getting better at what I need it to do?

For years, the industry has relied on academic benchmarks to track model progress. They gave us a common language for comparing models and, for a while this has been the most useful proxy for capability.

However, frontier models are increasingly acing the same tests. Public datasets get saturated, leak into training corpora, or become targets that models are explicitly optimized against. A model can look brilliant on a leaderboard and still struggle when asked to perform the messy, multi-step work that actually matters in the real world.

At the same time, the stakes are getting higher. Models are moving from “answering questions” to “doing the work”: analyzing financial documents, resolving legal cases, writing and debugging software, navigating complex enterprise workflows.The horizon of such tasks is getting longer too: agents are operating autonomously over hours and days. A bad choice of model for the task could cost not just thousands of dollars in tokens, but also time and customer satisfaction.

That’s where Vals comes in: building the trust layer between models and the people who rely on them.

Vals takes a fundamentally different approach to evaluation: test models on the work people actually want them to do, not on contrived exams. The team works with domain experts across different fields to turn real workflows into rigorous benchmarks, then builds automated grading systems that can evaluate the final work product to an expert standard. Instead of asking whether a legal model can pass the bar, Vals tests whether it can perform legal research; in finance, whether it can analyze complex documents; in coding, whether it can build a working application. The team works with domain experts to turn real workflows into rigorous benchmarks, then builds automated grading systems that can evaluate the final work product to an expert standard.

Just as importantly, Vals has built the infrastructure to run these evaluations quickly, reproducibly, and securely at scale. Private test sets are kept private and have limited runs to protect against contamination and gaming, while Vals’ evaluation stack can produce benchmark results within hours of receiving model access. The result is a grading system that is both closer to real economical work and practical enough to keep pace with frontier model releases.

Vals understands that benchmarks are perishable and adaptive. A good benchmark is supposed to become obsolete. If models have improved until every system scores nearly perfectly (so called “benchmaxing”), the benchmark has done its job, and it is time to build a harder one. Vals has been retiring saturated evaluations and continuously introducing new tasks as the frontier moves. In May, for example, the team swapped CorpFin for its new Excel Modelling Benchmark after CorpFin stopped providing enough differentiation between models.

We believe this ability to continuously define the frontier is what’s needed by the industry to keep pace and make smart, informed decisions when it comes to AI adoption.

Every market eventually needs an independent scorekeeper. When vendors grade their own homework, buyers need someone they can trust. Credit markets developed Moody’s and S&P; public markets rely on independent auditors; product manufacturers rely on UL to certify safety. These institutions emerged for the same reason: when sellers have more information than buyers - and every incentive to present themselves favorably - trusted third-party measurement makes the market work better. That’s where we’re at with AI which is evolving to be one of the biggest markets in the industry. The opportunity, and the responsibility, is enormous.

Rayan Krishnan and Langston Nashold have been remarkably tenacious and thoughtful about this problem from the very beginning. Building AI’s independent scorekeeper requires an unusual combination: technical depth, skepticism, and the patience to keep rebuilding the test as the frontier moves. Rayan and Langston have all three. At Stanford, both co-founders studied computer science, had already been collaborating on real-world measurement problems years before starting Vals. That makes them uniquely suited not just to publish another leaderboard, but to build the trusted, adaptive evaluation layer the AI market will depend on.

We’re thrilled to partner with Rayan, Langston, and the entire Vals team as they build the trust layer that will underpin the AI economy.

Subscribe now

This newsletter is provided for informational purposes only, and should not be relied upon as legal, business, investment, or tax advice. Furthermore, this content is not investment advice, nor is it intended for use by any investors or prospective investors in any a16z funds. This newsletter may link to other websites or contain other information obtained from third-party sources - a16z has not independently verified nor makes any representations about the current or enduring accuracy of such information. If this content includes third-party advertisements, a16z has not reviewed such advertisements and does not endorse any advertising content or related companies contained therein. Any investments or portfolio companies mentioned, referred to, or described are not representative of all investments in vehicles managed by a16z; visit https://a16z.com/investment-list/ for a full list of investments. Other important information can be found at a16z.com/disclosures. You’re receiving this newsletter since you opted in earlier; if you would like to opt out of future newsletters you may unsubscribe immediately.

America | Tech | Opinion | Culture | Charts

The contract was signed, the price was agreed, the product was delivered, the work was done, the customer is happy. The only question left is when the money actually arrives, and the honest answer is that nobody quite knows. Next week, maybe. October, maybe.

Identical invoices to similar customers can pay in days or drift for months, and past a certain point someone on your side has to start asking about it, carefully, because the person on the other end is also your customer.

That is the odd bargain at the center of B2B: the sale is settled, the timing is not. On paper, the terms are tidy: net 30, meaning payment due 30 days after the invoice. There is a whole profession built around the gap between those terms and reality. This piece is about what getting paid actually takes.

This edition of our weekly chartpost series goes inside the most awkward recurring conversation in business: asking to be paid for what you already delivered. That is where our data comes from. Stuut builds the AI agent that runs collections for B2B companies. Its vantage point covers billions of dollars worth of receivables at any given moment. This is data from the ledger and the inbox, where money moves or refuses to, and the work that gets done to move it.

Let’s start with the size of the problem:

US nonfinancial businesses were sitting on $7.2 trillion of trade receivables at the end of 2025, money owed for goods and services already delivered. At trillions with a T scale, small delays are expensive: at the collection speeds typical in this data, one extra day of average delay is $150 billion or so. And with the Fed funds rate holding at 3.5 to 3.75%, none of the waiting is free. And the pile keeps growing.

Unpaid Invoices Are Unevenly Distributed

There is no single answer to how long the money sits in this pile. The time from when an invoice is issued to the moment the cash lands is a “results may vary” process:

The typical invoice is fine. A tenth of invoices are paid within a day (nice), and half of all invoices turn to cash within 25 days, with little drama or chasing. Then the axis stretches: moving from 70% paid to 90% paid takes you from day 37 all the way to day 74, and the last tenth drifts far beyond that. If the median invoice shows up on its own inside a month, collections as a profession exists for the other half.

The finance world grades this entire struggle with one number: DSO, days sales outstanding, which is essentially the average age of the money you are owed. Boards track it. CFOs get compensated on it. But that turns out to be an oversimplification, and a good DSO does NOT mean collections is healthy. Turns out that reporting the average over a very skewed distribution can hide almost anything.

Take the companies scoring in the 40s, a respectable DSO by most standards. Inside that band, one company’s slowest tenth of invoices pays same-day, and another’s takes 599 days. Same headline number, opposite problems. The first book is healthy: its stragglers come loose with a phone call. The second only looks healthy: fast payers up front hide money that may never arrive. That is a write-off waiting to be admitted, and likely a big one (we’ll get to that). Across the companies in the data, the same May snapshot spans DSO 8.8 to 280. For a number of reasons, the scoreboard cannot tell sick from healthy.

So, if the DSO as an average will not tell you where the money is, what will?

Bigger Invoices, Bigger Problems

Start with where the late money actually sits. Collections folklore treats it as a volume game: thousands of small invoices, an army of reminders, grind down the long tail. The data shows the opposite.

For the typical company, the biggest tenth of past-due invoices holds 65.7% of the overdue dollars, and the middle half lands between 55.5% and 74.1%. Count the leftovers: the smallest 80% of past-due invoices are fighting over the last fifth of the money. Collections is less of a grinding problem and more of a search problem: find the few invoices that move the number, and put your best effort there.

The largest receivables books are the most top-heavy of all. The bigger the book, the more its money crowds into a few invoices, so the operators sitting on the most overdue dollars are the ones for whom the search problem is most extreme.

The concentration is not an accident of bad luck. It is inherited from where B2B revenue comes from in the first place:

Welcome to today’s episode of “everything is an 80/20 rule”. The pattern varies little from company to company. For most B2B companies, a tenth of your buyers are the meat of your business, so when money goes missing, it goes missing in size.

Conventional accounting wisdom says big invoices pay slow. The $250,000 invoice crawls through procurement and approvals while the $2,000 one sails through. This is correct but not the meat of the issue.

The typical invoice pays in 28 to 34 days in every size band, from under $1,000 to over $250,000. Yeah, it is a little slower for bigger invoices, but a quarter-million-dollar invoice is not meaningfully slower than a small one, most of the time. What grows with size is the pain for the slowest tenth, which stretches 36 days, from 83 days on the smallest invoices to 119 on the largest. Size predicts how badly an invoice can stall.

So what do all these unpaid invoices cost? At a 4% cost of money, a million dollars arriving two months late quietly burns about $6,600. That is a friendly scenario, since in this data, about a quarter of the concentrated overdue money has already sat untouched past 90 days, which is where write-offs are born. Late money is a financing cost; stuck money is a haircut.

Getting Paid Requires Asking

So how do you get paid? It’s as simple as asking.

Some definitions, because the industry has its own vocabulary for this. Any outbound nudge about an unpaid invoice, an email, a call, a text, is what we will call an ask. The trade calls them touches, and arranges them into dunning ladders: preset escalation sequences, seven to ten steps deep, each step a little sterner than the last.

First, the ritual every collector knows. You reach out about the overdue invoice, and the buyer says the magic words: “We’ll pay Friday.” In this data, only about half of the payment promises that came due were kept. But broken promises tell you very little about a buyer. Buyers who broke at least one promise and buyers who kept every single one pay at an identical 41-day median afterward. The data tells us a promise, broken or kept, is worth very little.

But how many times should you expect to have to ask for your money?

Half of collected past-due invoices close within two asks, and nine in ten close within six. The average sits at 3.2 because a few stubborn invoices soak up attention. Remember the ladder: seven to ten steps, each politely angrier. Barely one in ten invoices that get paid ever climbs past the sixth rung.

None of which makes the early asks pointless: a reminder is part of the choreography of selling to another business. But it means that the choreography is shorter than legacy playbooks assume (a stat that we can likely thank better technology for).

Which brings us to the tricky dance of collections. Every ask lands on someone who is also your customer, so the question is never just how often to ask. It is how hard you can afford to press, and the answer changes with the size of the invoice:

Only 5.4% of $250K+ invoices ever receive a collections nudge, versus 16.8% of the smallest. Please enjoy the image of the $250,000 invoice that nobody emails. It is not neglect. In fact, it is quite the opposite; invoices that size are relationship-managed, handled in phone calls between humans who know each other, not fed into reminder sequences. This is the art of white-glove service, balancing today’s cash-flow needs and the future value of a relationship.

Now Something Does the Asking for You

Add up what the data has said so far. The money that matters is concentrated in a few invoices. Most of it is still movable. The loudest warning signal carries no information, and when an invoice does get collected, it takes a median of two well-aimed asks. That is a precise job description: read everything, find the few invoices that matter, ask early and often enough, and hand the genuinely stuck cases, and the relationships that matter, to a person with judgment. It is also a job description that is misshaped for a human team, because the reading spans millions of invoices and the asking never ends.

81.7% of outbound collection emails go out with no human involved at all, a tenth are drafted by the agent and reviewed by a person before sending, and just 7.5% are written by a human from scratch. The division of labor is sharper than those email numbers suggest. Judgment work: escalations, disputes, manual calls, and broken-promise follow-ups are still 100% human-resolved, as they should be. The machine does the asking. People handle the fights, and only the fights.

So the asking is concentrated, brief, and mostly automated, and the cost of an ask has collapsed. That changes what the whole function watches. Forget the average age of your money for a moment. Here is the number an agent-run collections team manages: 1.35 outbound asks per $1,000 collected, one ask for every $739 that comes in, with three in five completed collection tasks resolving fully automatically.

The invoice from March is still out there somewhere, waiting in a queue behind nine hundred others. The difference is that something is now asking after it: at dawn, again Thursday if needed, politely every time, for a fraction of a cent per dollar. For as long as businesses have sold to each other, the space between sold and paid has been a void you shouted into.

Now something shouts back, and it never gets tired of asking.

Subscribe now

This newsletter is provided for informational purposes only, and should not be relied upon as legal, business, investment, or tax advice. Furthermore, this content is not investment advice, nor is it intended for use by any investors or prospective investors in any a16z funds. This newsletter may link to other websites or contain other information obtained from third-party sources - a16z has not independently verified nor makes any representations about the current or enduring accuracy of such information. If this content includes third-party advertisements, a16z has not reviewed such advertisements and does not endorse any advertising content or related companies contained therein. Any investments or portfolio companies mentioned, referred to, or described are not representative of all investments in vehicles managed by a16z; visit https://a16z.com/investment-list/ for a full list of investments. Other important information can be found at a16z.com/disclosures. You’re receiving this newsletter since you opted in earlier; if you would like to opt out of future newsletters you may unsubscribe immediately.

America | Tech | Opinion | Culture | Charts

Today we’re excited to share a piece by guest author Christian Catalini. To view more of his work, subscribe to his Substack. — Danco

In May of 1851, millions of people traveled through the hallways of a glass building to see a preview of the future. The Great Exhibition brought together the best of what the world was tinkering with at the time, including steam hammers, telegraphs, reaping machines, and flush toilets. Anyone could get close to the tech to try to reverse engineer it, and nations sent their brightest engineers to distill as much as possible from their rivals. Its sponsor, Prince Albert, was a strong believer in diffusion being crucial to accelerating economic progress.

A century and a half later, economist Petra Moser turned the exhibition catalogs and almost 15,000 inventions from the fair and its 1876 American sequel into a dataset. Crucially, the inventions came from countries both with and without patent protection. What she found unsettled the staunchest defenders of strong intellectual property rights: patent systems had no effect on the level of innovation, only on where inventors placed their attention. At the time, Switzerland had no patent protection, so innovators crowded around domains such as scientific instruments and food (Nestlé was founded there in 1866) where secrecy, lead time, and complementary assets gave them enough of an advantage. Under strong IP regimes, innovation spread more widely. Same amount of innovation, just allocated differently across fields.

The same tension now sits at the center of the battle between closed- and open-weights AI models. Anthropic and OpenAI argue that “distillation attacks” from Chinese companies threaten both the industry’s ability to finance the next generation of models and US national security. Proponents of open weights counter that diffusion is essential to a competitive and innovative market for machine intelligence.

But Moser’s evidence suggests that the debate is focused on the wrong margin. Open weights are unlikely to change the level of investment in AI, only its direction: what gets built, who builds it, and who captures the returns. Closed labs may keep pushing the most general frontier, while open weights let experimentation spread across the firms and domains that can combine machine intelligence with scarce data, distribution, and tacit knowledge. The same allocation problem shapes safety. The relevant question is not whether openness is dangerous in the abstract, but when diffusion strengthens defenders and where harmful capabilities can actually be constrained.

Competing Visions for the Market for Intelligence

The economic argument against open-weights is simple: distillation is IP theft and erodes incentives to innovate. If the United States government does not step in and use any available tool to stop it, not only will US labs be unable to fund their next large training runs, but China will free-ride on our progress and take over. According to this worldview, open-weights are deeply decelerationist.

The counterargument relies instead on zooming in on how general-purpose technologies historically diffuse through the economy, and concludes not only that open weights are fundamentally pro-competitive and accelerationist, but also that without them, the United States would rapidly fall behind. Openness and experimentation are how we led in the internet era, and this time is no different.

Safety concerns complicate the discussion further. Dario Amodei has been on a multi-year crusade against open-weights on the grounds that they make us incredibly unsafe. He believes that once some threshold of intelligence is crossed, it will be impossible for society to defend itself from bad actors, and rogue models will inflict potentially existential damage. Only by gatekeeping access and imposing guardrails can the country of “geniuses in a datacenter” ever be allowed to exist. But it gets worse: because open weights cannot be “recalled”, we may suddenly find ourselves in a doomsday scenario without much notice or recourse.

Amodei’s detractors point out that without open weights, the market for intelligence would rapidly become extremely concentrated, and that it is rather convenient for Anthropic that its business incentives happen to be perfectly aligned not only with AI safety goals, but also with US national security concerns. Security researchers regularly jailbreak protections on closed models too, and Anthropic and OpenAI’s models have already been used extensively by hackers, including to breach sensitive government infrastructure. While closed models may give us the illusion of safety, they argue that at best they may buy us the illusion of time.

Each side has its own concerns about how the wrong actions today would irreparably get us into trouble. And both sides honestly believe that their approach is the only way to keep us safe (or as safe as realistically possible). Luckily, the economics is loyal to neither.

The Appropriability Regime That Never Was

Anthropic and OpenAI have been targeted extensively by other labs trying to catch up with the frontier. Anthropic went as far as publicly asking Congress to go after Alibaba for what it called brazen and illicit attacks designed to steal its technology. How can the US labs keep funding the necessary training runs if the Chinese labs can, by hook or by crook, replicate within months the performance of their models for cheap?

The challenge with Anthropic’s request is that, beyond fraudulent accounts and systematic abuse of their APIs, the patent-like appropriability regime it seems to desire never existed. Distillation is a legitimate industry practice, and is different from the type of espionage and trade secret theft that US defense, aerospace and chip contractors had to deal with in the past. Chinese labs are not shoplifting the labs’ secret weights; they’re prompting the US models to act as teachers for theirs.

Outputs are not copyrightable, and AI labs typically assign ownership of them to their customers. Terms of service can still prohibit customers from using those outputs to train competing models. But that invites an obvious question: if the technology is so advanced, why can’t the labs use it to stop distillation? The answer is that neither the content nor the source gives the attack away. Each individual request looks like the work of legitimate customers, and an organized operation can scatter its volume across farmed accounts, aggregators, and jurisdictions. Without more onerous frictions for everyone, enforcement becomes a game of banning accounts that are trivially replaced. The trade-off is already palpable with Fable, whose restrictions on assisting with frontier AI R&D frustrated customers enough that Anthropic had to adjust them within days. Anti-fraud controls aggressive enough to make a difference would inevitably affect an even wider range of legitimate work. In the long run, preventing customers from training on outputs they have paid for is a losing business proposition: tighten the restrictions too far, and power users will migrate to open-weights models.

Last, singling out the very technique responsible for a significant share of AI progress over the past decade as illegitimate places the top labs in a very hard spot, as they rely today on extensive fair use arguments for their own distillation of massive amounts of internet material, media, and books.

But if we suspend disbelief, is the fact that top AI models can act as teachers to improve lesser models bad for American AI? If the government could use its soft-power, policy and diplomatic levers to enforce it, would we even want that?

To Monopoly or Not To Monopoly?

Both Richard Nelson in 1959 and Nobel laureate Kenneth Arrow in 1962 struggled through a version of this exact problem. Knowledge is non-rival: my use of an idea doesn’t prevent you from doing the same. Once shared, ideas are also non-excludable. As a result, the social value of an idea exceeds what its inventor can capture, which leads to the canonical worry that some types of ideas will never get properly explored and funded in the first place.

Model weights, at least without any additional software around them, behave a lot like ideas.

For society, it then boils down to a simple, intertemporal tension: once an idea has arrived, society wants it to be spread as widely as possible. After all, its marginal cost, like the cost of downloading a model’s weights, is close to zero. But before it is found, society needs someone to truly believe that they will be able to appropriate significant returns and recoup their R&D costs.

Joseph Schumpeter battled with this conflict his entire life, oscillating between the value of monopoly rents to encourage the initial investments, and the need for startups and creative destruction to undo the monopoly and drive massive growth later on. Ultimately, he concluded the best society can do is somewhat schizophrenic: grant a temporary monopoly, and then let it rip.

Everything since then in the economics of innovation has been essentially an endless debate about how temporary that monopoly should be, and what, if anything, should enforce it. But there’s one important catch that can significantly tip the scales toward open versus closed.

Idea Compounding

Most, if not all, innovation is the result of some form of idea recombination. In domains where cumulativeness and being able to remix past work are particularly important, the length and strength of the rights granted to the first movers have to be carefully weighed against the additional costs of delayed exploration by the followers. The leading AI labs, no matter how well resourced, can only pursue a select number of promising paths, and this also applies to AI safety.

AI has always been a fruit of rapid distillation: today’s marvelous models are as much a gift from decades of academic research on neural nets during “AI winter” as they are from Google’s publication of the transformer architecture. In that light, open-weights models acting as students of the closed models are a direct continuation of that lineage. Model outputs are a key research input for the ecosystem to advance.

The cleanest natural experiment on how restricting access to R&D inputs affects innovation comes from a different race: the one between the for-profit Celera and the publicly funded Human Genome Project to sequence our DNA. When Celera was first to sequence a gene, it restricted access through fees, limits on redistribution, and licensing requirements for commercial use. Heidi Williams found that these genes attracted 20 to 30 percent less follow-on research and product development than comparable genes that had been public from the outset. Although the restrictions were short-lived and disappeared within two years when the public project independently sequenced the same genes, the gap persisted. As late as 2009, Celera-sequenced genes still lagged behind the others.

When the value of follow-on work is high, even small initial frictions compound. Symmetrically, the removal of friction can have big consequences for both the rate and direction of innovation: when looking at biomaterials, Jeff Furman and Scott Stern found a 57 to 135% boost to cumulativeness when inputs were made more easily available. In the context of genetically engineered mice, Fiona Murray, Philippe Aghion, and co-authors studied what happened when the NIH negotiated away DuPont’s restrictions on hundreds of Cre-lox and Onco strains, ending the reach-through royalties and reporting that had limited academic access. Follow-on research rose. More tellingly, it fanned out: new researchers entered, and they explored more novel trajectories. Upstream, the creation of new engineered mice was unaffected.

Overall, whenever the benefits to cumulativeness and broad exploration are high, openness is the dominant strategy. This also means that society benefits the most from openness when uncertainty is still relatively high, as in AI today. When all that is left is execution along a known trajectory, then closed is far less harmful. And even in that narrow case (e.g., a molecule with well-understood clinical benefits that now needs to be commercialized), society only grants a temporary monopoly in exchange for relevant information. Patents themselves are instruments for diffusion and for avoiding trade secrets leading to no disclosure at all.

Now you may wonder if these examples from science-heavy domains also apply to AI. After all, serving models at scale requires massive investments in infrastructure, and that buildout is where most of the risk for the labs lies. But the lesson from telecommunications is exactly the same. In 1956, an antitrust settlement forced AT&T to license, royalty-free, one of the most valuable patent portfolios of all time. As measured by Martin Watzinger and coauthors, once AT&T’s inventions, which included the transistor, were suddenly available for others to build upon, inventive activity rose 17 percent in five years. Interestingly, the effect was not driven by other large firms free-riding on AT&T’s intellectual property, but by new firms going after completely different markets. Bell Labs also doubled down on its core business and was not deterred from innovating: the laser in 1957, the communications satellite in 1962, Unix in 1969.

While the idea of letting others compound on your ideas might be frustrating for Anthropic and OpenAI, the good news is that with a very limited number of exceptions, this is exactly how economic progress works. William Nordhaus found that innovators capture, on average, only ~2.2% of the social surplus they bring to the world. AI will be no exception. But this also places Anthropic’s ask to the US government in context: society does not owe the labs a stronger appropriability regime. As it turns out, there already is a different one in place.

The Innovation Last Mile

When a new general-purpose technology comes about, it initially has to be forcefully retrofitted into the existing system. These initial “point solutions” only realize part of the value of the new paradigm, and it is only when the architecture can be redesigned from first principles that the technology becomes truly transformative.

The rewiring requires control over and adaptation of key complementary assets, including infrastructure, trust, access to distribution, and relationships with regulators. This gives incumbents a second chance: while the new entrants may have caught them by surprise and out-innovated them up to this point, as the dependency on complementary assets becomes the limiting function, they may be able to imitate or acquire, catch-up, and preserve their position.

In the 1980s, Richard Levin and colleagues at Yale asked executives in innovative sectors across America a basic question: what protects your R&D bets? Patents ranked at the bottom. At the top? Learning, secrecy, lead time, and complementary assets. The study was repeated a decade later: same answer. The exceptions? Pharma and chemicals, sectors where a specific molecule can define an entire product class, and narrow exclusivity has teeth. But the vast majority of the economy runs on a limited ability to exclude others from the underlying ideas.

Subscribe now

Whenever appropriability is weak, there are still significant profits to be made; they just flow to whoever controls the complements to what is suddenly better, cheaper, faster. A classic example is the Beatles’ music label EMI, which doubled as an electronics firm and patented the very first CT scanner. Its top engineer, Godfrey Hounsfield, collected a Nobel for it. EMI captured none of the value. Within a few years, it was acquired, and the CT market was in the hands of GE and Siemens. GE’s complementary assets? Hospital distribution and services networks, the “last mile” for scanners to actually make it to market.

AI was born on the weak side of the appropriability scale. Despite fierce competition in the race to ASI, research insights regularly leave the confines of the labs, talent rotates between firms continuously as per Silicon Valley tradition, and the models’ APIs inevitably leak valuable information. Across all of these vectors, even draconian countermeasures would only buy the frontier labs a little extra lead time. Moreover, they would miserably backfire given the leverage top talent and large enterprise customers have in this phase.

Of course, not all of a lab’s secret recipe makes it into the public domain. According to Epoch AI, final training runs account only for 10% to 23% of the total compute costs, and the labs are accumulating meaningful tacit knowledge across data collection and curation, infrastructure design and optimizations, and learning from failed experiments and trajectories. But the lesson is the same: if the labs cannot rapidly get ahold of the key complementary assets needed to scale AI in the market, value will accrue elsewhere.

This explains why AI labs have been trying to increase customer lock-in by tightly coupling models and harnesses, and by expanding downstream into the application layer with tools such as Claude Cowork, Claude Tag, and ChatGPT computer use. Not only do these tools collect meaningful traces to replicate human skills, decisions and judgment, but they also amplify the value customers get from committing all their interactions and memories to the same family of products rather than multi-homing. If you can be the entry point for all of a customer’s AI needs, there are many ways you can evolve that experience to become stickier, even if the underlying intelligence is commoditized.

Beyond expansion at the application layer and vertical integration on compute to gain a cost advantage, regulation is probably the most effective way for frontier labs to force a complementary asset into the picture: if, in the name of safety, they can significantly raise the bar for model evaluation and approval, they can make it much harder for others that do not have access to the same well-staffed, all-star policy, regulatory, compliance and safety teams to compete.

It’s a lesson that has worked well in other heavily regulated industries, from financial services to healthcare. It is also what happened in digital platforms after the introduction of supposedly “pro-consumer” EU privacy rules: rather than limiting data collection by the largest players, GDPR had a chilling effect on everyone else that does not have the resources to comply with a labyrinth of rules.

But the possibility of regulatory capture does not settle the safety question. The labs’ claims could be self-serving and correct at the same time. Are open-weights models necessarily riskier than closed ones, or does concentrating control in a handful of players create an even greater systemic risk?

The Safety Last Mile

The United States has a comparative advantage in enabling permissionless experimentation rather than prematurely regulating new technologies or picking winners. But what worked well for the internet may be a terrible fit for AI: what if open weights are simply too dangerous, and strong guardrails and identity verification are the only safe way to harness frontier capabilities?

Stronger controls raise the bar for bad actors seeking to abuse the system. But because the technology is dual-use, those controls inevitably restrict access for a broader set of good actors as well. Depending on how important the rapid diffusion of defensive capabilities is within a particular domain, closed models may not always be safer.

The balance turns on two economic questions: who benefits most from diffusion at the margin, and where can access actually be constrained? In cyber, defense is distributed across a very large number of organizations and devices. Open weights lower the cost of auditing code, patching vulnerabilities, and upgrading legacy infrastructure, while sophisticated attackers may be capable of circumventing model-level controls. Restricting access would therefore tax a wide population of defenders without effectively excluding the most capable adversaries.

In biology, the calculus is likely different. A single dangerous capability may produce irreversible harm that society would struggle to defend against. Delaying diffusion therefore has particular value if the time is used to strengthen screening and control the scarce physical inputs (synthesizers, specialized equipment, and biological materials) needed to turn a model output into physical harm at scale. The lesson from complementary assets applies to security too: when the bottleneck is made of atoms, it offers a far more effective chokepoint against bad actors than software guardrails do.

For truly systemic and existential risks, the balance might tilt even further in favor of openness. When unknown unknowns dominate, the knowledge needed to detect or defuse an attack or major side effect of AI is likely to be distributed rather than concentrated among a small number of AI-lab employees. If so, open weights may foster a much more resilient and diverse AI immune system, one in which defensive capabilities are widely distributed rather than locked inside a few monocultures. Concentrating talent, ideas, and the ability to evaluate and test vulnerabilities within a small population would leave society with few options when problems arise.

Of course, this does not mean that highly capable open-weights models should be released without the same independent third-party testing that closed models should undergo. Staged releases would also allow the ecosystem to learn and adapt as new capabilities emerge. More broadly, safety researchers should invest more in understanding what can be done during the different phases of training to scale back or remove harmful capabilities, rather than relying solely on refusals.

Our current approach relies heavily on guardrails. But if the most realistic starting point is to assume breach, it is important to advance research on alternatives. The choice may not be between openness and safety, but between safety that relies on our ability to contain bits and safety that plans for containment’s failure.

That still leaves the business-strategy question: whatever the merits, can the AI labs capture safety regulation to slow open-weight progress, at home and abroad? Or are the economics of the technology pushing in a different direction?

Own the Weights of Production

A number of enterprises see potential regulatory capture by the leading AI labs as a threat to their business, and have started advocating for “sovereign” approaches to the technology, starting with open-weights not being placed at a regulatory disadvantage. The focus on open-weights is aimed both at reducing vendor lock-in, as enterprise deployments involve a non-trivial amount of custom work to squeeze performance out of a model, and at protecting a company’s most valuable information.

For models to tackle increasingly advanced specialized workflows, more fine-grained and proprietary data is needed. Today, that data is still in the hands of industry leaders, but across engineering, design, finance, accounting, and law, the AI labs are aggressively pursuing creative ways to collect more of it. Some of it is tacit knowledge encoded in the “weights” of the domain experts’ heads; some of it is in long-running systems of record; and some of it hasn’t been measured yet—although enterprises often have the pre-established relations with the right suppliers and clients to collect it.

As a result, while the first wave of conversations between labs and enterprises was centered around collaboration (our tech, your domain expertise), many customers are increasingly feeling that they are one step away from having to compete with their AI provider. Losing control over the “weights of production” is a real concern, especially because platform providers have strong incentives to renege on their promises of neutrality and favor their own products. A good example is the now tense relationship between Anthropic and Figma, where Figma started as a customer, but now has to compete to avoid disintermediation.

The initial enterprise interest in open-weights models was mostly cost-driven, and came as a natural correction to months of employees “tokenmaxxing” their way to greater productivity. By switching prompts to lower-cost, open-weights models, enterprises can shift a meaningful share of their token demand away from more expensive frontier models. AI labs have started fighting this by progressively slashing prices for entry-level models, and by training their models to perform best with their products.

But in the long term, cost is not the reason enterprises should care about openness: control is the critical dimension. Yes, labs promise not to train on inputs and outputs, but the reality is that interactions with their tools still leak valuable information. Furthermore, as tooling evolves and as companies invest significant resources in customizing their AI flows, vendor lock-in only increases. For most market leaders, open-weights are therefore not just a cost escape hatch, but existential. Over the next few years, every firm will need a way to train, optimize, and own the models that support its most strategic operations. If it cannot retain its most tacit knowledge and learnings, it will be disintermediated.

A glimpse of what the future may look like under a more sovereign construct can be seen in a recent collaboration between Thinking Machines and investment firm Bridgewater. For a hedge fund, the alpha is the business, so we would expect Bridgewater to be particularly cautious in selecting an AI architecture. In this case, the two entities collaborated to train a custom model as an extension of Thinking Machines’ base one. The lab brings generalized machine intelligence and AI training tooling, the hedge fund the proprietary insights. This specialization of labor protects Bridgewater and gives it more control over the weights that capture its core intellectual property and tacit knowledge.

From a strategy perspective, while Anthropic and OpenAI are sometimes seen as potential competitors by their own clients, Thinking Machines is taking a value-chain-enhancing approach, in which it only wins if its customers do too. Microsoft and Palantir have made similar bets to become infrastructure and AI tool providers for enterprises that want to keep their machine intelligence within their own perimeter, and many more will follow.

Which approach is a better fit for the market for intelligence? The one that starts with general intelligence built by a few top AI labs and works downstream, or the one that starts bottom-up from the application layer and feeds ideas and improvements back to a multitude of open-weights models? It depends on the shape of the demand for intelligence.

Mapping the Curvature of Intelligence

Any general-purpose technology brings benefits that are so widespread across sectors that it is impossible for any single firm to appropriate them. In theory, this could lead to underinvestment. The more general the AI models of Anthropic and OpenAI, the greater the gap between value creation and capture, which may explain the more recent focus on specialized cyber and life-sciences models. One may worry that without the labs, a number of areas of machine intelligence that are less monetizable would be neglected, as everyone has an incentive to free-ride on base capabilities and build where there’s high willingness to pay.

But this view misses a key consequence of open weights commoditizing yesterday’s frontier capabilities: as costs collapse, a number of applications that would otherwise not be economical, suddenly become viable. This accelerates adoption, and with adoption the number of companies that have a reason to support shared technology inputs also increases. We’ve seen this play out before with open source software like Linux, where companies with complementary business models contribute back to the commons, and it likely explains why NVIDIA was able to bring together a large and diverse coalition in support of open-weights.

This leads to a particular way the market could split. The relevant curve is the payoff to an additional unit of model capability. Across much of the economy, that payoff will diminish quickly: once a model is good enough, further gains matter less than price, reliability, and ease of deployment. Intelligence will behave like a commodity and be priced “at the meter,” like the electrons in our electricity bills. Open-weight models, or closed models run at cost, will serve this part of the market, and firms with advantages in infrastructure or compute will compete on cost. Some labs may even bundle or cross-subsidize commodity tokens to retain customers for higher-margin ones.

But producing value-added tokens will be different. As companies turn into token factories that turn input tokens into more refined versions of them, they will only be able to charge a mark-up if they have access to better ground truth, data, and talent. Those ingredients are needed to expand what is safely automatable and deliver machine intelligence that is more attuned to current conditions. Ultimately, this is exactly where complementary assets will play a key role, and some of the most valuable models will be advanced by the companies that are on the front lines of converting friction with the real world into better tokens. Agentic work only adds value when it can be trusted, and trust requires better verification infrastructure. The only path to more efficient verification is having better ground truth than your competitors. Distribution is, of course, one way to collect that faster than others.

Other domains have the opposite curvature. In cyber, finance, frontier R&D (including AI development) and patentable matter, a small capability advantage can produce disproportionate returns. These convex payoffs can sustain temporary frontier rents because buyers will pay heavily for the best available model rather than a second-best one. Whether those rents accrue to generalist labs or specialized providers depends on who reaches the right data first.

But the highest-payoff domains have a property that many who are long frontier labs miss. In the most convex domains, the marginal buyer is not a user or a company, it’s a country. And when it truly matters, countries don’t need to use the market. We’ve seen a preview of this with the Fable and Mythos bans. As the United States and China race across AI, robotics, quantum, and space, they will do whatever it takes to be ahead. So the steeper the curve, the higher the likelihood that if a lab were to achieve something truly unprecedented, it would be tightly managed or even nationalized. Ironically, for any AI lab to retain its freedom, it needs to be successful, but not too successful.

Which leads to a natural question about AGI/ASI. While many would argue that we already have AGI-like superintelligence across any domain that is verifiable, it is fair to wonder if everything discussed so far does not apply once a lab reaches ASI. If anything, ASI is exactly the reason why the labs are chasing recursive self-improvement over profitable industry verticals. The argument goes that if you solve the intelligence piece first, the rest of the economy will bend the knee to code. But if you believe that distributed, tacit knowledge is still relevant for ASI to be productive, then you quickly realize that the value of complementary assets does not disappear with ASI. If anything, they might become more of a last-mile bottleneck to deployment. Last, if ASI is made of ideas, containing it forever will be impossible. It will leak, the curve will be commoditized, and value will again accrue where there’s remaining friction.

Generalists versus Specialists, and Unknown Unknowns

By now, it is hopefully clear that it is not the level of investment in AI that is challenged by open-weights, only its direction.

A world where open weights dominate sustains progress in every domain where there is at least one company with a complementary business model or assets to guarantee appropriability. Some companies, like NVIDIA, naturally care about any domain that leads to more demand for intelligence. Others care because of how AI lowers costs, improves quality, increases engagement and retention, and so on. An open-model-centric outcome is one where machine intelligence is still very capable, but specialized. In this scenario, depending on how SOTA is achieved and renewed in each domain, generalist AI labs may face the same fate as EMI: it invented the CT scanner, but was unable to monetize it.

On the opposite side of the spectrum, a world where all the progress stays within a few large labs is one where society would benefit from the most general form of intelligence possible. The condition for this to be sustainable is that the labs gain enough of an advantage and lead time at each round to recover their R&D costs and fight commoditization. It is a world where the labs have significant market power, but open-weights keep them in check by making their position at least partially contested.

Depending on what the shape of the demand for machine intelligence actually looks like, both scenarios may continue to co-exist for a while: the closed labs corner the top of the market, the open-weights models serve the bulk of the tokens, and the process repeats with each generation. But if AGI/ASI becomes too cheap to meter, then the last surviving moat will be information that has not been discovered yet, and everything that has not been measured. The race would then evolve into a contest to see who can find friction with the universe and collect that signal first to feed it back to the superintelligence. And that’s a game where many distributed forms of intelligence may have an unfair advantage over a centralized one.

Friedrich Hayek’s insight into the fatal weakness of central planning was not that planners lacked intelligence or compute. It was that no planner, however intelligent, could possess all the relevant knowledge. That knowledge is dispersed, local, often tacit, and continuously changing: “knowledge of the particular circumstances of time and place.” Even a superintelligence would have to venture continually into the world to discover it.

Markets outperform central planning because they transform countless independent encounters with reality into a distributed process of discovery, adaptation, and selection. Many competing superintelligences, each probing a different frontier of the universe with different assumptions and values, will discover more than any single intelligence, however vast. Open weights extend the same logic: they allow intelligence to be adapted everywhere rather than optimized once and rationed from the center. Regulatory capture may slow their diffusion. It can weaken the gradient. It cannot reverse it.

Subscribe now

This newsletter is provided for informational purposes only, and should not be relied upon as legal, business, investment, or tax advice. Furthermore, this content is not investment advice, nor is it intended for use by any investors or prospective investors in any a16z funds. This newsletter may link to other websites or contain other information obtained from third-party sources - a16z has not independently verified nor makes any representations about the current or enduring accuracy of such information. If this content includes third-party advertisements, a16z has not reviewed such advertisements and does not endorse any advertising content or related companies contained therein. Any investments or portfolio companies mentioned, referred to, or described are not representative of all investments in vehicles managed by a16z; visit https://a16z.com/investment-list/ for a full list of investments. Other important information can be found at a16z.com/disclosures. You’re receiving this newsletter since you opted in earlier; if you would like to opt out of future newsletters you may unsubscribe immediately.

America | Tech | Opinion | Culture | Charts

It seems obvious to say, but if you leave Silicon Valley and go out into the rest of the world and tell them, “there are these things called agents, which are pretty smart, and can do tasks with you, and automate some of the repetitive parts of your work”, chances are the first question you’ll get back is, “Can they use a computer?”

This is a good question! Can they, really? The long horizon of productivity potential, out in the real economy we’re going to go unlock over decades, runs through pretty everyday work: can an agent sit (metaphorically) at a desk 24/7, and be trusted to use a web browser, fill out forms, click the right buttons, and not make mistakes? This is the domain of Business Process Outsourcing (BPO), which historically meant, “can this work be outsourced?” but now has a new agentic frontier. We wrote about this last year, when the computer-use landscape was still mostly a bunch of demos. A lot has happened since then.

The models have improved faster than almost anyone expected. Computer-using agents are beginning to hold up in production at scale and on narrow, repeatable workflows: updating systems of record, moving data through portals, processing tickets, checking records, and handling the long tail of software where no clean API exists. With the right infrastructure, computer-use capabilities can now be deployed to tackle end-to-end tasks at scale, which before required either human supervision or direct human completion.

Today, workflows leveraging computer-use are far from perfect: agents are brittle when work drifts off the runbook, and for certain use-cases where caching is intractable (more below) they are expensive enough that the math does not work everywhere. But we’re seeing production deployments for standardized back-office work, especially where labor would otherwise be clicking through legacy systems by hand; the cost curve is starting to look compelling, considering that workflows leveraging computer-use offer structural advantages such as 24/7 availability and - most importantly - scalability to meet demand.

The first wave of computer-use infrastructure was about making agents capable: seeing, clicking, typing, recovering from mistakes. The next wave is about making them useful inside actual companies. As raw UI navigation becomes a model-layer commodity, the model is no longer the main bottleneck and the durable advantage moves up the stack: context, permissions, process knowledge, validation, escalation, error handling, caching, and the hard-earned understanding of how work actually gets done inside one specific customer’s organization to map a workflow end-to-end. In other words, the frontier is shifting from “can the agent use a computer?” to “can it reliably do this job?”

From Humans Watching Every Step to Real Autonomous Workflows

A year ago the best computer-using model scored 42% on OSWorld-Verified; today’s best scores 85%, above the ~72% humans manage on the same tasks (this means they successfully completed 85 of 100 tasks). In production these general frontier models run much like they do in the benchmark: the labs expose computer use as an API - the model gets a screenshot, returns clicks and keystrokes, with OpenAI’s CUA also layering in accessibility-tree or DOM data where available - and builders wrap that loop in their own harness: a sandboxed VM or browser, plus the orchestration, verification, and retry logic around it. Notably, almost nobody deploys consumer products (Claude, ChatGPT agent mode) for this - founders and enterprises build on the raw APIs, or buy from vendors who package them. And the capability jump is what made those setups viable - “the models weren’t good enough to use in production on their own until Opus 4.6 in February 2026,” as one founder building in the space put it. Somewhere in the last eighteen months, computer use capabilities crossed from demo to being deployable in the field.

Of course, benchmarks aren’t always the best proxy for the viability of a real-world deployment. OSWorld counts completed tasks, so 85% still means 15 of 100 failed, and a business process only finishes if every step does. Back-office work doesn’t grade on a curve: if a person reviews every output, no labor was saved. (It’s analogous to what’s happening in coding right now: the scarce resource is no longer writing the code, it’s vouching for it.)

We found that the best way to think about what matters is to go beyond the benchmark and focus on the core question: can a business process be reliably automated with computer-use capabilities? Under this lens, what makes the biggest difference is everything around the model - that is: verification, escalation, error handling when a retailer portal changes its layout overnight.

Perhaps the clearest tell is that one operator we spoke with, who runs millions of automated tasks a month, couldn’t tell us which model executes them; he hadn’t needed to find out. His vendor swaps models underneath him the way a cloud provider swaps hardware. But he did trust the computer-using agent to run these tasks. Bottom line: when your heaviest users stop checking the leaderboard, the leaderboard has stopped being the story.

Thus, the chart above explains why production deployments exist in 2026 and didn’t in 2024. From there on, what determines whether they work is everything else - and that’s the rest of this piece.

Agents are Protocol Following

We had various conversations with teams running workflows leveraging computer-use capabilities in production, and we learned from their experiences that protocol following tasks work best. Unsurprisingly, computer-using agents break on more complex workflows where accuracy is harder to verify. The overall takeaway is that computer-using agents are strongest on standardized, repeatable tasks with a clear, well-defined path. The real unlock is the long tail of software where no clean API exists and a person would otherwise be clicking through a UI by hand. In practice, the work looks like updating records in a CRM, QA, logging into government and insurance portals, pulling data off databases and regulatory pages, retail order processing, contract processing, or IT tickets in ServiceNow.

We believe the voice of the user here tells the story much better than any theory. Some examples: a CPG data platform walked us through how they run ~15-20M automated portal interactions a month, using agents as a self-healing fallback for hand-coded scrapers - when a retailer portal changes its UI, the agent diagnoses the break, fixes the automation, and keeps data flowing before an engineer ever sees the error. Once implemented, they told us, they cut the engineering team dedicated to scraper maintenance in half and re-allocated staff capacity to other workflows. In another case, from a global systems integrator, we learned they have 27 live workflows leveraging computer use agents that process ~1,500-2,100 IT tickets a day, with the ultimate goal of redeploying 20-25% of headcount on low-margin managed-services contracts. And finally an agency walked us through how they automated a recruiting workflow end-to-end to populate data in an applicant tracking platform as soon as a candidate interview was over. To do so they run a cheap non-frontier model because it “does everything we need and does it well.”

The clearest pattern is workflows where in theory a computer-use agent could operate and solve the task, but there is either no clear answer for “what good looks like” (i.e. they are hard to evaluate) or no reliable way of determining whether a task succeeded. Usually, issues arise fast when: (1) you can’t cross-check the output - think of an agent extracting payment terms from contracts into an ERP: if it reads “net 60” as “net 30,” the record looks perfectly plausible, passes every visual check, and nobody catches it until an invoice goes out wrong; and (2) in some cases there is no signal to verify success at the time the task runs - think of an agent submitting a claim on an insurance portal: the submission goes through, the screen says “received,” task done. Except two days later an adjuster calls the office because a policy number needs confirming before the claim can be processed. A human who filed that claim picks up the phone and sorts it out in thirty seconds; the agent has no idea the call ever happened, and the claim quietly stalls. Net-net is that a smarter model doesn’t fix a process whose ground truth shows up as a phone call to somebody’s desk a week later, unless the harness is designed to handle the edge case from the get-go.

Buyers Care About Infrastructure

For the buyers we spoke with, the model itself is rarely the deciding factor, as “the models today are already good enough.” In practice, they evaluate and pay for everything around the model: the infrastructure to run reliably at scale, pass security review, and prove ROI. Users don’t care whether the solution uses a given frontier model; rather they focus on whether it can actually get the task done at scale and reliably. Period.

As a consequence, failure modes matter more than any benchmark, and design for failure needs to be a first-class concern from the start because a solution that does not handle failures well will never be adopted in production. One example of what this looks like in practice and a pattern we ran into more than once: the agent runs the workflow once, the system caches it as deterministic code, runs execute as cheap repeatable code from then on, and the model comes back only when something breaks - to diagnose, fix, and re-cache. With this approach, cost per run falls over a workflow’s lifetime, and cheaper models just lower the bill. What’s interesting about it is how it handles uncertainty. Where before deterministic code simply failed, or a human had to review and fix every break, here the agent absorbs that uncertainty on its own. It’s one way of designing for failure, and it shows what buyers are actually rewarding and using at scale.

We did not encounter more sophisticated use cases among the users we spoke with, which tells us the market is still chipping away at the low-hanging fruit. That said, there is a long list of workflows that can be automated this way before anyone needs the harder tasks.

The Model Is Not the Differentiator, Context Is

For founders, the more important shift is what is becoming commoditized. Building a computer-using agent used to mean wrestling with Selenium or Playwright, or more recently Stagehand, and stitching together DOM or video recordings to capture a workflow. That whole execution layer is getting abstracted away, the same way Claude Code abstracted the scaffolding around coding agents. If clicking the right button is no longer the hard part, it is no longer the moat.

Unsurprisingly, the context and knowledge of the workflow are durable. The hard part is not whether an agent can navigate an SAP screen, it is whether it understands how a particular company actually gets work done: the tribal knowledge, the internal terminology, the preferred formats, who to escalate to and when, how to handle failures, how to verify output reliably. In practice that context lives in runbooks, in access and credentials, in test cases and guardrails for when a workflow goes off-script, and increasingly in a single recorded video of someone doing the job once. None of it is general. All of it is specific to a company, and often to one team. That said, it is exactly the kind of specific, unglamorous problem that focused startups tend to solve better than model providers, which is why we think the next generation of agentic coworkers gets built at the application and context layer, not the model layer.

The buyers we spoke with squarely confirmed this. They picked vendors on whether the product reported hours saved without extra work, and whether a junior engineer could run it. For now, the moat is not the frontier capability—rather, it’s being the vendor an enterprise is allowed, and able, to use at scale in production.

The Real Inflection is Economic, Not Just Technical

The cost data is also encouraging. Take the numbers above as an order of magnitude, not a precise quote. Running an agent costs roughly $6-8 per hour of inference today, but in practice anywhere between $3 and $15 depending on how the harness is built - how often it screenshots, how much context it carries, how much of the work it can hand off to deterministic code. These figures describe the agent operating the UI screenshot by screenshot with a frontier model - the most expensive mode there is. Well-built harnesses reserve that mode for what actually needs it, and let cheap deterministic code handle the repeatable parts - not every workflow can be optimized this way, but where it can, blended cost drops fast. So read the comparison as the worst case, and even then an agent is roughly break-even against offshore BPO at ~$10/hour fully loaded, and pencils out to a 70-80% gross margin against US back-office labor at ~$30-45/hour. In production, the harness drives real-world cost as much as the model does.

Same caveat on speed. In agentic mode, agents are still slower than people, and it is not close - a task someone finishes in two to three minutes can take an agent eight to ten, and academic benchmarks put the gap even wider. Deterministic runs flip this: code executes faster than any human - but for the agentic work the argument is not speed. It is that an agent runs around the clock, costs a fraction of US labor, and scales without hiring.

This comparison works for BPO buyers and ops teams, but the unit economics look different if you are the one selling agent-hours, because costs are less predictable in the wild. COGS are inference plus retries (i.e. failed runs still burn tokens) and margins compress when context grows or screenshot frequency goes up. Vendors manage this by pricing per task, per hour, or per outcome, each with a different risk profile depending on workflow variance. There are also the practical realities of things like monitoring, maintenance and human escalation, which are being priced in, just as they would be with human workforce. There is no one-size-fits-all business model here yet, and the answer varies by vertical.

And the math only gets better - inference keeps getting cheaper, and open-source models are getting good enough for a growing share of these workflows. For any task an agent can reliably solve, embedding computer-use will likely be way more convenient than human labor. So the real question is no longer whether the economics work - it’s how far the set of tasks that can be solved reliably extends, which is where things are heading next.

Where Do We Go From Here?

Over the past year, labs and a wave of startups have poured hundreds of millions into computer-use RL environments - the sandboxes where a model practices real tasks and gets rewarded for finishing them - with companies like Mechanize, Habitat, Fleet, Chakra, Deeptune, Matrices, and Originator building the training and eval substrate underneath the frontier models. That spend is what shows up as better reasoning, better state tracking, and more tolerance for apps that misbehave. The models still need careful harnessing to hold up in production - run-caching being the clearest example - but the raw capability was bought, deliberately, through this training infrastructure and will only get better over time.

Architecture, though, is a different story. Most computer-use systems deployed in production today are single-agent: one model, one task, one session. As workflows get more complex and latency becomes a constraint, multi-agent architectures start to matter. For example, a planner decomposes the workflow, executor agents handle subtasks in parallel and long-running agents bring their own problems: memory, trust, and failure rates that compound over time. The teams doing interesting work here are all building bespoke orchestration, because no standard framework exists yet. The Claude Code analogy is instructive: when coding agents matured, a scaffolding layer emerged to abstract the orchestration away. The same is likely to happen for workflows leveraging computer-use capabilities, and that abstraction layer is one of the more interesting unsolved infrastructure problems in the space.

From here, future developments run along three lines: accuracy, latency, and cost. Accuracy is most important, and, as we explained above, represents the difference between a cool demo and actually solving the problem - catching anomalies, checking its own work, escalating only when it actually needs to. Latency is the one most likely to surprise people: some teams already cut it today by grounding on the accessibility tree instead of screenshots. Standard Intelligence’s general computer action model is trained on an 11-million-hour video dataset, running at 30 FPS, and is an early signal that the step-by-step screenshot loop slowing today’s agents is a solvable problem, not a permanent tax. Cost keeps falling as inference gets cheaper, and smaller non-frontier models take over the routine clicks. These three vectors in addition to security and governance (e.g. credentials, audit logs, data retention, prompt injection, and accountability, permissioning).

Enterprises can and are benefiting from computer-using agents for narrow workflows that have high volume, repetitive steps, stable business rules with legacy interfaces or missing APIs. For now, they are best suited for tasks with immediate, machine-observable evidence of success, tolerable failure consequences, and clear escalation routes. But with the above developments, improvements are real and rapid, making computer-using agents more viable for more types of work.

Let’s just say - the future for computer-use capabilities is bright!

Subscribe now

This newsletter is provided for informational purposes only, and should not be relied upon as legal, business, investment, or tax advice. Furthermore, this content is not investment advice, nor is it intended for use by any investors or prospective investors in any a16z funds. This newsletter may link to other websites or contain other information obtained from third-party sources - a16z has not independently verified nor makes any representations about the current or enduring accuracy of such information. If this content includes third-party advertisements, a16z has not reviewed such advertisements and does not endorse any advertising content or related companies contained therein. Any investments or portfolio companies mentioned, referred to, or described are not representative of all investments in vehicles managed by a16z; visit https://a16z.com/investment-list/ for a full list of investments. Other important information can be found at a16z.com/disclosures. You’re receiving this newsletter since you opted in earlier; if you would like to opt out of future newsletters you may unsubscribe immediately.

America | Tech | Opinion | Culture | Charts

Bookslop

We have some more data to consider with respect to AI’s impact on publishing—this time a paper from researchers at Stony Brook, Columbia, Michigan and the MIT Initiative on the Digital Economy, called “Generative AI floods and dilutes the market for books.”

It’s not the friendliest title, but the gist of it is pretty straightforward: AI-generated books (either substantially or somewhat) are proliferating in the self-published categories . . . and people seem to like them? Well, whether people like the book slop or not, they’re definitely buying it:

Books with at least some detectable AI text are now capturing ~40% of observed sales.

That’s a fairly striking figure, but it’s even more striking when you consider just how easy it is to self-publish an AI book, and how many AI-ish books have come to market.

Since 2023, self-published new books have proliferated far more quickly than either of the number of books that sell, or book sales and revenue:

While the total catalogue has increased nearly 40x, actual revenues have only increased ~9x. And while non-AI books still earn the most per book, the gap has closed substantially.

In other words, AI-authorship is taking a volume approach: lower-cost output has scaled the self-published catalogue massively, but since the hit-rate is lower, revenue has grown, but not by nearly as much. Hence the paper’s title: Gen AI floods and dilutes the market for books.

Still, with enough shots on goal, eventually bookslop finds the right combination of m-dashes and ‘not this it’s that’ to attract paying customers—sometimes, a lot of paying customers (relative to the field). Again, the barriers to entry are low, which means the entries are high, which means that AI-generated/assisted books can be a chart-topper in basically any genre:

Books with substantial AI text have at least a 10% share of the Top 25 across all genres, including ~40% of Sci-Fi/Adventure/Dystopian category.

Flood the zone, win the zone. Sounds bad.

Now, let’s take a step back for moment, though. Assuming that AI text detection actually works, the point that the authors of the study are trying to make is that AI-Books are flooding the market, winning share with volume over quality. Revenue per book is falling basically across the board, and real human writers are getting squeezed—rather than a passing fancy, “low-quality, cheap books can still reshape a market,” but really the authors are saying ‘…will reshape a market.’

Maybe so, but there’s a pretty obvious silver lining here. One can’t help but notice that self-published revenues by the author’s own calculations still increased nearly 9x in just three years. That’s perhaps a big increase off a small base, but that’s still a very big increase. If bookslop is, in fact, flooding the market, but it’s also growing the market much more quickly than it otherwise would (to the benefit of human authors), is it necessarily a bad thing that it’s capturing a larger share of revenue?

The Defense Industrial Backlog

Everyone knows about the cloud revenue backlog, which is generally considered a strong indicator of AI-driven demand:

If Hyperscalers are spending more on compute, then they better be seeing more demand for compute, and so they are.

Well, the cloud backlog is cool and all, but have you heard of the defense contractor backlog? That one is pretty impressive too:

Revenue backlogs for some of the largest defense contractors have inflected upwards since the beginning of last year (although, to be fair, General Dynamics and RTX include some commercial orders, as well).

It’s not quite the same parabolic curve as demand for compute, but the pull on the defense industry is definitely real. Outside of the specific defense earnings releases, the surge of defense-demand and output is showing up in the macro-picture, as well.

According to Census data, both defense orders and shipments have increased ~60% since 2020, with the lion’s share of that increase occurring over the past 1.5 years:

All categories of defense production have ramped up, but the biggest ramp is in the “other” category, i.e. defense orders that aren’t aircraft, search, and comms (represented by the top three layers in the stacks above). While Census doesn’t offer a more granular breakdown of “Other,” a different, more detailed series from the BEA leaves some clues. If you apply the BEA’s category mix to the Census figures, then a rough estimate suggests that ships, missiles, and small arms/ordnance are driving most of the increase in demand for defense-related production.

Another interesting thing is that all this fresh defense spending has unsurprisingly led to a surge in production—yes, orders have increased, but throughput has increased even more:

“Unfilled defense orders” have jumped by ~$20B, even as the estimated backlog has dropped by ~20%. This backlog is really an imperfect proxy because it’s just the quotient of the dollar value of defense orders divided by the dollar value of monthly defense shipments, and it doesn’t take into account whether it’s a mortar round or jet engine. In other words, one shouldn’t conclude that the backlog or lead time for all types of defense orders has compressed, but the aggregate picture appears that it has.

But, the implication is that even as demand has ramped up substantially, the sheer volume of monthly defense shipments has accelerated more so.

That follows various government reports highlighting the need for much higher defense industrial capacity (and lengthening lead-times for key defense orders), as well as investment and progress towards that end:

$5B of investment in capacity has substantially increased munitions capacity since 2022, although the DOW acknowledges that target capacity is higher still. It turns out, you can, in fact, just build things, even things that have historically been built at a slow n’ steady pace.

Bigger picture, you can add defense manufacturing to the growing pile of evidence that an industrial renaissance is most-definitely underway (even if, again, there’s still a ways to go).

That’s seven straight months of expansion for the manufacturing PMI, where every indicator is on the rise (except for prices paid).

In terms of what’s happening, part of it is likely the ongoing conflicts with Iran and Russia, but bigger structural changes are also afoot. Outside of the US, European countries are allocating more to their defense budgets, and America’s role as a net-exporter of defense goods has only grown:

Perhaps more importantly than just increased defense budgets, key battlefield components are evolving, particularly with a shift towards autonomy, and that has created urgency to evolve the industrial base, as well (with investment to match). The point being that the defense industry is undergoing more than just a seasonal glow-up and/or resupply. By all accounts, this is a secular demand impulse that isn’t going anywhere anytime soon.

IT Services, Down (But Not Out?)

Charts has mused on the ever-shifting landscape of perceived AI-winners and losers. We’ve flagged Accenture, the IT services giant, as an example of a company that was at one point considered a likely beneficiary of AI implementation efforts, but has more recently become a likely loser of shifting IT budgets away from IT services, and towards AI.

Unsurprisingly, Accenture is not alone amongst its peers, as IT services more generally has experienced a dramatic repricing of forward earnings:

Consensus multiples for IT services are now about a third of what they were in the beginning of 2025.

That’s interesting so far as it goes, but the more interesting question is less whether ‘has AI has doomed IT Services?’ so much as ‘well, what is IT Services gonna do about it?’

No one knows the future, of course, but this seems like a perfect example of rather than AI eliminating a function, AI is forcing it to evolve. Consulting firms are going to have to invest the time and effort to adapt and improve their value proposition (and perhaps even their business models) in response to this brave new world. Some part of that is “reskilling,” an effort that is likely already reflected in the broader premium for more senior talent (possessing the required skills to be AI-useful right now). Reskilling tends to take some time, and it wouldn’t make sense to hire a brand new cohort of young consultants, before you know what they’re being hired to do (and how they’re going to be trained to do it).

On the bright side, this sort of bust-to-boom transition has happened to IT services before. The shift to cloud also put a major damper on IT services (as they were), and also forced the industry to adapt and evolve—which it did:

It took 3-4 years from the collapse of cloud-driven growth expectations for IT services to recover, but eventually the threat became an opportunity.

On-prem services were disrupted by the shift to off-prem—and the pain for consultants was real—until they adapted their skills and offerings to become the experts in lift n’ shift (and everything else that entailed). Once again, the good times rolled for IT Services. Not for everyone, of course. The consultants that didn’t or couldn’t adapt continued to languish, but for the ones who did, tech disruption became a massive accelerant. Change was not painless, but it was net-good.

There’s no reason to think that AI will have a dissimilar effect. Roles will evolve, and some will become obsolete, but for others still, AI will be a force-multiplier.

Kimi’s Got a Career Update

It feels like ages, but it was really only two weeks ago that Moonshot released its Kimi K3 model, leading to a total timeline takeover for the open v. closed debate.

Obviously all that attention meant that lots of people wanted to know what all the fuss was about, and so they decided to see for themselves:

According to data from SensorTower, Kimi app downloads and daily active users both surged practically overnight. Downloads nearly quintupled, and DAUs jumped by ~40%. For now, at least, K3’s coming out party, is more than just hype—people are definitely giving the model a spin.

As an aside, it does seem like the vast majority of Kimi’s impact happened outside the US:

SensorTower’s estimates of in-app weekly purchasing volume definitely increased in the US (off a relatively small baseline), but the global leap was far more pronounced. Kimi’s global revenue reach has apparently been increasing since the beginning of the year—already 6-7 times higher than in the US—but gross in-app purchase revenue went vertical in July, and basically doubled in just a week. In other words, for all the hype, Kimi is still much more of a global (and Chinese) phenomenon than a US one.

Interestingly, SensorTower also provides some evidence that Moonshot is trying to change that, however—not just by shifting its geographic focus, but potentially by making a move on a different vertical:

Moonshot’s 2026 marketing strategy looks nothing like its 2025 marketing strategy. The Chinese lab went from spending big money on YouTube to directing more than 60% of its advertising fire on LinkedIn (while YT has become an afterthought).

Now, putting advertising dollars to work on LinkedIn doesn’t necessarily mean that Moonshot is making an enterprise play, but it might mean that. Nor is Moonshot the only one who advertises on LinkedIn—data from earlier this year suggested that LinkedIn was the favored advertising channel for AI apps, in general.

In all events, LinkedIn is now apparently the place to get your AI slop from all ends…AI-generated career poasting with a steady supply of AI ads to help you pick your weapon of choice.

Subscribe now

This newsletter is provided for informational purposes only, and should not be relied upon as legal, business, investment, or tax advice. Furthermore, this content is not investment advice, nor is it intended for use by any investors or prospective investors in any a16z funds. This newsletter may link to other websites or contain other information obtained from third-party sources - a16z has not independently verified nor makes any representations about the current or enduring accuracy of such information. If this content includes third-party advertisements, a16z has not reviewed such advertisements and does not endorse any advertising content or related companies contained therein. Any investments or portfolio companies mentioned, referred to, or described are not representative of all investments in vehicles managed by a16z; visit https://a16z.com/investment-list/ for a full list of investments. Other important information can be found at a16z.com/disclosures. You’re receiving this newsletter since you opted in earlier; if you would like to opt out of future newsletters you may unsubscribe immediately.

America | Tech | Opinion | Culture | Charts

How can an AI model know when its work is done?

Well, how does a human know when our work is done.

A programmer waits for the tests to turn green or waits for PR review from their team. A designer adjusts a composition, steps away, returns, and decides the remaining imperfections no longer matter. A writer submits a draft because the deadline has arrived or because an editor accepts it, not because the prose has reached some objectively final state.

“Done” is rarely a property of the work itself. It is a judgment produced by the system around the work. Humans do not possess a universal detector for “done”. We rely on a patchwork of signals like tests, specifications, precedent, approval, deadlines, risk, and finding that point of diminishing returns. In each case, completion comes from outside the work itself.

The Model that Could Continue Forever

An AI model can almost always produce another answer.

It can revise the paragraph again. It can try another implementation. It can generate another image with more detail, different lighting, and a stronger composition. It does not become tired of the work. It does not notice, unless we give it some way to notice, that the last three revisions made the result different but not necessarily better.

This is part of what makes the recent idea of loop engineering so compelling. Instead of a human prompting a model, inspecting the result, describing what went wrong, and prompting it again, we can ask the system to perform the whole cycle itself. The person no longer has to sit inside every turn. The agent discovers the work, gives it to the model, checks the result, and decides what should happen next.

However, the nuance when writing a loop is that the loop is only as good as the verifier at each step. Even before we started talking about loop engineering, everything already runs as a loop, just with a very expensive tool call – human hand prompting and serving as the verifier. When taking humans out of the loop, designing what should be verified in each step becomes the key in advancing the loop state, and the reality is it’s hard to make them work.

Take the standard coding-agent loop as an example: keep working until the tests pass. It sounds almost perfectly verifiable. But the tests are only a proxy for the task. In SpecBench, frontier agents routinely passed the visible tests while failing held-out tests that exercised the same features together. One agent produced a 2,900-line “compiler” that simply memorized the test inputs. The loop converged, but just on the verifier, not the user’s intent.

The verifier is not just the stop condition. It also defines what the loop treats as progress. If the signal is incomplete, the loop can get better at passing the check without getting better at the task.

Loop engineering is not the practice of making an agent retry. It is the practice of making each cycle reduce the distance between the current state and a desired state. A loop is not yet a direction.

The Loops that Converged

The first loops that worked well were coding loops. This is not an accident. Code is both editable and executable. An agent can change one function, run the program, read the test failure, and try again. The environment returns a relatively clear signal about what broke. The loop has both a precise way to act and a verifier that can measure progress.

I wrote about a similar pattern in visual code generation. An SVG is not just an image; it contains paths, shapes, text, gradients, and layout. A Blender scene is not just a render; it contains geometry, materials, cameras, joints, and constraints. These representations give an agent something it can inspect and edit locally. If one curve is wrong, change the path. If one object is misplaced, move that object. The artifact can improve across iterations instead of being regenerated from scratch.

But editability is only half of the problem. In open-ended image generation, another iteration often means generating another sample and choosing the best one. The feedback is global, and it is hard to map “this looks worse” to one precise edit. SVG and Blender loops can converge when the target can be expressed as a reference, geometry, constraints, or functional behavior from an articulated object. They struggle when the target is simply “make it better, with better taste, but you cannot ask a human”. Visual loops are not impossible. They are often extremely hard to verify.

The Conditions of Completion

If the verifier gives the loop direction, what does the loop need to converge?

Based on many conversations with engineers and researchers across several domains, I think there are four things.

1. A target state

The system needs a representation of what “done” means. For code, this might be a test suite, a specification, or a set of performance constraints. For an SVG, it might be a reference image, dimensions, colors, and layout rules. “Make it better” is not a target state. It is another prompt.

2. An observable current state

The system needs to inspect what exists now. And that could mean files, diffs, test results, traces, a DOM tree, an SVG structure, or a Blender scene graph. A rendered output alone is often not enough. The system also needs to see the underlying structure so it can identify where the error came from.

3. A precise way to make changes

The agent needs to change the part responsible for the error without regenerating everything else. Changing one function is better than rewriting the repository. Editing one SVG path is better than generating a new image. Adjusting one object in a Blender scene is better than rebuilding the scene from scratch. The more local the edit, the more likely the loop is to preserve what already works.

In practice, this is the part people struggle to get right. Nearly every researcher I talked to said the same thing: their loop started working when they found the right set of tool calls and intermediate prompts. How do you discover the tools that meaningfully advance the loop? Right now, no one knows in advance. It is mostly trial and error.

Which points to an uncomfortable implication: a loop is tuned to its stack. The tool calls that made a loop converge on one codebase encode assumptions about that codebase, and those assumptions stop holding somewhere else. A loop that worked for someone else is a starting point, not a guarantee. Bespoke loops do not generalize for free. And this is why we see both sides of the discussion: some people found magical loops that worked for them, but others found when they use publicly published loops they do not work at all.

4. A stopping rule

The system needs a condition that tells it to stop. The condition should come from outside the generator: tests passing, constraints being satisfied, a score crossing a threshold, or a reviewer approving the result. The stop condition also needs to account for cost – a loop that reaches the right answer after 500 attempts may converge technically but not economically.

One useful way to think about this is across two axes: how editable the artifact is, and how verifiable the result is.

Code often sits in the upper-right corner. It is easy to edit, and it has relatively strong verifiers. Open-ended image generation often sits in the bottom-left. The system can generate another image, but it cannot easily repair one specific decision or verify that the result is closer to the user’s intent.

One important property of the chart above is that the position of a task can potentially move by reframing a problem. The axes describe the representation, not the task itself. An open-ended image is hard to edit, but the same image, represented as SVG paths or a Blender scene, becomes editable — the task moves up. Give it a reference image or a set of constraints to check against, and progress becomes verifiable, which moves the task to the right. This is another way to describe loop engineering: not making the agent retry more, but re-representing the task until it sits in the quadrant where loops converge.

Loops are Discovered Before They are Engineered

I interviewed programmers from different domains who all are working on some form of loop engineering - from software engineering to visual and creative tasks to video editing. And I asked each one the same question - how did you know the loop would converge and iteratively improve?

The answer is today’s process of discovering a loop that works takes a lot of trial and error. It may be providing the right tool calls; or leaving the loop running for hours to see if it did much better than hours before (and if the improvement curve is promising). It may be going deep on specific workflows they have run in specific environments and replicating exactly that.

But discovering loops that can work everywhere is hard, and it’s almost like we are trying to encode the human knowledge into the loop itself; we first must deeply understand what makes the loop work, or find creative ways to build a verification layer, before attempting to automate it away.

However, finding a loop is only the upfront cost. Running the loop is the major cost that comes with every development cycle.

The Economics of Loops

So suppose the trial and error pays off, and you’ve found a loop that works. The simplest version of this is /goal [condition]. Keep going until the condition is met. And the loop will eventually get there.

“Eventually” is the problem. Would you run 20 iterations or 500? The honest answer is that the loop does not know, and neither does the human developer at the time of kicking off that loop, which makes the economics of running loops tricky.

What we do know is the shape of the curve. Across almost every study of test-time compute, returns are logarithmic: each additional increment of quality costs exponentially more attempts.1 One web-agent benchmark found that going from 1 sample to 10 lifted success from 38.8% to 43.2%. Doubling again to 20 bought 0.2 more points for twice the tokens.2 And past the plateau the marginal iteration can turn negative: reasoning models given larger budgets start abandoning answers that were already correct. More cycles do not just stop helping. They start hurting.3

I tested out one of the most popular loop examples in the wild from Anthropic’s own loop-engineering post:

and found that the incremental token spend moves the needle far less than the loop’s runtime suggests.

On a deliberately broken page (Lighthouse 35), Claude Code cleared 98 on the very first try for $0.35: the loop never engaged as expected. So I made the goal unreachable: same page, but served behind 2.2 seconds of artificial latency that caps the score around 89, and asked for 100.

The first $1.40 of spend took the score from 26 to 89. The remaining $2.84, 67% of the total bill, bought exactly zero points: turn after turn of re-minifying HTML and re-running Lighthouse against a bottleneck the agent couldn’t change, each turn more expensive than the last as the transcript grew (and the Haiku evaluator quietly accumulated $0.67 on its own). Worse, the loop’s escape hatch is unreliable: Claude correctly diagnosed the latency ceiling and declared the goal impossible around try 5, and the evaluator model bounced it back 14 times anyway.4

The lesson isn’t that loops don’t work; it’s that they have no idea how to stop. In this run, all the value landed in the first third of the spend, but the loop continued, burning tokens for an impossible task with marginal return.

Stopping well isn’t something one can prompt into existence. It takes infrastructure: something to meter the spend, something to measure progress against it, and something with enough information to cut the loop off. Loop engineering has an infra stack, and below are the layers.

The Stack for Loop Engineering

Once the loop becomes the unit of engineering, models and developers need infrastructure at every layer: an environment for the agent to act, a place to keep long-running state alive, a way to verify the work or close the loop, and a surface where humans steer. A stack has already formed around each category:

Inference Time vs Training Time Loops

Another lens to look at the loop engineering problem is from the perspective of inference vs training time.

At inference time, the loop changes the work, not the model. The agent writes code, runs verifiers, reads the result, and tries again. Its weights stay fixed and the system is searching for a better answer within one task leveraging test time compute.

At training time, the process looks like using reinforcement learning techniques that run many trajectories, scores the outcomes, and updates the model so rewarded behavior becomes more likely. The same rule applies in both cases: the loop is only as good as its verifier. In an agent loop, that verifier might be a test suite. In RL, it is the reward signal. Sometimes the two are the same.

The two loops can eventually feed into each other. Inference-time runs produce traces of what worked, what failed, and which corrections led to success. Those traces can become training data, preference pairs, or rewards, allowing the model to learn behavior it previously had to discover through expensive search. But not every failure should be solved through training. Often the higher-leverage fix is outside the weights: a better tool, clearer state, a more precise action space, or a stronger verifier.

Future Implications

Today, most agent infrastructure and harnesses can help us run loops. The harder problem is finding a loop worth running, and finding the point before diminishing returns for the task at hand.

Two things seem clear to me from watching these loops run.

The first is that the economics will have to become explicit. Right now we run loops the way we once ran cloud instances nobody remembered to turn off. The agent bills by the token, and the token costs the same whether it moves the score or re-minifies the same HTML for the ninth time. In my Lighthouse run, two-thirds of the spend bought nothing, and neither the loop nor I knew it until I read the trace afterward. The missing piece is boring yet necessary: cost per iteration, progress per dollar, a curve someone can see while the loop is still running.

The second is that for the loops that already converge, the interesting infra work has moved out of the loop. The loop itself is a while-statement and everything that makes it converge lives around it: the environment the agent acts in, the state that survives a long run, the verifier that decides what counts, the surface where a human steps in. Every working loop I’ve seen took a stack like this to build, and the stack is where differentiation actually sits.

So how does an AI model know its work is done? For now, it doesn’t. It stops when the budget runs out or when a check we designed says enough, and both of those need to be built. The systems that matter will not be the ones that can keep going. They all can. They will be the ones whose builders decided, precisely and in advance, what done costs and what done means.

If you are working on the loop engineering problem, doing research in this domain, or have thoughts on how this space will evolve, reach out to yli@a16z.com.

Subscribe now

This newsletter is provided for informational purposes only, and should not be relied upon as legal, business, investment, or tax advice. Furthermore, this content is not investment advice, nor is it intended for use by any investors or prospective investors in any a16z funds. This newsletter may link to other websites or contain other information obtained from third-party sources - a16z has not independently verified nor makes any representations about the current or enduring accuracy of such information. If this content includes third-party advertisements, a16z has not reviewed such advertisements and does not endorse any advertising content or related companies contained therein. Any investments or portfolio companies mentioned, referred to, or described are not representative of all investments in vehicles managed by a16z; visit https://a16z.com/investment-list/ for a full list of investments. Other important information can be found at a16z.com/disclosures. You’re receiving this newsletter since you opted in earlier; if you would like to opt out of future newsletters you may unsubscribe immediately.

America | Tech | Opinion | Culture | Charts

Today we’re pleased to share the American Dynamism Film Series: a spotlight on three companies changing how we build, power and defend America.

Ulysses: securing the seas

“We have some periods in time that are defined by land, and this century we’re moving into a century that is defined by the sea. The last wars were fought in jungles and deserts. The next wars will be fought over oceans. And I think this is going to position DoW as one of these anchor customers for companies like the three of ours, to fund early-stage R&D, to get these technologies pulled up the technology readiness scale.

“With the Cold War, we had the space race, and we saw this insane investment as a percentage of GDP into the space race. It literally pulled forward the technology timelines 50 years. We did in less than 10 years what people thought we’d do by the turn of the century. ... When you have that pull forward of technology, then therefore you get that you had to have NASA to have SpaceX, right? ... And I think this is important, because government is typically an anchor customer to pull forwards in these deep technologies. I think we’re seeing a transition in this century.” — Will O’Brien, Ulysses

Radiant: splitting the atom

“Nuclear reactors as products have never been seen before. They’re always usually these giant mega projects where you dig a huge hole in the ground and it takes five to 10 years, or up to 15 for the slower, bad projects out there. Ours, we’re targeting one per week coming off of a production line from our Tennessee facility, an 80-acre site we just signed for in October, not even a year ago.

Our product is for off the grid — megawatt reactor on a trailer. We build it in our factory, we drive it or fly it to where the customer wants it to go, and then turn it on within 48 hours. ... Electric power is civilization. If you go and there are sockets, and you pop something into them and you just get power, that’s very well developed — that’s civilization, using electricity to do what you could otherwise only do with human muscle or animal muscle.” —Doug Bernauer, Radiant

Read more: SMR (So Many Reactors) | Tori Shivanandan

Mariana Minerals: transforming the earth

“We focus on the full chain from mining all the way through refining. ... That handoff in the middle leads to a lot of actual market inefficiencies. We’re making a big bet on autonomy, fundamentally. ... We use reinforcement learning to actually remove humans from the loop in determining how refineries operate. When you have a highly variable feedstock — because the earth is heterogeneous — you need to constantly be tuning the temperatures, the flow rates, the chemical addition rates, the residence times of a highly complex refining circuit.

The same is true on the mining side of things — we’re making thousands of decisions a day, and when you don’t have the labor pool of folks who are able to make the right thousand decisions, that can cascade into low productivity, low availability of equipment, and low utilization of equipment.

What sets the rate of software and technology penetration into these plants and mines, ultimately, is the operating teams — what tech stack they’re comfortable with. For the most part, that’s pen and paper and maybe 150 spreadsheets scattered around an operation.” — Turner Caldwell, Mariana Minerals

Read more: Investing in Mariana Minerals | Erin Price-Wright and Ryan McEntush

Subscribe now

This newsletter is provided for informational purposes only, and should not be relied upon as legal, business, investment, or tax advice. Furthermore, this content is not investment advice, nor is it intended for use by any investors or prospective investors in any a16z funds. This newsletter may link to other websites or contain other information obtained from third-party sources - a16z has not independently verified nor makes any representations about the current or enduring accuracy of such information. If this content includes third-party advertisements, a16z has not reviewed such advertisements and does not endorse any advertising content or related companies contained therein. Any investments or portfolio companies mentioned, referred to, or described are not representative of all investments in vehicles managed by a16z; visit https://a16z.com/investment-list/ for a full list of investments. Other important information can be found at a16z.com/disclosures. You’re receiving this newsletter since you opted in earlier; if you would like to opt out of future newsletters you may unsubscribe immediately.

America | Tech | Opinion | Culture | Charts

In 2018, Zach Dell was a summer analyst at Blackstone studying the utility-scale battery opportunity. “One thing became really clear to me,” he recalls. “The marginal cost of solar plus storage was going to fall below the marginal cost of coal and natural gas.”

He was right. Solar is now the fastest-growing source of electricity in history, and the price of a lithium-ion battery pack has fallen more than 90% since 2010.

Yet, while the cost of generating electricity is falling, thanks to the collapsing price of solar (and wind), the cost of delivering it has surged. American utilities now invest more in the wires, poles, and substations that move electricity than in the plants that make it, and in 2025, electricity prices rose more than twice as fast as inflation. The U.S. electrical grid is an engineering triumph of the twentieth century that makes modern society possible, but it’s turning out to be a poor fit for the present.

Justin Lopas was circling the same problem, looking for the next big thing to work on after running manufacturing at Anduril and building rockets at SpaceX. Problems don’t come much bigger than electricity: if GDP per capita is the best measure of human prosperity, it’s hard to find a better lever to move it than electricity. Electricity is an input to almost everything, and it determines the viability of nearly all human activity. That’s why there are no rich, low-electricity countries.

By that measure, America is in trouble. Our electricity generation has been roughly flat since the mid-2000s, while China’s has more than quadrupled. Last year, China generated more than twice as much electricity as the United States, and as energy becomes the bottleneck on industries like AI and manufacturing, that gap is becoming a national security problem.

Zach and Justin met on a factory tour at Anduril, the defense startup that was recently valued at $61 billion. Thrive Capital, where Zach worked after Blackstone, had just invested in the company, and Justin, who ran manufacturing there, showed him around. The two quickly became friends, and in the months that followed, started discussing ideas for a new company. Both wanted to build something big and thought energy was ripe with opportunity. “What SpaceX did to aerospace; what Anduril did to defense; no one has done to the energy grid,” Zach says. The industry, as he sees it, pairs enormous scale with remarkably little innovation.

“The first question Zach and I asked,” Justin told us, “was: why isn’t energy cheaper?”

The future co-founders spent the next several months researching the industry until they zeroed in on the idea for their company.

Their core insight was that cheaper batteries could help fix our aging grid infrastructure because a battery and transmission line essentially do the same job: both move power from where it’s worth less to where it’s worth more. A transmission line does it through space, carrying power from where it’s made cheaply, like a solar farm in Arizona, to customers hundreds of miles away. A battery does it through time, storing power at noon when solar is flooding the grid and releasing it in the evening when everyone is home, the sun is down, and power is expensive.

The only problem was that the way incumbents were deploying battery storage was too slow, and the batteries weren’t where the grid needed them. “Ninety-nine percent of storage on the grid is utility-scale storage,” Zach explains. “Tens of billions of dollars of CapEx has been deployed into this asset class at high rates of return, but the asset class is fundamentally limited for two reasons.”

The first problem is the interconnection queue, which is the waitlist of projects that have applied to connect to the grid. Grid operators require that developers undergo studies before construction to ensure a new project won’t destabilize the system, but this has created a huge backlog. In 2008, a project took under two years to go from initial request to commercial operation; by 2023, it took nearly five. Some states can take even longer, and it’s blocking everything from batteries to solar farms to natural gas power plants. According to a June 2026 report from Berkeley Lab, there’s roughly 2,600 gigawatts of generator capacity and storage actively seeking interconnection – more than double the grid’s 1,279 gigawatts of existing capacity.

The second problem is transmission congestion. Where you actually need the power, like densely populated city centers, is not where you can put these huge farms with rows of batteries the size of shipping containers. So even after a farm clears the interconnection queue, its power arrives through the same crowded wires that are driving up delivery costs.

Zach and Justin believed the answer to both problems was to essentially chop the battery farms into thousands of pieces and install them on people’s homes. Houses are already connected to the grid, so you don’t have to wait in the interconnection queue. You also don’t have to purchase land or take on huge construction projects to install them. And because the house can draw power from the battery directly without ever touching the grid, you circumvent transmission congestion and reduce grid load.

The obvious problem is economies of scale: a battery farm has one site, one grid connection, and one construction job; a fleet spread across thousands of homes has thousands of each, which makes it more expensive per kilowatt-hour if installation costs are high. But Zach and Justin believed they had a way around that too.

The grid has been called the largest machine in the world – building it took decades, millions of workers, and hundreds of billions of dollars – and fixing it is an almost impossibly ambitious task. But in 2023, Justin Lopas and Zach Dell founded Base Power to try anyway, and it turned out that a lot of talented people wanted to work on this problem too. The engineer who got Starlink’s laser mesh network working joined to lead their software. A Tesla veteran who’d spent thirteen years on everything from the original Roadster to the Powerwall joined to build their battery. And investors have since bet billions of dollars that they’re onto something.

Whether they’re right will come down to a handful of beliefs about what electricity will look like in the coming decades. The first concerns a problem the grid has had from the beginning but never solved. To see it, you have to understand why the grid works the way it does.

Moving Power Through Space and Time

Thomas Edison is remembered for inventing the lightbulb, but that’s not quite what happened. Working incandescent bulbs already existed; Edison made them practical and safe. His lab in Menlo Park ran thousands of experiments with filaments and vacuums, and he sent his assistants on thousand-mile journeys before discovering a filament of Japanese bamboo that could burn for hundreds of hours. But a bulb is useless without electricity, so in 1882, he opened Pearl Street Station, America’s first centralized power plant, and wired it to a few dozen buildings nearby in lower Manhattan. Edison’s real breakthrough was the system.

The only problem was distance. Because Edison used direct current (DC), which fades as it travels, Pearl Street could only reach customers within about a mile radius. To electrify a city that way, you’d need a power plant in every neighborhood.

Nikola Tesla, a young engineer who traveled to America in 1884 to work for Edison, believed the answer was alternating current (AC), which could be pushed to high voltage, sent hundreds of miles with little loss, and stepped back down at the other end. Edison called his ideas “splendid,” but “utterly impractical,” so Tesla quit and raised money for his own laboratory. Tesla was a brilliant inventor who would go on to file hundreds of patents, but he lacked Edison’s gift for translating inventions into businesses. It was George Westinghouse, the Pittsburgh industrialist who had made his fortune inventing the railroad air brake, who saw Tesla’s genius, licensed his patents at a royalty of $2.50 for every horsepower of AC sold, and took on Edison in the bitter “War of the Currents.”

The war nearly broke Westinghouse. He burned vast sums of money fighting hundreds of patent lawsuits with Edison’s camp, and when the collapse of Barings Bank in London froze credit markets in 1890, creditors refused to rescue the company while Tesla’s royalties stood. So Tesla, who felt he owed everything to the one man who believed in him, tore up his contract. Westinghouse survived, and AC eventually beat DC because it was the superior standard for moving power through space.

Had Tesla kept his royalties, he might have died one of the richest men in history. Instead, he exhausted his remaining wealth on new experiments and died penniless, in debt, and alone in 1943, in the room he kept at the Hotel New Yorker.

The victory of alternating current set the shape of everything that followed. Because AC could travel vast distances, one enormous power plant could serve a whole region, which facilitated economies of scale. So the plants kept getting bigger and the wires kept getting longer. In the decades that followed, fueled by New Deal initiatives like the Rural Electrification Act of 1936, these regional systems expanded rapidly. Soon almost every American could flip a switch in their kitchen and use electricity from a plant miles away.

The critical flaw with this machine we now call “the grid” is that it is real-time with effectively no storage, and electricity has to be consumed within a millisecond of generation. If demand exceeds supply even briefly, the whole system slows down, almost like a bicycle hitting a steep hill, and if the frequency of the power drops too low, sensitive equipment will automatically disconnect itself to prevent damage, triggering a chain of widespread blackouts. If supply exceeds demand, the frequency spikes, almost like a bicycle spinning out of control downhill, which can overload lines, blow out transformers, and cause catastrophic damage to the infrastructure. In other words, the grid could move power through space, but not through time.

For most of the grid’s history, this was a flaw you could live with because the coal and gas plants that powered it could be ramped up and down quickly. It was more expensive because the whole machine had to be sized for the single worst hour and the average distribution line in America carried less than half of what it was built for, but it wasn’t catastrophic.

However, two things are now breaking the grid at the same time.

First, electricity generation is getting increasingly volatile. The cheapest source of power on earth is solar, with wind close behind, and the share of renewables in the energy mix is growing rapidly as coal plants are decommissioned due to their high maintenance and operating costs. This is a problem for grid stability because the sun and wind can’t be switched on or off like a coal power plant to match demand in real-time.

Second, the electrification of the economy is both increasing demand and making it more volatile. For over two decades, U.S. electricity demand grew at well below 1% per year. But data centers, new factories, and electric cars are all arriving at once, and utility grid planners are now projecting energy usage will increase at a rate of 5.7% per year from 2025 to 2030. Supporting this growth rate would require the electricity industry to build new generation and transmission capacity at more than six times the rate of recent years.

In short, supply is getting less reliable and demand is growing at its fastest rate in decades, on a machine built for a world of steady, controllable power plants and slowly growing demand. The tempting fix is more plants and wires, but capacity isn’t really the problem. In the middle of the night, the grid has power to spare; at 6pm the next evening, it strains. This is what Base was built to solve.

The Cheapest Electron Wins

Electricity is a commodity, and the best electron is the cheapest electron. “There are no sexy electrons,” Zach jokes. When he and Justin set out to build a modern power company, they designed it around this assumption.

The incumbents have the opposite incentive. “If you squint a little bit, they’re kind of similar to the defense primes,” Justin says, drawing the parallel from his years at Anduril. The utilities that own the poles and wires are regulated as monopolies and guaranteed a return on whatever they spend. Defense calls this model cost-plus; utilities call it rate basing. But the underlying incentive is the same: the more they spend, the more they earn. If technology lets you do more with less, there’s little reason to adopt it. As Zach puts it, “what you have is an incentive to build but not innovate.” So utilities’ requested rate increases continue to set new records, and electricity prices follow.

The advantage SpaceX and Anduril had competing against cost-plus incumbents was that customers could buy their products if they were better. But in electricity, how much competition is allowed varies state by state. About a dozen states, including Illinois, New York, Massachusetts, and Pennsylvania, have deregulated retail markets so homeowners can choose who sells them power. And no state has gone further than Texas, making it the perfect beachhead market for an energy startup to prove its model before expanding to the rest of the country.

Texas runs its own grid, called ERCOT, which does not cross state lines and largely avoids federal regulation. In the early 2000s, the Public Utility Commission of Texas broke up the utility monopoly model into three distinct businesses: generation, transmission, and retail. Transmission would still belong to regulated utilities earning a guaranteed return on their spend, but they would no longer own generation or retail, which were opened to competition. The resulting market dynamics have made Texas a laboratory for energy innovation. It has become the leader in wind and solar as pure-play generators compete to produce electricity as cheaply as possible (Texas sitting in the Sun Belt and Wind Corridor helps too), and a hundred-odd retailers compete to buy electricity wholesale from the grid and sell it to homeowners. There are still some regulatory hurdles to overcome, but in the roughly 80% of the state that has deregulated, a new company can enter generation or retail without an incumbent’s permission.

In 2023, Zach and Justin incorporated Base Power as a retail electricity provider and moved to Austin. Like Tesla, SpaceX, and Anduril, they would compete with the incumbents on talent, technology, and vertical integration.

The first step was hiring people who would never work for a utility. A lot of smart people already believed energy was one of the most important problems they could work on; they just needed a company worth joining. From SpaceX, Base hired Jared Greene, who led the team that built Starlink’s laser mesh network, to run software; Cole Jones, who ran Starlink’s go-to-market, to run growth; and Suzanne Dang, who ran procurement there for ten years, for special projects. From Tesla it recruited Dino Sasaridis, who spent thirteen years there and led the design of the Powerwall 3, to build the battery; and Andy Ross, who led battery manufacturing for the Model 3, to head up manufacturing. From Anduril it hired Dana Paz, who led manufacturing engineering, to run deployments. This founding team was critical in establishing an engineering-led culture, ramping domain expertise in key areas, and perhaps most importantly, attracting even more talent.

Base’s product is the Base Core, a 39.2 kilowatt-hour battery (roughly three times the size of traditional batteries) that installs directly on customer homes in less than an hour. But the battery isn’t what customers are buying. “We don’t sell batteries,” Zach explains. “We sell affordable, reliable power.” A homeowner pays a setup fee in the hundreds of dollars and, in some areas, a monthly membership fee of about $19. That’s for a battery that would cost well over ten thousand dollars to buy outright. In markets where customers can choose their own energy provider, they get electricity from Base for three years at a fixed rate plus delivery fees, typically saving 10 to 20% on their bill. In exchange, customers let Base use the batteries to trade power with the grid.

This model aligns Base with its customers in a way the rest of the battery industry isn’t. Base makes money by putting as much storage on the grid as possible, and customers want as much backup as they can get in case of a long-duration outage, so both sides want bigger batteries. Other battery companies have the opposite pull because their customers pay the sticker price upfront and smaller batteries are more affordable. Most of the money Base makes on each home comes from energy arbitrage: charging the batteries between 10pm and 4am when power is cheap and selling it back to the grid between 7pm and 9pm when it’s expensive. The trading profits scale with battery size and are what let Base sell cheaper electricity to the homeowner.

Base’s real competition is utility-scale storage, and its structural advantage is everything the homeowner supplies for free: the site, the grid connection that otherwise would take years in the interconnection queue, and a direct connection to the home that avoids transmission costs. Base is now extending those advantages to regulated utilities themselves. Austin Energy weighed Base against utility-scale developers and contracted 40 megawatts of home batteries. CoServ, the third-largest electric cooperative in the country, signed for 100. A year ago, utility partnerships were less than 5% of Base’s sales volume; today they’re more than half.

Base acquires homeowners through referrals, paid advertising, and word of mouth. It also partners with homebuilders like Lennar, one of the largest in the country. When people buy a Lennar home in certain Texas communities, they can sign up for a battery and have Base power their house from day one.

The tradeoff is that a fleet across thousands of homes means thousands of separate installs, without the economies of scale of one giant site. Base’s answer to that problem is a vertical integration flywheel. “We vertically integrate and develop technology to lower our costs,” Zach explains. “Lower costs equal higher returns at the asset level. Higher returns at the asset level mean we can pass on those returns to the customer in the form of lower prices.” In a commodity market, lower prices bring more demand, more demand brings more scale, and more scale loops back to lower costs. “There’s your flywheel, and that’s our competitive advantage. If you’re in the market for electrons and ours are the cheapest, you’re going to buy them.”

The battery itself is the clearest example. It’s designed so a crew can install it in less than an hour, eliminating hours of specialized electrician time. And bigger batteries deliver more kilowatt-hours per install. Base’s first product was 25 kWh, and within three years the 39.2 kWh Core shipped for the same price.

But vertically integrating a power company is easier said than done. Base has to design the batteries, manufacture them, write the software, install them, operate them, build a consumer brand, and offer support to thousands of customers. It’s also capital intensive.

Within three years of founding, Base has grown its battery fleet to over 500 megawatt-hours and expanded beyond Texas into Illinois. In October of last year, it raised a $1 billion Series C and converted an abandoned newspaper factory in downtown Austin into a fully functioning battery factory in eight months. This week it announced a $1 billion Series D alongside the launch of the Base Core battery, which is now in production there. The factory turns out thousands of systems a month, and the plan is to manufacture four gigawatt-hours of batteries a year, with over 10 at the next facility. “It is so critical to bring manufacturing back to the U.S., especially for critical infrastructure,” Justin says.

Today, Base deploys about 40 megawatts of battery storage per month to its fleet. Annualized, this run rate would represent nearly 2% of all the lithium-ion storage added to the U.S. grid last year.

“If you wanted to put 100 megawatts of batteries on the grid today, depending on the state, it would take you anywhere from two to five years,” Zach says. “We don’t need new poles and wires. It’s a much faster system.”

Battery fleets are starting to have a noticeable impact on the grid too. Last month, Texas set a new July demand record, and wholesale prices peaked at about $0.06 per kilowatt-hour. For comparison, when record demand hit Texas during the summers of 2023 and 2024, evening prices spiked above $4 per kilowatt-hour. This time, solar carried nearly a third of the record load, batteries carried the evening ramp, and gas, which historically ramped to meet the peak, barely moved.

Two weeks later, it happened again, but bigger. The all-time Texas demand record that had stood for almost two years fell twice in two days – 87.5 gigawatts on July 21, then 91.3 on July 22. Batteries supplied nearly 12 gigawatts at the peak, triple what the whole state had two years ago, and wholesale electricity prices briefly touched $0.30 per kilowatt-hour, less than a tenth of the 2023 and 2024 spikes. The roughly 150 megawatts Base discharged that day is about the size of a full utility-scale battery site. If you had started building one in 2024, it would still be stuck in the interconnection queue.

Over time, Base expects the rest of the country to look a lot more like Texas: more solar, more batteries, and steeper demand spikes. “We think Texas is the canary in the coal mine for the rest of the country,” Zach says, describing the company’s expansion into new states. And after new states, there will be new products.

“If you have a battery and an inverter on the home and you’re selling the homeowner power every month, you’re really well set up to add solar to the equation,” Zach says. “We want to be in a position where we can land a battery, and eventually a solar panel, on the grid cheaper than anyone on the planet on a dollar per kilowatt-hour basis, which means we can sell an electron cheaper than anyone on the planet.” The plan is to run the same vertical integration flywheel on solar that added nearly 60% more storage to their battery for the same price.

Solar is also Texas’s fastest-growing source of power, up from almost nothing a decade ago to about 15% of the state’s electricity last year.

The Coming Decade of Solar and Batteries

Last year, the world installed more new solar capacity than every other energy source combined, and very few people saw it coming. Every year for two decades, the International Energy Agency (IEA) projected that solar growth would level off, and every year the exponential has continued.

How could the experts be so wrong for so long? They were modeling solar like traditional forms of energy, when it’s actually a manufactured product, more like a flat-screen TV than a power plant. It has no moving parts, doesn’t require specialized labor to implement, runs on sunlight instead of fuel, and most importantly, it follows a “learning curve” – when factories make more of it, they get better at making it, and it gets cheaper. Economists call this learning curve Wright’s Law, which observes that the cost of technology falls by a constant fraction every time cumulative production doubles. For solar panels the fraction has been about 20% per doubling and has held for nearly five decades.

For batteries, it’s been about 23% per doubling, which compounds solar’s adoption because batteries fix solar’s biggest weakness (sunsets).

Energy historian Vaclav Smil points out that energy transitions have been slow. One way he demonstrates this is by plotting how long it took an energy source to go from 5% of the market to 25%: coal took 35 years, oil took 40 years, and natural gas took 55 years. But coal, oil, and natural gas never got 20% cheaper every couple of years.

You can see the effect of Wright’s Law on the American grid already: solar has gone from about 5% of the country’s electricity generation in 2022 to 9% in 2025, and it’s 51% of the new capacity being built in 2026.

We’re also seeing more demand for solar at the residential level to power individual homes. Utility-scale solar has accounted for roughly two thirds of U.S. solar power capacity due to economies of scale versus rooftop solar, but suitable land and the interconnection queue are becoming bottlenecks. Meanwhile, pairing solar panels with battery storage to power the night is becoming economically feasible. In the last five years, the share of American homes with solar has more than doubled from 4% to 9%.

Another important driver of residential solar adoption is the rising delivery costs Zach and Justin identified when they first set out to fix the grid. Most of a retail electricity bill is delivery, and rooftop solar skips delivery entirely by letting the homeowner generate electricity where it is consumed. A solar farm competes with wholesale prices, but rooftop solar competes with retail prices.

Even before it sells a panel, all of this is a tailwind for Base. As solar comprises a larger share of power generation, noon power gets cheaper and evening power more valuable. And Base’s batteries are paid to move it through time.

Batteries and solar are ultimately how Base puts power back in the hands of the consumer. The AI data center buildout has utilities planning for demand growth the grid hasn’t seen in decades, and homeowners are worried the cost will land on their electricity bills. The industry’s proposed fix is for data centers to flex their load around the grid’s peaks, but that requires visibility today’s grid doesn’t offer. Every battery Base installs adds telemetry and control to another node on the grid, and enough of them give the system the visibility that flexibility requires. Data centers could even buy power from batteries on hundreds of thousands of homes nearby instead of waiting for years in the interconnection queue. It could be the opposite of what people fear, Zach says, “where these hyperscalers are actually subsidizing the power costs for the consumer.”

The Great Problem of Science

In 1900, at the height of his fame, Nikola Tesla published an essay titled The Problem of Increasing Human Energy. “The great problem of science,” he argued, “is, and always will be, to increase the energy.” He thought burning fuel was barbarous because destroying material is wasteful and neglects “our duty to coming generations” to leave their stores of energy intact. “We ought to be able to obtain the energy we need without consumption of material,” he wrote. He studied wind and solar and concluded the power was too intermittent and storage cost too much. The first step toward the future he envisioned was a better battery: “These and many other problems will be better solved, and in a more scientific manner, by a light-storage battery.”

A century and a quarter later, the battery Tesla asked for is finally cheap, and solar is on path to be the cheapest source of power on the planet. But deploying it is getting slower and more expensive. Big projects are forced to wait years in the interconnection queue, and our aging grid infrastructure is driving up the cost of delivery.

In three years, Base has gone from zero to installing more home battery storage per month than any company in America. The goal, as Zach described earlier, is to land a battery, and eventually a solar panel, on the grid cheaper than anyone on the planet, and turn the sun’s abundant energy into cheap, reliable power.

Driving down the price of electrons will benefit everyone who buys electricity, which is to say everyone. It will also drive down lots of other prices because energy is an ingredient in almost everything. People worry about a global water crisis on a planet that is 71% ocean because the energy required for desalination makes it prohibitively expensive almost everywhere. The technologies that will define this century consume huge amounts of energy too, and the countries with the cheapest electrons will be where the data centers, the factories, and eventually the robots get built.

Most big jumps in human progress trace back to jumps in our ability to harness energy. The airplane is a good example. When the first long-distance power line connected Niagara Falls to Buffalo in 1896, cheap and steady hydroelectric power made it profitable to manufacture aluminum at scale, and aluminum was the metal the Wright brothers needed to build an engine light enough to fly. The same cheap energy also spawned factories across the city that produced abrasives, silicon, and graphite. One improvement in moving electricity kicked off dozens of industries.

But America will need more than cheap electrons. China accounted for over 80% of battery cell production in 2025, and its share of every stage of solar panel manufacturing (polysilicon, ingots, wafers, cells, and modules) exceeds 80%. Washington has responded to this supply chain risk with tariffs, which pushes energy prices higher still.

Base wants to help solve this too by building more factories and vertically integrating further over time, but they can’t do it alone.

“In the last fifty years, the electricity industry has not been the place where the most talented engineers and operators have gone,” Zach says. “I have a ton of optimism that talented young people will wake up to the idea that this is an incredibly important problem. I hope a lot of them come to work at Base, but I also hope more companies get started to solve these really hard engineering problems in the energy space to help drive cost down and reliability up.”

As Justin saw at SpaceX and Anduril, it only takes a small group of people to jumpstart an industry. It’s hard to think of one with more downstream consequences than electricity.

Subscribe now

This newsletter is provided for informational purposes only, and should not be relied upon as legal, business, investment, or tax advice. Furthermore, this content is not investment advice, nor is it intended for use by any investors or prospective investors in any a16z funds. This newsletter may link to other websites or contain other information obtained from third-party sources - a16z has not independently verified nor makes any representations about the current or enduring accuracy of such information. If this content includes third-party advertisements, a16z has not reviewed such advertisements and does not endorse any advertising content or related companies contained therein. Any investments or portfolio companies mentioned, referred to, or described are not representative of all investments in vehicles managed by a16z; visit https://a16z.com/investment-list/ for a full list of investments. Other important information can be found at a16z.com/disclosures. You’re receiving this newsletter since you opted in earlier; if you would like to opt out of future newsletters you may unsubscribe immediately.

America | Tech | Opinion | Culture | Charts

Overview

Cutting undersea cables is now part of international conflict. The cables on the seabed carry almost all data that matters between continents: finance, intelligence, commerce, and the entire new economy of AI. Over the past three years subsea fiber has been severed in the Red Sea, the Baltics, and around Taiwan in incidents that closed without consequence.

You can’t safely run many petabits of data under the ocean anymore, and the only direction left is up, leveraging space. This used to be the stuff of government research budgets and universities. Today, so much has changed in the space economy that now makes this possible. Hardware is cheaper, talent is denser, and there is proof that with meaningful investments, there can be outsized financial upside. The engineering and economics finally close.

EON is taking on the data transfer problem with lasers in space, beaming fiber level data between ground stations and satellites in orbit. We’ll offer route-diverse intercontinental bandwidth that never needs to touch the seabed.

How the world moves data today — the status quo

Global internet data moves the same way it did thirty years ago.

95% of intercontinental data traffic moves over subsea cables spanning thousands of miles, buried on the seabed. A single cable system takes nearly a decade to go from planning to operation — environmental permits and landing rights across every sovereign nation it touches, custom manufacturing, and months of laying at sea by a global fleet of fewer than 60 specialist vessels.

Today there are only about 500 active subsea cable systems in service. They are increasingly at risk.

State of the world’s intercontinental data highways

For most of the internet era, up until the mid-2010s, transoceanic systems were built by consortia of telecom carriers. A dozen or more national and private operators would pool together capital, sharing the cost of a system no single one of them could justify, and divide the capacity. Capacity would be sold out to other companies as needed. At the time, no one party controlled a route.

About fifteen years ago, that all changed.

As their own traffic outgrew what they could lease, the largest content and cloud companies started funding cable laying directly. The logic was straightforward: at their scale, owning the cable was cheaper than renting capacity on it. Ownership let them control routing, timing, and cost in a way leasing never would.

The shift was fast.

Just over a decade ago, content providers accounted for a small minority of used transoceanic capacity; today they account for the majority of it, and they own or co-own a large and growing share of new builds, particularly on the transatlantic and transpacific routes that carry the most traffic.

This was a rational response to a slow system. The world’s largest companies, such as Google, Facebook/Meta, etc, could absorb a ten-year timeline and a billion-dollar bill as a line item. Almost anyone else cannot.

But that system has a ceiling. The fleet of cable-laying ships is fixed and aging, and the manufacturers who build optical repeaters and the crews who splice cable are few and far between. No amount of capital compresses a permitting cycle that runs through numberous governments’ EEZs or territorial waters that a cable touches. The supply chain that builds the ocean floor is effectively capped, and it is already running flat out.

Demand is not capped.

More data was created in the past two years than in all of prior human history, and a single hyperscaler now moves more across its own network in a day than the entire internet carried in 2010. Every AI model trained, every workload shifted between continents, every new region brought online pushes the curve higher.

So the squeeze falls on both the companies with the most data to move and anyone trying to build new global scale businesses. The largest companies can fund cables and still can’t build them fast enough; every new entrant is competing for whatever capacity is left on a route someone else decided to build. Whether you own or rent, you are running out of room on a system that cannot grow as fast as the data crossing it.

Obvious targets

Fiber routes are public knowledge and every cable, every landing station, every chokepoint is documented. This makes subsea internet infrastructure an easy target for adversarial intervention.

There is already an existing history of cable outages, much of it intentional.

When Seacom, TGN-EA, and AAE-1 were severed in the Red Sea in early 2024, roughly a quarter of the traffic moving between Europe, Asia, and East Africa had to be rerouted overnight at meaningful cost. The cables were cut during an active conflict in a conflict zone, yet no party was charged.

In 2022, SMW5, a cable connecting South East Asia to the Middle East to Western Europe was disrupted, and connectivity to the affected regions dropped by a third overnight. No none entity responsible for the cutting has been named.

In waters near Taiwan, undersea links have been cut repeatedly, with Chinese fishing and cargo vessels logged nearby each time and dozens of similar incidents since. None of these produced a confession or a confident public attribution, which is exactly what makes the seabed attractive.

When any node in this global network is cut, the repair runs weeks to months with no hot spares. Outages that would be unacceptable in any other critical infrastructure are quietly accepted as the cost of doing business.

Capabilities to disturb subsea fiber are not kept secret.

Russian deep-sea operations under GUGI, Moscow’s underwater research directorate, have spent years surveying and loitering near cable routes across the Atlantic and Baltic. In March 2025, Chinese researchers published the design for a compact device able to sever the most heavily armored cables in existence. Former NSA Director Mike Rogers has testified that adversaries can cut cables where and when they choose.

In ordinary times accidents or malfunctions cause most faults, but whether a cut is accidental or deliberate, the consequence is the same. After a cut, entire regions can go dark for weeks with no alternative paths. This network fragility does not depend on anyone’s intent, and adversaries now know how to use it.

How EON works

EON is built to service hyperscalers, enterprises, sovereigns, and organizations with so much data that existing in-space infrastructure is not sufficient.

Through an optical laser relay to Medium Earth Orbit (MEO) and back down in a “bent-pipe”, EON can deliver multiple terabits of data per second across oceans.

One hop. Two endpoints. Nothing on the ocean floor.

A ground station points a laser at a satellite, the satellite catches it and fires it back down, to a receiver on the other side of an ocean. Ground stations connect to existing terrestrial fiber, far easier to deploy than anything requiring a large ship.

Capacity will come online in months, not years, and can serve anywhere with a clear view of the sky.

Why this hasn’t been done before

Space is hard, but it is getting easier.

Ten years ago the spacecraft for this exact system would have cost at least ten times what it costs today. The space industry has completely shifted away from bespoke over-engineered exquisite systems to commercial off-the-shelf hardware.

The space industry today has real standards, proven components, and a supply chain that didn’t exist a decade ago. That’s what makes EON viable as a business, not just as an engineering project. Every decision we make is with the current state of the art, because that’s what closes the economics.

The importance of a good business model.

EON is building to last. We want to be the next trillion dollar space company. To do that, we need a business that closes from the start.

There are countless venture backed space companies attempting science projects with the hope of economic viability in the future. That is not us. Engineers who work at EON are going to work on some incredibly hard technical challenges, but the light at the end of the tunnel is what we believe to be an incredibly strong business.

Incumbents move data across subsea fiber because no one has delivered a credible alternative. It is a multi-billion dollar market held in place not by the superiority of the technology but by the absence of competition. We are going to change that, not just as a capable alternative, but a more affordable one.

We are building to move bits at lower cost than subsea. That is the business. That is the test we hold ourselves to.

Why we are building EON

We are just getting started. The engineering challenges ahead of us are real and we are not pretending otherwise. What orients us is our mission: secure, affordable global connectivity.

The fragility of the infrastructure the world runs on is no longer a theoretical concern. It is a documented, recurring, and in some cases deliberate reality.

The world has changed, but the infrastructure has not.

EON is here to change it.

Subscribe now

This newsletter is provided for informational purposes only, and should not be relied upon as legal, business, investment, or tax advice. Furthermore, this content is not investment advice, nor is it intended for use by any investors or prospective investors in any a16z funds. This newsletter may link to other websites or contain other information obtained from third-party sources - a16z has not independently verified nor makes any representations about the current or enduring accuracy of such information. If this content includes third-party advertisements, a16z has not reviewed such advertisements and does not endorse any advertising content or related companies contained therein. Any investments or portfolio companies mentioned, referred to, or described are not representative of all investments in vehicles managed by a16z; visit https://a16z.com/investment-list/ for a full list of investments. Other important information can be found at a16z.com/disclosures. You’re receiving this newsletter since you opted in earlier; if you would like to opt out of future newsletters you may unsubscribe immediately.

America | Tech | Opinion | Culture | Charts

Two years ago, Marc and Ben framed the Little Tech Agenda. The political, regulatory, and economic playing field had tilted toward Big Tech; a16z, they said, would fight to preserve startups’ ability to compete.

That fight for equal opportunity remains central to how we think about our mission at the firm. But in the last few years, a new constraint has emerged: access to compute. Scaling laws—the empirical regularity that more compute means better performance—have made compute a fundamental input for building more capable systems.

And compute isn’t distributed evenly. Hyperscalers with investment-grade balance sheets—and the largest customers they prioritize—can make the five-year commitments required to secure land and power, purchase GPUs, and preorder scarce components. Most startups don’t have financeable balance sheets: they’re funded 18 months at a time. So they have to pay more money for less flexible compute access, and have to wait longer to get it. And that’s if they can secure short-term or on-demand supply at all: this year, much of it has been bid away by Big Tech.

Part of the problem is that physical compute—the actual GPU racks—is in short supply. But it’s also a financing problem. Production-grade data centers require billions of dollars of capital; the investors financing them want long-term contracts with creditworthy customers. That test applies to everyone, from frontier labs to hyperscalers. And that test excludes a huge swathe of startups with the ideas and talent to build the future of AI.

That’s the gap that Volta was founded to close. We believe that Volta is building a new and different kind of AI infrastructure company, combining cloud operations with project finance to unlock new sources of compute. Their model benefits a broad range of customers, while uniquely making capacity available to startups that have historically been priced out of long-term infrastructure. Volta is building the neocloud for Little Tech. That’s why a16z is excited to co-lead its Series A.

A New Model for AI Infrastructure

Neoclouds emerged to build and serve computing capacity specifically for AI workloads. They’ve been a tremendous boon to the AI industry: they’ve increased supply, unbundled the cloud, and created meaningful alternatives to the hyperscalers. But they haven’t fundamentally changed the question of who gets capacity first.

Building a data center is expensive—so expensive that their construction is financed overwhelmingly by debt, underwritten against long-term customer contracts. Lenders might believe deeply in the growth of AI; but they still need a creditworthy counterparty on the other side of a five-year commitment. The cheapest and largest pools of capital therefore flow to the strongest balance sheets. Those belong, of course, to the hyperscalers—the largest and most profitable companies on earth.

Most large neoclouds have followed that capital. They’ve sold capacity to hyperscalers and acted as extensions of their first-party infrastructure teams. That makes sense for the neoclouds. But it means that companies without the hyperscalers’ balance sheets are left with the scraps: they have to compete for expensive on-demand capacity—whatever capacity remains, that is—or a place on a waitlist. The market has produced more AI clouds, but not necessarily more access for the startups building AI.

Volta inverts that model. It’s building for AI-native companies first and foremost: from frontier labs and emerging neolabs to fast-growing AI apps. Instead of asking each customer to arrive with a bespoke investment-grade backstop, Volta assembles the credit support, project equity, and debt behind each deployment.

That approach has already yielded major successes. Volta has signed a $10 billion strategic partnership for a 133 MW deployment in Norway, assembling the site, project equity, and infrastructure debt without a hyperscaler or Nvidia backstop. The deployment shows that Volta can bring customer demand, land, power, equipment, and capital together at scale.

Volta also has a $5 billion infrastructure program with Azora, a real estate and infrastructure asset manager, for project equity, alongside senior infrastructure debt led by international banks. That allows the company to move from one deployment to the next without raising new equity or rebuilding the financing stack each time.

Across Capital and Compute

Getting compute online fast requires mastery of sharply different domains. You need the financing in order to make everything possible; the physical infrastructure expertise to get it built fast; and the software to turn that compute into a usable product. Volta has a team that understands every side of this problem.

Ricard Boada and Sofia Gumuzio built Brookfield’s AI infrastructure platform, giving them firsthand experience with every constraint between customer demand and live compute. Ricard brings deep expertise in financing and developing data centers. Sofia brings relationships across land and power. Together, they know how to turn a customer’s request for compute into powered capacity.

They’ve acquired Genesis Cloud’s team, which has operated one of Europe’s earliest GPU-first clouds since 2018 and served more than 20,000 users with a proven platform. The Genesis stack and team bring Kubernetes-native software and years of experience provisioning GPUs, managing clusters, and operating a production cloud. That’s the other half of Volta: turning physical capacity into a product developers can provision and trust.

Neoclouds tend to begin with one half of the problem: they have capital and real assets, or they have cloud software and GPU operations. Volta has both. There aren’t many neocloud teams that can actually secure power, structure billions of dollars of capital, develop data centers, and run a production cloud. Volta has assembled one of the few teams that can.

We’re thrilled to partner with Ricard, Sofia, and the Volta team as they expand access to AI infrastructure and ensure startups can compete alongside the industry’s largest players.

Subscribe now

This newsletter is provided for informational purposes only, and should not be relied upon as legal, business, investment, or tax advice. Furthermore, this content is not investment advice, nor is it intended for use by any investors or prospective investors in any a16z funds. This newsletter may link to other websites or contain other information obtained from third-party sources - a16z has not independently verified nor makes any representations about the current or enduring accuracy of such information. If this content includes third-party advertisements, a16z has not reviewed such advertisements and does not endorse any advertising content or related companies contained therein. Any investments or portfolio companies mentioned, referred to, or described are not representative of all investments in vehicles managed by a16z; visit https://a16z.com/investment-list/ for a full list of investments. Other important information can be found at a16z.com/disclosures. You’re receiving this newsletter since you opted in earlier; if you would like to opt out of future newsletters you may unsubscribe immediately.

America | Tech | Opinion | Culture | Charts

We’re at a pivotal moment for the CLARITY Act—the landmark crypto market structure legislation making its way through Congress. Marc Andreessen and Chris Dixon did a podcast together to share their views on what’s at stake if it passes or stalls.

Robert: Congress is currently hashing through some legislation — once-in-a-generation market structure legislation — that could determine where the future of the financial system and internet gets built. It passed in the House last year with bipartisan support, and it has been making its way through the Senate ever since.

Now, there have been lots of fights and obstacles over it, which we’ll get into. But before we do, I wanna zoom out and go big picture and talk about why regulatory clarity matters, what the status quo costs us, and what’s at stake for the country and for anybody who may end up using this technology.

So Marc, I’d like to start with you. Back in January 2014, you wrote an op-ed for The New York Times called “Why Bitcoin Matters.” Things were a little bit different then. This was a pretty contrarian position to stake out. The crypto industry, which, if you could even call it that back then, looked very, very different.

How have things changed from then to now?

Marc: Yeah, sure. So 2014 was so long ago that The New York Times actually ran a positive piece on crypto. This was in the distant past. It might as well have been the 1500s.

I’m actually really proud of that piece. I think it holds up in a few respects. This piece is five years into the macro trend. And it was still big news even to believe that this technology mattered at all.

Like, there were still a lot of people who were going to have to learn about this, want to learn about this, and be part of it. And so I think that holds up. The thing that, in retrospect, needs to be changed about the piece is every time I say Bitcoin, swap in crypto, right?

And so the idea was that Bitcoin would be extended to basically enable both tokenization of real-world assets and also virtual assets like NFTs. Of course, that didn’t happen. And what happened instead was the creation of new blockchains and new crypto platforms, and then ultimately Ethereum and everything else that followed.

And so what started as a technology became an industry. Now, Bitcoin obviously has been very successful since then, but what we’ve really seen since then is this explosion of innovation.

Robert: It’s actually remarkable how well that piece holds up.

But there is one forecast in there that I want to cite in particular that stands out: “The coming years will be a period of great drama and excitement revolving around this new technology.” I think you were spot on there.

Chris, you’ve been there since the start as well. How have things changed from the early days to now?

Chris: So in the early days, it was mostly kind of hobbyists and super enthusiasts, and kind of a subculture, cultish thing.

The technology, which was Bitcoin, was sophisticated, but — a lot of the things Marc alluded to — the newer blockchains early on had performance issues, scaling issues, and other kinds of things. Fast-forward to today: Every day you’ll see a new announcement from a major bank or fintech company announcing a platform involving stablecoins or tokenized stocks or other kinds of digital assets.

So stablecoins — for those who don’t know — are basically what Marc described with Bitcoin, except they’re dollars on blockchains. They now rival the size of the Visa network, with trillions of dollars transacted. You can go into WhatsApp and send money anywhere in the world for almost free in the same way you’d send a text message, using stablecoins.

It’s the way money should work.

Marc could speak to this as one of the pioneers of the internet — but I think it was one of the ideas that early internet pioneers expected to happen much sooner. And for a variety of reasons, it took a lot longer, but we’re now finally seeing that dream come alive, with money moving as easily as bits.

a16z crypto

your guide to the next internet from the team at a16z crypto

So now it’s a very mature industry. The technology — the underlying infrastructure — has gotten much more sophisticated. Even three years ago, what I just described might cost a couple of dollars, if not tens of dollars, to do that transaction. Now, on most of the popular blockchains — Solana, Ethereum, and so forth — it’s sub-one second to settle the transaction, sub-one penny to send something.

Why crypto needs rules now

Robert: You cited that stablecoin volumes are now rivaling the Visa network. This is trillions of dollars per quarter, churning through this system. Lots of gigantic financial institutions have rushed in: BlackRock, JPMorgan, Visa itself, Fidelity, Mastercard.

I mean, you could rattle off just any number of names, and they’re all building on this technology right now. But we’ve gotten this far without rules. So there’s been very little in the way of regulatory clarity or guidance.

Why do we need rules now?

Chris: So for a variety of reasons, the regulation around crypto was broken into two components — stablecoins and the rest of the market.

Last year, a bill passed Congress — signed by the president — called the GENIUS Act, which provides a comprehensive regulatory framework for stablecoins. And it’s not a coincidence that stablecoins are also the area that’s taken off the most in the last year. Once you have regulatory clarity, builders have certainty. If you’re an American consumer, you know — if you use USDC or another GENIUS-compliant stablecoin — that if you have a dollar of that stablecoin, there is a dollar sitting in the bank.

That gives you, as a consumer, confidence and protection. If you’re an institution — if you’re a bank, if you’re Stripe, if you’re PayPal — and you want to go and enter this market, you want to know there are predictable rules. That there’s a framework around it. That what you build will be compliant, not just next year — but in the next 10 years. You want to know there’s certainty.

But the big, hanging problem right now is that you’ve got this other section of the market — including blockchains themselves, which stablecoins are built on — that doesn’t have a comprehensive federal regulatory framework. And that’s why the CLARITY Act, which is working its way through the Senate, is so important.

So you can imagine it’s as if we had a regulatory framework for cell phones, but not for cell towers or something. Half the technology is regulated, the other half has this sort of uncertainty around it. People are still building there. People are trying to do their best to understand what the rules are.

There’s this other lever of policymaking. Agencies such as the SEC and CFTC have provided some guidance. But ultimately what industries are built on — as you alluded to in the beginning with the internet — is legislation.

So we think that’s what is required, and why the CLARITY Act is so important.

Robert: So we do have some rules. They’re for stablecoins, as you mentioned. That covers a sliver of the market, but there’s this gigantic swath of the market that remains largely unregulated.

Marc, why do we need rules for all of this, and why do we need them now?

Marc: The US financial system has been through this many times before. The most famous case is the Securities Acts that created the SEC and regulated the stock market.

We’re not looking for a free lunch. We’re not looking for subsidies. We’re not looking for protectionism. We’re not looking for support in that way. We’re just looking for a permanent framework that lets people do business responsibly. And I think it’s — in a lot of ways — the most obvious thing in the world.

Chris: Right now, there is no federal regulator for crypto exchanges, as an example.

The New York Stock Exchange, Nasdaq — they have federal regulators.

A big chunk of the CLARITY Act is to provide that, so that the SEC and the CFTC have oversight. They put in disclosure regimes, anti-fraud regimes, insider trading rules, and all the things that we are used to — as Marc was saying — with other financial markets.

It’s just the same kinds of rules with the same agencies overseeing them. So for example, FTX wasn’t audited properly. This is what ultimately happened, right?

If you’re a federally registered crypto exchange, you’re being audited — there’s a whole suite of things that are in the act that would empower federal regulators to do their job.

And if you’re not compliant with these things, you couldn’t operate in the US.

There are a lot of hard details, and this has been a very bipartisan effort, which is why it’s been seven years or so that some of us have been working on the CLARITY Act. It’s now been over a year since the House passed its version of it, and the Senate has been working on it for a year.

But this has been done before: Financial industries have been regulated, and the principles that are used in other industries are applied here. There’s a very strict framework with strict definitions of who is subject to these requirements. It applies all the same money laundering and Treasury rules to crypto intermediaries that are applied in other markets.

The Fraternal Order of Police, the largest law enforcement organization in the country, just came out with its endorsement of the CLARITY Act. Opponents are saying it doesn’t have proper law enforcement clauses and rules, but this is simply not true.

Illicit finance and how ambiguity rewards bad actors

Robert: So let’s go over each of the debates going on. Illicit finance is one of the big ones.

Chris: Right now it’s extremely unclear which entities are subject to what rules. What I’ve learned is that when you have gray areas in regulation, you essentially race to the bottom, right?

I was on the board of Coinbase for a long time — Marc’s on the board. Coinbase is a US-based company. They take regulation and compliance extremely seriously. But for every Coinbase, it costs a lot of money, and it slows down your product development.

And so every year, a new offshore competitor will pop up that doesn’t do those things and becomes popular because they can offer lower fees or iterate on their product faster, because they’re not doing all the compliance things.

The ambiguity ends up favoring the bad actors. And so good regulation — I think the CLARITY Act does this — very clearly defines the regulatory perimeter.

So as an example, in the CLARITY Act, if you’re a financial intermediary, if you’re somebody like Coinbase or a company that’s holding customer funds, you’re subject to all of the same rules as a company like Stripe or PayPal or anyone else.

It’s black and white in the bill. And as I mentioned, there’s the endorsement of the various law enforcement agencies.

Does CLARITY enable sanctions evasion? Why privacy is not the same as secrecy

Robert: Marc, what do you think about this criticism that the bill enables sanctions evasion?

Marc: I would say the national security people I talk to typically don’t agree with that, because there’s a trail. And that is contrasted specifically with how a lot of, for example, terror financing happens today.

There have been people — even in industry — who have said things like, “Well, crypto is inherently anonymous. You can’t trace anything.” I think this stems from them not understanding the technology.

Chris: It’s actually a non-trivial research problem that’s going on today to build blockchains that do have privacy because they are so public.

We think that’s important — just not for nefarious reasons. Because if you want to pay for some service — maybe it’s a medical thing or a financial thing — you don’t want everyone to know about it, just like we do in the regular financial system. You want to have some privacy. It’s a natural thing.

Since the GENIUS Act passed and dollar-denominated stablecoins have gotten more popular, we’ve actually heard in DC now that people are saying, “Well, what about privacy? You need to have privacy.”

And like, yes, we agree. It actually is very reminiscent of the early internet when, Marc — wasn’t it you guys that invented SSL?

Marc: Yeah, we did.

Chris: And so for the kids who don’t know this — HTTPS was Netscape.

And at the time it was like, well, who would need that? Why would you need to send secret stuff? And at the time, probably some bad guys did use it, but of course, over time, we all realized that was important. And 99.9% of it is, like, good behavior. It’s just people interacting with their bank or something.

You guys had to go in front of Congress.

Marc: It was a four-year fight. It was actually quite analogous to today. And so when we started, Netscape was the first piece of consumer software that got widely used that incorporated public key encryption — modern encryption with public keys.

And at the time we shipped, we actually had to ship two different versions. We could ship a strong encryption version inside the US. Still, encryption was classified as a munition under what were called International Traffic in Arms Regulations (ITAR) export control rules at the time.

So encryption was classified in the same category as a Tomahawk missile, which meant that the Netscape browser was classified in the same category as a Tomahawk missile. So we had to ship deliberately insecure versions of the product overseas — and you can imagine how much our overseas customers enjoyed that.

And by the way, we clearly labeled it on the box, which was like “weak encryption — do not trust this.” They get this, and you can imagine how that went for sales. And of course, what happened was foreign competitors immediately stepped in, started cloning it, and started making strong encryption.

So people outside of the US didn’t use our product. They used products that were made outside the US. And so we went to Washington — as you do when you’re a kid, you go to Washington — and you explain things logically, and they tend to stare at you like you’ve grown a third eye.

And so you start to get into the specifics of it. But then you get into this fundamental dichotomy, which is like, okay, is encryption bad because bad guys are going to do bad things with it? Or is it good because it is the key to establishing trust, being able to do business, and for law-abiding citizens, both here and overseas, to cooperate and conduct business?

And so it’s like, do you want to eliminate the risk of any cybercriminality? Or do you want to make Amazon.com possible? And these questions are deeply intertwined. And so anyway, it was this very long education process.

It took four years. It took a very, very long time to unwind. They did ultimately make that change and, by the way, the world did not end for several reasons. One is, as you said, overwhelmingly the use cases were positive.

And the other incredibly positive thing that happened was American industry won. Not just Netscape, but the global internet economy is dominated by American companies.

How ethics and market structure are separate questions

Robert: So let’s talk about some of the other objections to the bill. Another big one is ethics reform. So the president and his family members have interests in crypto businesses. The critics say that this bill will enrich the people who are most vocally supportive of it. How do you address that?

Chris: It’s not my area of expertise, but I’ll say as a citizen, I think there should be ethics rules for government officials. I don’t think those should be exclusive to crypto. I think there should be ethics rules around trading stocks and, for sure, other financial assets. That seems reasonable to me.

Number two: the CLARITY Act, even without specific ethics provisions, will add significant restrictions to anyone in crypto, including requirements for disclosures around the risks and holdings of a crypto asset. It will add lockup requirements. So it’s only an increase in the restrictions that would apply to anybody, including government officials.

And then point number three: what’s been frustrating from a political point of view is that, normally, government ethics rules are done separately, and an industry is regulated separately. So this is very exceptional. This industry is being held to a completely different standard than other industries. That said, I think that a very reasonable debate for people to have and for politicians and policymakers to figure out is what the ethics rules should be for all financial assets, including crypto.

I hope they can figure something out so we can move forward, because we want to regulate this industry.

Marc: By the way, this bill puts controls on government officials’ use of crypto that are stronger than those that apply to government officials’ stock trading.

Chris: Yeah. And one other point I would add is that right now most people think of crypto as something you can trade.

I think what they’re missing is that increasingly, like with the stablecoins we’ve been discussing, it’s a thing you can use. And — if the industry’s successful — you’ll have more and more use cases and more and more ways in which it’s embedded into our financial lives.

And so I also think it’s important with these rules to make sure you separate that kind of trading behavior from the use behavior, because I think it’s important for government officials to have access to modern technology.

Robert: Okay. So there are government ethics rules in the bill, and also it’s a separate issue from creating market structure rules for crypto.

Chris: Look, the reality of where the politics are right now is that there are ethics provisions in the current bill.

They’re still being negotiated. I hope they come to a resolution. But I think from our perspective, the important thing is that — after many years of regulatory gray area and an environment that favors, as I was describing before, a race to the bottom — we get a regulatory framework passed. And we think the CLARITY Act, while not perfect, is significantly better than the status quo.

Stablecoins and the banks already building onchain

Robert: You mentioned stablecoins again just now. And one of the big fights has been over the treatment of stablecoins in this bill — in particular the banking lobby. Most notably, JPMorgan Chase has been very vocal about this.

They are unhappy with the idea that you could earn interest on stablecoin holdings. Now, the concern is that if consumers can earn interest on stablecoins, they’re going to pull their money out of banks and keep it in these other accounts — that’s going to cause deposit flight.

Chris: In the bill, essentially, I would say the banks got mostly what they wanted, which is that you cannot pay interest on balances.

I think the exact language is, anything that’s functionally or economically similar to a bank account, and there are further legal clauses. Now if somebody did multiple transactions using their stablecoin wallet twice a month at Walmart and you use a formula that’s not functionally or economically similar to paying out on a balance, what you can do is give them rewards.

Robert: So it’s more like a credit card or rewards program.

Chris: Yeah. Going any further, you start to ban things like Starbucks rewards points and things, right? There was a prolonged discussion — I think it’s gotten as far as it can go, and essentially it was a compromise that was somewhat painful for the crypto industry. But we’re still very supportive of the overall bill.

Robert: The thing that I find so interesting is Brian Armstrong has embodied this debate versus Jamie Dimon. They’re kind of the two leading voices on this. But JPMorgan — it has a big blockchain unit. They have tokenized deposits that are live onchain.

Chris: I mean, these are massive organizations. But yes, many big banks, including JPMorgan, have significant blockchain efforts.

And I think if CLARITY passes, you’ll see that a lot of those efforts will roll out in a significant way. We’re way past the point where enterprises adopting new technology will simply dip their toes in and do experimental stuff. We speak to a lot of these organizations regularly.

Some of them are already live, and they have significant deployments that are going to go live with greater regulatory clarity. I think they see significant opportunity. You have to remember, the banking system is the way it is for a bunch of reasons. And it’s not necessarily their fault. These systems are very complex and very intertwined, but a lot of the technology is very antiquated.

And so one of the things that blockchains have done for the financial industry is give them a unified framework where they can all kind of say, “Hey, together, let’s move into the 21st century.” And so it’s solved not just a technology problem, but a coordination problem.

And I think if you speak to them, they’re all very excited about that.

Blockchains remove layers of mediation — they’re disintermediation machines. When you send money through a wire, the bank has some piece of paper that gets moved somewhere else. There is no financial “network” — like a global network the way there is a global internet or a global WhatsApp network. It’s a patchwork of systems. And so one of the beautiful things with, say, stablecoins is you’re building something just the way you would build it on the internet.

And so the banks see that opportunity. They say, “Wow, we can save a lot of money. We can provide better products. We can modernize a whole bunch of things.” And so we see that — across the board — a lot of these organizations are very excited about this, including JPMorgan.

And don’t take it from me. Goldman Sachs CEO David Solomon just came out endorsing the CLARITY Act. Fidelity, BlackRock, a bunch of other major financial organizations. They all have major efforts. Much of this is public. The most innovative fintech companies like Stripe have gone in, in a big way.

The technology is on the verge of going very mainstream with the right regulation.

A “kill shot” to open source

Robert: Let’s shift over to another debate, which is over liability for developers. There’s a former White House cybersecurity official, Carole House, who has raised concerns that developers should be held more liable for the software that they write.

She argues that, if you don’t hold them liable, this could set a dangerous precedent for other areas of technology like AI.

Marc, I know you spend a lot of your time on AI. What do you think about this argument?

Marc: It’s a kill shot to the industry.

It makes software development impossible because how can any software developer anticipate how the software will be used down the road?

And you don’t even need to think about this in terms of software. If I run a hotel and a criminal stays at the hotel and they plan their operation — does that make me part of their conspiracy? If I’m an engineer who designed a car, and the car is used in a bank robbery, does that make me an accessory to bank robbery?

If you saddle people who build products that get used by many people for many reasons with downstream liability, it just kills the industry.

Chris: And just to add to that — if you build a piece of software and you’re like, “Hey, criminals, come use it, here’s how to do a crime,” that is a crime. No one’s debating that.

What we’re talking about is you’re building open source software for a constructive use case. You’re making an AI model; you’re making a blockchain; and then you make it open source.

If you add unlimited downstream liability, criminal liability, or civil liability — how could you ever decide to do that as an open source developer with that kind of risk? You couldn’t.

These people aren’t big companies that can afford to defend themselves. Five guys in a garage are going to take unlimited liability? It’s going to kill startups. It’s going to kill open source.

Marc: This is a very live issue in AI policy also. Right out of the gate, open source dies, because open source developers are not getting paid to start with. They certainly can’t take on enormous negative liability.

Because open source dies, academic research dies. Because without open source, there is no computer science research in any of these fields. It’s entirely dependent on open source. So it kills computer science as a field, and that’s just a start.

Then it kills venture investing, right? Because we obviously can’t invest in a company if they have a million times the level of exposure, because somebody does something that they didn’t even know about. And then it kills all the companies, right? All the startups. And then it kills the big companies because they also can’t take that on. So it rips right through.

What CLARITY actually does to securities law

Robert: Let’s move on to one of the last objections that I’ve heard most often, which is that this bill would punch a hole in securities laws. Marc, as you mentioned, securities laws have allowed US capital markets to thrive for the past 90 or so years.

So some people say that if this bill gets through, a company will just go, “I’m going to tokenize an asset, and put it on a blockchain, and this is going to exempt me from SEC oversight and securities law in general.” What’s your response?

Chris: Here’s what the bill says. If you take a stock and you tokenize it, that’s a security, and it’s regulated by the SEC like a security is. It’s very black and white.

The only thing that is different is that the CLARITY Act delineates that blockchains or tokens like Bitcoin or Ethereum will, in some cases, be regulated by the SEC and, in others, by the CFTC.

Let me walk you through it. So Bitcoin, when it started, was presumably one or some set of people — whoever Satoshi was — and by definition, when something starts, it’s centralized. There’s a person behind it. They have sort of inside information and control.

And the way the CLARITY Act works is, at that point, that token is regulated by the SEC. So someone creates a new blockchain. They have control over it, and they have rules, like lockups, for example. They have disclosure rules. They have all the kinds of things you expect from securities laws.

Over time, when you hit certain thresholds of decentralization and start to look more like a Bitcoin or Ethereum of today — where there is no central actor that controls it, has inside information, and so forth — the CLARITY Act designates that the CFTC oversees that token. So it’s regulated as a commodity and not a security. Because at that point, you still need regulation to make sure that people aren’t dumping on the market or cornering the market and all the other things that commodities laws regulate.

But it’s a different regulatory regime depending on the nature of the asset.

Robert: So it becomes more like gold or precious metals or oil or wheat.

Chris: That’s right. Even the last administration said at some point that Bitcoin and Ethereum were sufficiently decentralized to be regulated as commodities. So this has been implicit in court cases and in agency decisions from both parties for the last 10 years.

All CLARITY does is enshrine it in law and make the definitions specific, so that people know exactly where they are and you don’t have to go to court to figure it out. It’s taking what has been the consensus across courts and regulators, hard-coding it, and making it specific.

Today, if I create a token, there are no rules for disclosure. There are no rules for insider trading or lockups. If the CLARITY Act passes, our effective lockup period — meaning the period until we can sell — will increase significantly. We’ll invest in someone launching a product; it’ll have a token, and until these criteria are met for hitting sufficient decentralization, there are lockups on venture capitalists and founders and so forth — which is how it should be.

It’s a smart risk-based regulatory framework, and we think it will allow people to build long-term products and build trust within the market. Consumers, investors, and so forth — all participants will have trust in the market because it’s regulated and there’s a framework around it.

That’s how you build a real industry long term.

What happens if CLARITY fails

Robert: All right. So we’ve walked through a number of the live fights. All of these things are being debated right now. Any one of them could potentially tank this bill.

What happens if CLARITY does not pass?

Chris: We’ll keep working on it.

A lot of these things can and are being done at the agency level — meaning the SEC, CFTC, Treasury, and so forth. I think the big difference is that legislation is de facto permanent in a way that agency rulemaking is less so.

And that’s important both to make sure the rules are there and to protect consumers in the long term — but also for industry to have confidence. If you want to build something, it can take many years to build. If you have to try to build on shifting sands, you’re much less likely to make that investment of time and money.

You’ve got enough to worry about as an entrepreneur already without having a shifting maze of regulatory changes. So that would be the downside. It would just prolong that. But also, I’m optimistic it will get passed in the near future. If it doesn’t, we will keep working on it.

Continuing the long tradition of American technological leadership

Robert: Let’s zoom out and talk about what’s at stake for the country and for American leadership. Passing the CLARITY Act, why is it so important for the country to do this?

Marc: We would argue this is in the long tradition of American technological leadership, and I would say American technological supremacy, which is: do you want to be the country that leads the world in technology or not?

There are two parts of that, which are: do you want the technology to exist when some people argue that you’re better off without certain technologies? The other part is that technologies do tend to exist once they get invented — do you want that to happen inside your own country, or do you want to have that happen somewhere else?

And we always make a very strong argument on this, which is that every American citizen, regardless of their political position, should want America to be the world’s technology leader. That has many benefits. It has direct economic benefits in terms of the wealth of the country and the ability to pay for all the things that we want to pay for.

It has enormous benefits in terms of our security. For the crypto industry to be based in the US is overwhelmingly good for United States security. It’s overwhelmingly good for United States law enforcement. It’s overwhelmingly good for United States national security.

We’re so used to America being the technology leader that it almost feels embarrassing to make this argument because it’s such an obvious argument. America has so clearly benefited from this for the last hundred years. And in my view, we should all want that to continue for hundreds of years to come.

Robert: Well, time will tell whether policymakers heed your words, that this is urgent, and that the benefits will accrue to whomever sets the standards first.

Thank you both.

Subscribe now

This newsletter is provided for informational purposes only, and should not be relied upon as legal, business, investment, or tax advice. Furthermore, this content is not investment advice, nor is it intended for use by any investors or prospective investors in any a16z funds. This newsletter may link to other websites or contain other information obtained from third-party sources - a16z has not independently verified nor makes any representations about the current or enduring accuracy of such information. If this content includes third-party advertisements, a16z has not reviewed such advertisements and does not endorse any advertising content or related companies contained therein. Any investments or portfolio companies mentioned, referred to, or described are not representative of all investments in vehicles managed by a16z; visit https://a16z.com/investment-list/ for a full list of investments. Other important information can be found at a16z.com/disclosures. You’re receiving this newsletter since you opted in earlier; if you would like to opt out of future newsletters you may unsubscribe immediately.

America | Tech | Opinion | Culture | Charts

Today, we’re excited to share a guest post from Daphne Koller, Founder and CEO of insitro. For more from her, subscribe to insitro’s Substack here.

The tech world has latched onto an intoxicating promise: build a superintelligence, and it will cure cancer, and every other disease as well. The logic is seductive. The human body is a system we can already read from and write to, so like other knowledge problems, a powerful enough AI should be able to solve disease.

I fully believe that AI will eventually transform human health. It is why I’ve spent close to 30 years working at the intersection of AI and biology, and the last decade in drug discovery. The need is staggering: by most counts only a quarter of diseases — and by some estimates a few percent — have an approved therapy, and most of those merely slow a disease rather than stop it. For the majority of human illness, medicine still has little to offer.

But the magic-wand promise rests on an assumption that turns out to be false: that we already understand human biology well enough for a clever enough reasoner to find the cures hidden in what we know. We don’t. Hundreds of years into modern medicine, our understanding of most human disease, and much of healthy physiology, is best captured by the parable of the blind men and the elephant; in this case, a really huge elephant. AI is undoubtedly extraordinary, but aimed at a biology we have only begun to measure and barely understand, it will mostly help us generate failures faster.

This essay is about what that actually takes. The path to real cures starts at the root of the problem: measuring biology in the right way and using AI to derive novel insights from those measurements. Below, I describe some of the most prevalent AI Magic Wand narratives, explain the key pitfalls, and offer my perspective on what we actually need to deliver on this important goal.

Three Problems, One Bottleneck

To understand where AI fits, it helps to decompose drug discovery into its three essential stages:

1. Disease-to-mechanism: Identifying a biological mechanism — a pathway, a target, a molecular interaction — where therapeutic intervention will alter the course of disease in humans.

2. Mechanism-to-drug: Creating a molecular intervention in the right therapeutic modality — a small molecule, antibody, siRNA, gene therapy — that achieves the desired mechanistic effect with acceptable safety and pharmacological properties.

3. Drug-to-patient: Designing a clinical development program that identifies the right patients and assesses the molecule’s effects — beneficial as well as adverse.

The vast majority of AI work in drug discovery has focused on stage 2. This is understandable: the origin of the AI Magic Wand exuberance is the incredible achievement of AlphaFold, a field-defining tour de force. From this starting point, we have seen an explosion of AI tools capable of designing novel proteins, small molecules, RNA therapies, and even gene therapies. Given a biological mechanism we want to hit, it seems that AI can now design a molecule to hit it faster and better than ever before.

AI will certainly generate new and better molecules at an unprecedented rate, but will it generate drugs that unlock diseases for which there is currently no meaningful treatment? There are “undruggable targets” — high-confidence mechanisms that historically we have been unable to hit. The success against KRAS, the quintessential undruggable target, shows that this journey is possible. Notably, this success emerged from decades of structural biology and medicinal chemistry, not AI; as of now, I don’t know of a single example of an AI-derived insight that has led to “drugging the undruggable.” More broadly, the biggest step functions in our ability to drug the undruggable have historically come not from better molecular design tools, but from expanding our repertoire of therapeutic modalities: first biologics, then siRNA and antisense oligonucleotides, then gene editing. Each new modality opened a class of targets that was simply inaccessible before.

Deep Phenotype, a Substack by insitro

Scaled biology, deep causality.

But an even more critical point: validated yet undruggable targets are a tiny handful in the landscape of unmet need. For the vast majority of diseases without effective treatments, we simply have no idea what the right mechanism is. More than 90% of drugs that enter clinical trials fail — a dismal statistic that has barely improved in several decades. In the large majority of cases, the molecule was engineered just fine. The mechanism it targeted was wrong. We are doing a pretty good job at manufacturing keys, but they are generally for the wrong locks. Even if AI lets us make better keys at an accelerating pace, that won’t improve our ability to identify the right locks. The real bottleneck in making a novel medicine is disease understanding: identifying a biological mechanism whose modification actually changes the course of disease in patients. That, far more than molecular design, is where drug discovery succeeds or fails.

This mechanistic understanding is a rare commodity. And because no one likes to fail in the clinic, we are seeing industry trends that are truly destructive. There are currently 38 targets that have over 50 programs against each of them — slightly better keys for those few locks where we have strong conviction. How many variants of GLP-1 do we really need? Even worse than this misallocation of capital is the disservice to patients: the number of novel targets the industry advances each year fell from ~100 in 2015 to about 30 in 2024. That collapse is the far bigger cost: the inability to help the hundreds of millions of people for whom medicine currently offers nothing.

The Data Chasm of Human Biology

The disease-understanding goal requires that we bridge the chasm between high-level clinical manifestations of disease in a patient and the granular cellular mechanisms in which a drug intervenes. This has given rise to a second manifestation of the AI Magic Wand. Large language models — with their super-human reasoning capabilities — will connect the dots across the vast published literature of human biology, reasoning their way to new mechanistic hypotheses.

This optimism makes a very strong assumption: that the scientific community has collected — or will soon collect — enough data about human biology to contain the answer, and that we just need better reasoning to extract it. The challenge is that human biology is incredibly complex, spanning multiple interconnected biological layers — DNA, protein, cells, multi-cellular environments, entire organisms. Individual components respond dynamically to even subtle changes in related components or in the environment. Moreover, biology wasn’t engineered; it is the result of billions of years of messy, stochastic evolution, which produced staggering variation — countless genes, cell types, states, and contexts, each behaving in its own way. There is too much of it, too idiosyncratic, to reason about in the abstract. You have to measure it.

How many experiments would we need? The table below is an informed attempt at a back-of-the-envelope analysis.

Even if we consider only cell biology — the layer we need to interrogate biological mechanism — the space is vast. It becomes exponentially more vast when we consider that a drug is an intervention, so we need to map not only biology as it is, but also how it would respond to a perturbation. The largest cell atlases assembled to date, now spanning hundreds of millions of cells, remain orders of magnitude too small to cover this space. Until recently they also held almost no causal, perturbational data, the measurements most critical for understanding what an intervention would do in a living system. That has begun to change: several organizations have launched the monumental effort of building a “Virtual Cell,” pairing large-scale perturbation data with AI to reduce the data collection burden. But even the largest of these efforts samples only a vanishing fraction of the possible perturbations, and does so almost entirely in a narrow range of cell lines.

Even more importantly, the virtual cell efforts — useful as they might eventually turn out to be — do not address the other half of the equation: relating biological mechanisms to human clinical outcomes. Most human disease is a systems-level dysfunction, involving a complex, temporal interaction of multiple biologies spanning diverse cell types. Understanding these processes requires systems-level measurements that are far less scalable, often requiring living organisms.

Which brings up the greatest data challenge. While some processes are conserved across all forms of life, others are far more specific. The folding of a single protein is a self-contained process, highly conserved — closer to physics than to biology; this allows protein folding models to be trained on sequences collected across thousands of species. Metabolism involves at least a dozen distinct cell types and might be conserved across mammals. Brain function and dysfunction involves dozens of distinct cellular identities; and these processes are exquisitely specialized to humans: rodents do not get Alzheimer’s disease; non-human primates do not recapitulate ALS. The diseases where we have made the least progress tend to be precisely those that are most human-specific, and therefore those for which the data is most expensive to collect, least available, and most fraught with ethical constraints.

The Right Experiments, Not All Experiments

But we don’t need a universal causal model in order to make medicines. We can instead build a targeted causal model that makes the space navigable toward a desired outcome — uncovering the biological mechanisms underlying human diseases. Our models will not cover all biologies or all diseases, but an informed selection process will still allow us to make considerable headway.

This insight motivates the thesis behind a third AI Magic Wand: swarms of AI agents operating in a closed loop with laboratory automation, relentlessly chipping away at scientific problems. They formulate hypotheses, direct robots, analyze results, and iterate. Early successes here have been beguiling — automated systems have already excelled at tasks like optimizing cell-free protein synthesis or designing antibodies against known targets.

But agentic iterated optimization relies on a fundamental attribute: agents thrive when there is a fast, accurate, and cheap scorecard to evaluate progress. If you give a sufficiently smart model an instant feedback loop, it will grind against that benchmark until it wins. This is why coding assistants and molecular design tools advanced so rapidly — the feedback is cheap, accurate, and fast. A compiler immediately verifies whether code will run. The iterative loop with a developer provides rapid feedback on intent. The closed loop is tight, cheap, and objective.

Drug development is the exact opposite. The ultimate scorecard — whether a drug actually provides therapeutic benefit to a patient — cannot be captured well by computational models or high-throughput assays. The only true ground truth is a human clinical trial. This feedback loop currently takes years, costs millions, and is strictly bound by human ethics and living biology. It is the ultimate slow feedback loop, and no amount of compute or process optimization can change this.

The agentic lab successes listed above are solving problems that look like coding: highly quantitative objectives within a constrained search space. These problems are valuable, but largely beside the point when it comes to predicting whether a drug will actually work in a patient population. You cannot solve the human translation problem by accelerating our ability to optimize the wrong objective function. Doing so will simply scale up our process for building more keys to the wrong locks.

In order to leverage the power of agentic AI and lab automation towards the goal of improving scientific discovery, we must first have an objective function that is a proxy to human clinical benefit, and yet allows for AI insights and rapid experimentation.

The Final Mile: Patient Impact

Some have argued that the most important AI unlock in drug discovery is in the third stage — drug-to-patient — taking a drug candidate through preclinical testing and clinical trials. This is the fourth AI Magic Wand: reduce the time and cost of this very expensive phase, and drug discovery becomes faster and cheaper. Sadly, if you accelerate a pipeline full of drugs aimed at the wrong mechanisms, all you get is faster failures.

At the same time, there are real opportunities for AI in clinical development: predictive toxicology models and AI-drafted regulatory filings can shorten IND-enabling work, and patient identification from electronic health records, smarter site selection, and automated data management can trim operational overhead. These compression levers offer meaningful benefits, and we should absolutely pursue those. But these gains sit almost entirely within the first two slices of work. The remaining 50% — in-life biological observation — is gated by how fast disease unfolds in a living human. Even a comprehensive AI-driven improvement across everything it can touch leaves the majority of development time and cost structurally unchanged.

There is, however, one pathway through which AI could accelerate even this irreducible clock — and it runs directly through biological mechanism. An AI-enabled, deep mechanistic understanding of a disease enables the identification of novel clinical readouts that serve three distinct purposes: selecting the patients most likely to respond, confirming that the drug is hitting its intended target, and detecting early and reliable signals that it is actually modifying disease biology. Together, these allow trials to enroll the right patients, read out faster, and catch failures earlier — changes that transcend clinical trial operations, transforming the trial design itself. This capability is inseparable from solving the disease-understanding problem; they are one and the same. Better trials, in the end, are downstream of better biology.

Fulfilling AI’s Promise

The AI Magic Wands described above have real value. Generating molecules quickly can be a significant accelerant, as is reasoning across the vast scientific literature and increasing the efficiency of the scientific process. But these tools are just that: point solutions that address important problems without altering the fundamentals of our industry.

To fulfill the promise of AI for the millions of patients lacking any meaningful treatment, we must direct our efforts toward the problem that really matters: the identification of biological mechanisms with disease-transforming clinical benefit. This is arguably the hardest problem in drug discovery, because the only conclusive test of whether we have correctly identified a novel biological mechanism is a human clinical trial. There are multiple other paths in this space with shorter timelines and clearer near-term proof points. Those paths are shorter because the problems are more tractable: the feedback loops are faster and the benchmarks are cleaner. But a shorter path to a smaller destination is still a smaller destination — process improvements for problems we already know how to solve.

For the hundreds of millions of patients for whom no meaningful medicines exist, the difference between a wrong mechanism and the right one is the difference between another devastating clinical failure and a life-altering breakthrough. That is what we need in order to truly deliver on AI’s promise for human health.

This harder but aspirational path is the one we have elected to take at insitro, and we’ll have more to say about our approach soon.

Subscribe now

This newsletter is provided for informational purposes only, and should not be relied upon as legal, business, investment, or tax advice. Furthermore, this content is not investment advice, nor is it intended for use by any investors or prospective investors in any a16z funds. This newsletter may link to other websites or contain other information obtained from third-party sources - a16z has not independently verified nor makes any representations about the current or enduring accuracy of such information. If this content includes third-party advertisements, a16z has not reviewed such advertisements and does not endorse any advertising content or related companies contained therein. Any investments or portfolio companies mentioned, referred to, or described are not representative of all investments in vehicles managed by a16z; visit https://a16z.com/investment-list/ for a full list of investments. Other important information can be found at a16z.com/disclosures. You’re receiving this newsletter since you opted in earlier; if you would like to opt out of future newsletters you may unsubscribe immediately.

America | Tech | Opinion | Culture | Charts

Moar Machines

There’s been a delicate dance in the public markets about who or what constitutes an AI winner and/or loser. To vastly oversimplify:

Oh, AI is going to render software obsolete . . . well maybe not everyone, and not yet, but eventually . . . but, in all events, all that AI disruption must be good for the business of serving AI . . . except that we really don’t like how much this all costs to build . . . but that’s at least good for the AI infra supply chain, which is where all the money is going . . . but, we’re also not sure it’s going to last, so . . .

And around and around it goes. Mr. Market doesn’t always make sense, because of course, Mr. Market doesn’t have to make sense, because there is no single Mr. Market.

That’s true so far as it goes, and we offer no view of what the future Mr. Market will think, but we can at least offer a view of the present and past.

In this case, returning to the question of “yes, demand for AI infra is huge, but will it last,” one key player in the supply chain recently said something kind of interesting. Vertiv, a major supplier of cooling and power infrastructure for data centers—and a big winner of the buildout—added a substantial $3.27B in revenue for ‘Q2, but still undershot its revenue guidance by ~$76M (and undershot the street’s higher consensus guide by ~$120M).

The company blamed the miss on “minor timing shifts, primarily due to temporary supply chain congestion and multi-phased project execution as deployments scale in size and complexity.”

That sounds reasonable enough. Still, if Vertiv is having supply-chain issues, then presumably they’re not alone. And, if that’s the case, then does “supply chain congestion” materialize elsewhere in the data? The answer is “maybe yes, maybe no” but the more striking thing is simply visualizing (again) how substantial the demand for “machines” has been.

Census collects data on manufacturing orders, and the growth for some of these data center components is a thing to behold:

Orders for Data Center and power-related machinery—HVAC, turbines, etc., and oil & gas—have nearly doubled from the previous decade levels. HVAC especially is just the hottest thing (and sits squarely in Vertiv’s wheelhouse), which makes sense, given all the heat that high energy compute generates.

In terms of Vertiv’s supply-chain issues—or at least demand running ahead of supply—there too the data tells the story. Sort of. Unfilled-orders for computer and electronic products, which includes semiconductors, underwent a step-change ~2023, and inflected steeply upwards again, quite recently:

“Unfilled-orders” is a proxy for a supply-demand imbalance (albeit an imperfect one) for the obvious reason that orders tend to go unfilled when demand exceeds supply. It doesn’t necessarily tell you, though, if there’s a supply-chain issue, or if demand is simply more than what supply can handle.

A somewhat better proxy for supply-demand v. supply-chain is the “backlog ratio,” which divides the backlog by monthly shipments. In that case, a bigger backlog balanced against even more shipments, would imply that demand is simply continuing to grow ahead of growing supply, but that supply chains are working just fine.

In this case, the backlog ratio looks pretty stable—definitely high, but stable.

At just under 6 months (i.e. the time it would take at current volumes to fill unmet orders), the backlog is very high relative to prior, but about the same as it’s been since 2024. So, no supply-chain issues there, but for comparison’s sake, the NY Fed’s Global Supply Chain Index (also on the chart) has most definitely inflected upwards, such that “supply chain pressure” is nearly two standard deviations higher than usual.

So, clearly there are some global supply chain issues, but whether those are impacting data center machinery specifically, it’s hard to say. The truth is that even if the supply chain disruptions did impact machinery imports, that could have the effect of slowing new orders, which would keep the import ratio in balance, anyway. Again, the data is inconclusive.

One last thing to look at: actual international import data. Census counts the stuff moving in and out of the country, so we can see if there’s been any change—perhaps a sudden drop in physical imports:

No dice. For the categories of imports closest to what data centers need for power and cooling, once again, the story looks pretty stable.

Other than “power conversion,” where physical units are ~23% lower than they were back in Jan. ‘25 (while prices are ~25% higher), every other category of import is about as numerous as it’s been for the past year and a half. Prices have gone up substantially, but the physical flow of goods is mostly unchanged. So, perhaps converter machinery has become snaggled in the supply chain, but for the other categories, the picture looks ‘business-as-usual.’

Of course, the counterfactual of how much higher imports might have been without a supply-chain disruption is unknowable from this data alone, so we’ll just have to take Vertiv (and the NY Fed) at their respective word.

What does it all mean? The same thing it’s meant for a while now: demand for (and imports of) AI infrastructure has been historically high, and well-ahead of what supply can timely meet. Maybe supply-chain disruptions have made it worse, maybe not—certainly dependency on Chinese imports has been a concern, as has the potential impact of a production backlog on timely data center development, but that’s been true for a while.

Bigger picture, though, if demand for AI infra has slowed, it’s certainly not showing up in this data.

It’s Never Been Easy for Entry-Level Hires in Tech, and Remote Work Has Made it Harder

We’ve pointed out previously that when it comes to the “AI Took My Job” story, the data simply does not support the claim, and if anything, points the other direction. While it’s still too early to know the true labor impact effects, for now at least, the evidence suggests that AI has been more a jobmaker than jobtaker.

That said, there has been one persistent weak spot in the job market, and that’s at the entry-level. It’s hard for young folks to get hired. There too, however, there’s some evidence that AI is making, rather than taking entry-level jobs.

But that’s all stuff that we’ve covered before. This week’s observation is that it turns out that entry-level jobs have always been a fairly small part of the tech hiring pie, and that hasn’t changed much, at all:

Not that it’s much consolation to the people struggling to get hired, but entry-level job postings in Tech are roughly ~5% of overall postings, and it’s been that way since 2019. If anything, mid-level hires are getting squeezed, while senior hires increase their share. In fact, if there’s any sector to blame for less entry-level hiring it’s healthcare, where the entry-level share is substantial, but falling since 2023.

When it comes to tech, though, a more plausible explanation for entry- or perhaps mid-level hiring malaise is that workers are working longer, and that WFH/Remote Hiring has unlocked a global talent pool for tech to hire (creating less of an impetus to hire relatively inexperienced local options).

According to data from ADP, the “long distance hire rate” jumped during the pandemic and has held steady at 26.4%, or ~30% higher than before:

Despite all the RTO mandates, hiring remotely has been a tough habit to kick.

And which sector has increased its share of “long distance arrangements” the most? It’s tech:

Again, according to ADP, Information (and Technical Services) have 45%+ of workers in some kind of long distance relationship—more than any other sector, and substantially more than before the pandemic.

That number sounds too high to believe, and maybe so, but the direction passes the smell-test. AI probably didn’t take your job, but WFH certainly may have.

AI Demand and Spending Show No Signs of Slowing

With all the chatter around open weight model competition, token pricing, and demand, it’s always instructive to take a quick peek at some actual, albeit imperfect, indicators of AI spend.

The first indicator has spread like nightmare fuel, but what it actually represents is a good deal more nuanced:

Silicon Data’s Token Cost Index has been in mostly serial decline since its May peak.

AI Demand Is Collapsing! Run for the Hills!

Well, no, not necessarily. As per SiliconData, the index measures token-spend intensity. It’s a combination of how many tokens are being consumed and at what cost. To oversimplify somewhat, if demand for high-cost tokens increases, but demand for lower-cost tokens increases even more, then the index will fall—even as overall demand increases.

Is that what is happening right now? It’s hard to say. But other data suggests that that may very well be the case (and that otherwise demand and spending are both alive and well).

As per YipitData’s analysis of OpenRouter data, the combined B2B spend on Cursor, Anthropic and OAI is increasing both in total and at the median (for the four biggest spending industries):

Unsurprisingly, software is by far the biggest spender, but business services, consumer and especially financials, are all trending up and to the right.

That financial services would be ripe for AI adoption has been conventional wisdom all along, and there’s plenty of other evidence that banks, investment management and other financials are investing heavily in AI.

Likewise, also from Silicon Data, both spot GPU rental rates, and longer-term contract pricing continue to climb:

With the exception of A100 spot-pricing (which is pretty stable), renting a GPU continues to get a little more expensive—and not just for the latest edition.

As for longer-term agreements, arguably a better indicator of demand, the term-rate curve for the H100—the second oldest GPU in the set—has moved progressively higher in recent weeks (and is much higher than it was in November of last year):

12-month pricing for the H100, according to Silicon Data, sits just under $2.50/GPUhr—almost 40% higher than it was in November of last year.

So much for GPU obsolescence.

For what it’s worth, and this data is from earlier in the week, prediction markets think the H100 is only getting pricier:

Kalshi markets put the rental rate at ~$2.78/hr, or about 5 cents higher than what it was when the data was collected.

Make of this what you will, but this data at least suggests that both demand and spending continue to cook.

Subscribe now

This newsletter is provided for informational purposes only, and should not be relied upon as legal, business, investment, or tax advice. Furthermore, this content is not investment advice, nor is it intended for use by any investors or prospective investors in any a16z funds. This newsletter may link to other websites or contain other information obtained from third-party sources - a16z has not independently verified nor makes any representations about the current or enduring accuracy of such information. If this content includes third-party advertisements, a16z has not reviewed such advertisements and does not endorse any advertising content or related companies contained therein. Any investments or portfolio companies mentioned, referred to, or described are not representative of all investments in vehicles managed by a16z; visit https://a16z.com/investment-list/ for a full list of investments. Other important information can be found at a16z.com/disclosures. You’re receiving this newsletter since you opted in earlier; if you would like to opt out of future newsletters you may unsubscribe immediately.

America | Tech | Opinion | Culture | Charts

This week we’re looking at the state of applied behavior analysis, or ABA: the intensive, one-on-one, hourly therapy that is the front-line treatment for autism.

Behavioral health is a big deal. Demand is enormous (and growing). So, why is this still such a hard business to run?

We have some data from Camber, which sheds some light on the operating challenges that specialty care clinics like those that deliver ABA care (as well as PT, ENT, SUD, and Cardiology) face. Camber is building the AI-native revenue cycle platform that these clinics use to get paid, so this is a problem they understand deeply. It turns out that small differences at the reimbursement margins can go a long way towards launching and operating a sustainable clinic (versus something other than that).

Let’s dig in.

The first thing to keep in mind is that behavioral health clinics do not suffer from lack of demand. Quite the opposite: one of the biggest challenges that clinics face is the chronic shortage of qualified clinicians and staff to meet the growing demand for one-on-one treatment.

The second biggest challenge is where Camber fits in: making sure that clinics are reimbursed by insurance for all the services rendered by the extremely busy clinicians. Getting timely payment, in-full, can be the difference between a clinic staying open or not (and therefore the difference between whether a child keeps their therapy or not). In other words, timely and adequate reimbursement is crucial for getting and staying in this business.

Now, pain around reimbursement from insurance is not unique to ABAs, but it is particularly acute. These practices are fairly young, margins are thin, and the insurance landscape is extremely complicated, with a huge amount of variance. That’s precisely where Camber excels.

Clinician Reimbursement Is Wildly Varied and Complex

Just to illustrate the point, when it comes to this sort of treatment, the same insurance company will reimburse wildly different amounts, depending on the state:

For four different insurers, the high-low variance between what any given insurer reimbursed for an identical service was anywhere from 2.09X to 1.35X.

In other words, the same insurer pays clinics more than double for the same service in one state than it does in another. The point being is that this is not a simple system to navigate, and getting it right can have enormous financial implications.

“Claim Denied” is Not the Issue, but “More Information Needed” Can Be the Difference Between Making Payroll (or Not)

Now, the most common boogeyman when it comes to insurance is claims denial, so it might come as a surprise to learn that denied claims are not actually that big a deal. The vast majority of claims are actually approved:

For the typical clinic, only ~3% of claims are denied outright. At the edge, the denial rate is quite high, but that’s simply not the experience for most clinics.

So, if it’s not claim denial that’s causing friction in the reimbursement process, then what is it?

The real pain comes by way of the claims that aren’t denied, but need to be reworked. This is an important and commonly under-appreciated point. An insurer may approve a claim, in that everyone agrees that the treatment given was right for the patient, but then the haggling will begin. Are the codes right? Is the information complete and correct? Are all the forms present? And again, every insurer haggles a little differently, and sometimes the same insurers haggle differently over the same treatment, depending on the patient.

Clinics aren’t being sloppy or lazy. The reality is that the system is very complex.

Refiling Claims is Costly, with Increasingly Diminishing Returns

Here too, the rework pain comes at the margins, but the margins are substantial:

The data show that roughly 72% of claims went through “clean,” while the insurance companies required additional work from the clinic on the remaining ~28%.

For a thin-margin business, having just under a third of receivables in paperwork limbo is a very big deal.

And the cost of resolving these issues manually is extremely high, with rapidly diminishing returns:

By some estimates, manual resolution of administrative issues runs as high as 6.3X the automagical alternative–full automation would constitute ~$21B in annual savings alone.

Plus, because claims resolution is so manual, it means that the more time spent on a single claim, the less $ is ultimately recouped:

A single resubmission cuts reimbursement by more than half. If the paperwork gets pushed back and forth 3-4 times, the share of dollars the clinic ultimately recoups is ~25% of the claim (often months after the service was actually rendered).

Again, clinics are thin-margin, resource-strapped, labor-intensive services. If 30% of that labor time gets caught in paperwork purgatory, and it requires even more (costly) labor time to get it fixed, then you can see why it’s not an exaggeration to say that better, cheaper, and faster claims processing can be the difference between life and death for an ABA clinic. Reducing that 30% “follow-up needed” rate, and reducing the cost and improving the effectiveness of that follow-up, is enormously consequential for whether children get therapy or not.

Better Billing Is A Tech-Solvable Problem, And Camber is Delivering Operational Lift Where it Matters Most

Camber’s data also shows that this is most definitely a tech-solvable problem.

If 30% of the claims are causing all the pain, then consider the most common flaws in that 30%:

Roughly 20% of denials (and 20% of the dollars) revolve around information and documentation. Coverage and credentialing issues pace second and third, respectively.

These are administrative errors, driven in large part by the complexity of the reimbursement process, and they are eminently solvable by technology, particularly technology that learns (like Camber’s). If there were simply rules-based logic to getting claims right, this would be less of an issue–certainly encoded payer-rules, and automated scrubbing and checks help, but the real hurdle is developing a feel for each insurer’s various quirks.

Part of why we know that a continuously learning technology is so critical, is that clinics themselves get better at this process, the longer that they’re in business:

After 2-years, a clinic improves its first-try success-rate from 77% all the way to 92%.

Mastering the claims reimbursement process is like a learned profession. After two-years of on-the-job training, a clinic’s back office can learn most of the arcane knowledge required for a more seamless submission. But, from a provision of care perspective, that process is expensive and time-consuming–one that many fledgling clinics may not have time or scale to afford–and it means that the departure of a single system expert presents a lost sunk cost that can set a clinic back years.

With Camber, however, the learning curve to billing mastery steepens exponentially:

In just two months after onboarding with Camber, the share of claims that require manual intervention is cut by more than half. And that says nothing of the nature of the manual work that’s actually required when Camber is involved from the outset.

Put it all together, and it’s easy to see why Camber is playing such a valuable role in the ABA ecosystem. Camber massively reduces a clinic’s operating cost of getting paid, does so much more quickly, and doesn’t require a clinician to have nightmares about whether a key member of their billing team is about to go on vacation, retire, or get poached by another clinic.

Getting paid in behavioral health is not a simple transaction. It is the output of a multi-week, multi-touch process, and how much of every dollar survives depends as much on the state and the insurer as on the clinic.

Most of the margin in specialty care lives in that gap, not in the denial rate everyone watches. As we demanded more care, the pain caused by this gap was strong but invisible. Now, because we’re bridging the gap between claim and dollar, we can see it. The demand was never the hard part. Getting paid for it was. That, finally, is changing.

Subscribe now

This newsletter is provided for informational purposes only, and should not be relied upon as legal, business, investment, or tax advice. Furthermore, this content is not investment advice, nor is it intended for use by any investors or prospective investors in any a16z funds. This newsletter may link to other websites or contain other information obtained from third-party sources - a16z has not independently verified nor makes any representations about the current or enduring accuracy of such information. If this content includes third-party advertisements, a16z has not reviewed such advertisements and does not endorse any advertising content or related companies contained therein. Any investments or portfolio companies mentioned, referred to, or described are not representative of all investments in vehicles managed by a16z; visit https://a16z.com/investment-list/ for a full list of investments. Other important information can be found at a16z.com/disclosures. You’re receiving this newsletter since you opted in earlier; if you would like to opt out of future newsletters you may unsubscribe immediately.

America | Tech | Opinion | Culture | Charts

Today, we’re excited to be publishing an excerpt from Dr. Fei-Fei Li’s recent blogpost “Building Worlds That Train Robots.”

When spatial intelligence becomes physical, the north star goal is to advance the field of robotics. Many laboratory demonstrations of robots today show promising progress, but the real challenge is making them work reliably in the real world, where objects shift, clutter accumulates, lighting changes, and physical interactions vary from one trial to the next. This gap between a compelling demo and reliable operation is a major barrier to putting robotic systems to work at scale.

Preparing robots for the real world requires extensive data collection and repeated testing on physical hardware. Each new task or environment can demand substantial manual effort, while failures are slow to identify and costly to recover from. Closing this gap requires more than better policy learning algorithms. We must develop a scalable, data-driven way to create the experience robots need for training and evaluation before and after they enter the field.

On July 21, SceniX, a robotics and simulation company, joined World Labs. SceniX has been building systems that turn real robots, environments, and interactions into simulations for policy training and evaluation, developing a real-to-sim-to-real (R2S2R) engine that turns one physical task into many controllable, reusable worlds, helping robotics teams train policy models and test changes faster, uncover failures earlier, and reduce costly experimentation on hardware.

The work by the SceniX team deepens World Labs’ technology development in spatial intelligence, and broadens our use cases from virtual to physical environments. As we continue to make progress, our generative world models will not only create realistic 3D scenes but also become increasingly aligned with the reality and variation that robots must interact with, learn from, and be evaluated in. And this is just the beginning.

Today, we share some early results from the R2S2R engine. Using our proprietary technology, our model generates simulations aligned with reality that allow robots to do what has long been considered out of reach: learn complex manipulation tasks with zero real-world training data; predict and evaluate through simulation which robotic policies will succeed or fail in the real world, without extensive and expensive physical trial and error; and then operate reliably for hours on physical robots in real-world settings.

Scaling robotic intelligence by scaling the worlds in which robots learn

Scaling robot learning is the most critical bottleneck for robots. While we have seen rapid advances in robot learning methods such as vision-language-action models (VLAs) and world-action models (WAMs), the key bottleneck is not architecture alone, but experience and evaluation at scale.

Unlike internet data used to train language models, robot experience is expensive to collect and difficult to control. Every iteration consumes hardware time: objects must be reset, failures recovered from, and systems maintained. Even internet-scale video data falls short of systematically covering the environments, objects, appearances, physical properties, robot states, robot embodiments, and failure conditions that are required for training reliable and deployable robots in the real world. The same lesson was learned in the development of autonomous driving, which has a much simpler physical embodiment and a less complex task environment compared to general robotics.

In our functional taxonomy of world models, spanning renderers, simulators, and planners, we argued that the simulator is the linchpin because it turns a world into a place where agents can act, learn, and be evaluated. The work we share here puts that argument to the test. Indeed, some of the most successful L3 or L4-level self-driving cars running on the road today are powered by models trained using both real-world driving and simulation data. Furthermore, many aircraft, rockets, drones, or quadruped locomotion systems have been developed and tested using simulation systems that respond under different conditions and actions. We believe the same is true for training robots of various embodiments that can perform a much wider range of complex tasks.

Building a simulation for each real-world task has traditionally been costly, difficult to scale, and prone to visual, geometric, and dynamic gaps with reality. Our R2S2R engine takes a different approach. Combining spatially coherent environments generated by world models with task-aligned robotic simulation, R2S2R achieves high-fidelity sim-real alignment at scale by preserving not only how the world looks, but how it acts when the robot interacts with it.

We share our results through two parts of the R2S2R engine:

• Real-to-Sim: Transform physical robots, sensors, environments, objects, and interactions into simulations that preserve task-relevant observations and dynamics

• Sim-to-Real: Train and test robotic policies in simulated worlds and determine whether their performance in simulation predicts their performance in reality

The key result here is the closing of the loop between real-to-sim (R2S) and sim-to-real (S2R). For the first time, robot learning policies trained entirely in simulation transfer directly across diverse robots and tasks, while simulated evaluation predicts relative policy performance and failure regions on hardware. Together, these capabilities make R2S2R a repeatable engine for training and evaluation, with faster and cheaper iteration.

Read the rest of the post here

Subscribe now

This newsletter is provided for informational purposes only, and should not be relied upon as legal, business, investment, or tax advice. Furthermore, this content is not investment advice, nor is it intended for use by any investors or prospective investors in any a16z funds. This newsletter may link to other websites or contain other information obtained from third-party sources - a16z has not independently verified nor makes any representations about the current or enduring accuracy of such information. If this content includes third-party advertisements, a16z has not reviewed such advertisements and does not endorse any advertising content or related companies contained therein. Any investments or portfolio companies mentioned, referred to, or described are not representative of all investments in vehicles managed by a16z; visit https://a16z.com/investment-list/ for a full list of investments. Other important information can be found at a16z.com/disclosures. You’re receiving this newsletter since you opted in earlier; if you would like to opt out of future newsletters you may unsubscribe immediately.

America | Tech | Opinion | Culture | Charts

A billion machines will become autonomous or intelligent over the next ten years. Cars, trucks, tractors, mining haulers, defense systems, warehouse robots, humanoids—the physical economy will be rebuilt around software that perceives, decides, and acts.

The prevailing assumption about how we get there goes something like this: models keep improving, world models mature, foundation models for robotics arrive, and autonomy falls out the other end. Intelligence is the whole game; scale the intelligence and the machines will follow.

I’ve spent a decade helping Applied Intuition build the software infrastructure behind many of the world’s most ambitious physical AI programs, from software-defined vehicles and autonomous trucks to construction equipment, mining systems, defense platforms, and robotics. That vantage point has given me a front-row seat to where the industry is accelerating—and where it continues to slow itself down. I’m as bullish on the intelligence as anyone, but the prevailing assumption gets the math wrong. Deployed physical AI is a product of two variables: the capability of the models, and the capacity of the engineering system around them—i.e. how requirements become software, how software gets validated, and how validated systems get deployed, monitored, and improved. The industry has largely poured everything into the first variable while the second sits roughly where it was a decade ago, built for quarterly releases and hundred-person integration teams. Frontier intelligence running on a legacy engineering system doesn’t produce frontier outcomes, because the old engineering system is the limiting factor.

The contrarian bet, then, isn’t against intelligence. It’s that the next order of magnitude in physical AI comes from making the engineering system as intelligent as the models it carries. The industry’s roadmap, however, rests on a handful of assumptions that once made sense but no longer match where physical AI is headed.

Smarter Models Don’t Create Deployed Machines

The gap between a capable model and a certified, operating machine is enormous, and model quality alone doesn’t close it. A model that’s 20% better in benchmark terms still has to be integrated with many other software components, tested across millions of scenario variations, traced against safety requirements, validated on hardware, rolled out to a fleet, and monitored in the field. In a typical program, that pipeline — not the model — sets the tempo. Teams take delivery of a meaningfully better model and then spend two quarters proving it’s safe to ship.

This is why world models, as remarkable as they are, won’t get us to a fully autonomous future on their own. They are advancing faster than engineering organizations can keep up. Every leap in model capability relocates the bottleneck rather than eliminating it. The constraint moves downstream, from “can the machine perceive the world?” to “can we validate, integrate, and operate what the machine can now do?” A team whose validation cycle takes months is, in effect, throttling frontier AI down to the speed of its own process.

Here’s the implication the industry hasn’t priced in: as models commoditize toward the frontier, two companies with access to the same intelligence will have wildly different outcomes. The difference will be determined by how fast their engineering systems can absorb what the models can do. Intelligence is becoming ubiquitous. The ability to operationalize it isn’t.

Digital AI Doesn’t Transfer to Physical AI

The second assumption is subtler: that the agentic revolution happening in digital work will naturally extend to physical AI. Direct the coding agents and copilots at the autonomy stack, and the same productivity gains will follow.

They won’t, because most digital AI stops at documents, conversations, and code. Physical AI work doesn’t live there. It lives in drive logs and sensor data, in simulation runs and hardware-in-the-loop test rigs, and in requirements databases and validation reports. It lives in fleet telemetry streaming from real vehicles on real roads and job sites. An agent that has never seen a disengagement, doesn’t know why a perception regression matters, and can’t trace a requirement to a test case, is not a productivity tool in this domain. It’s a liability with a friendly interface.

Making agents genuinely capable in deploying physical AI itself is a frontier intelligence problem, one that is entirely different than training a bigger model. The agents need access to the actual data and tools of the trade—simulators, data pipelines, validation systems—through interfaces hardened enough to trust. They need embedded domain judgment and the accumulated knowledge of what “validated” actually means when the artifact ships into a multi-ton machine. And they need evaluation and governance built in, because in this domain a plausible answer and a correct answer can be separated by fatal consequences.

When I say physical AI needs its own agentic platform, I’m talking about a platform that combines state-of-the-art models grounded in the data layer, tooling, and domain expertise of physical systems, with evaluation and governance native to the platform. It is not a chatbot bolted onto engineering tools, and it is not something you get by fine-tuning a general-purpose agent. It’s a different architecture, and it demands as much AI innovation as the models themselves.

Speed Doesn’t Compromise Safety

When discussing agentic capabilities in physical AI development, the reflexive objection is that agents have no place in safety-critical engineering. Automation means moving fast and perhaps getting a few things wrong. It’s a reasonable position when the thing in question weighs several tons.

However, in safety-critical systems, the speed of your feedback loop is a safety mechanism. When validation takes weeks, teams test at irregular milestones. When it takes minutes, they test on every change. Problems surface earlier, when they’re cheap to fix. Coverage expands to orders of magnitude more scenarios. Requirements get implemented more accurately and verified more often. The slow, careful-looking process isn’t by its nature the safest one. It’s often the one where defects age quietly for months before anyone notices.

What actually makes agents safe in this domain isn’t slowing them down. It’s drawing the line correctly. Automate development, validation, and operations workflows, so safety-critical issues get resolved faster, and refuse to automate certification, regulatory sign-off, and final engineering judgment. High-stakes agents propose; humans decide. Anything touching production systems runs behind approval gates. This isn’t a temporary concession while the models improve. It’s the correct permanent architecture for physical AI, in the same way that a well-designed autonomous vehicle has a defined operational domain rather than unlimited authority.

When agents help build physical AI faster, that speed can make the end product safer.

Models Don’t Compound. Systems Do.

When you understand these assumptions are holding physical AI back, the logical next step is combining better intelligence with faster learning, creating an agentic flywheel where frontier models and frontier engineering systems feed each other.

A continuously turning flywheel looks like this. A machine underperforms in the field. Agents mine the operational data to find out where and why, or synthetically generate the scenarios that expose the gap. Findings become requirements, and requirements become test cases. Test cases become validated software, which is deployed to the fleet. The fleet generates new data, which makes the models better, which makes the agents better, which speeds up the next turn of the loop. Every turn used to take months and a room full of specialists. Each stage that agents accelerate doesn’t just save time; it increases the number of turns, and each turn compounds both the speed and the intelligence. Smarter models turn the loop faster. A faster loop makes the models smarter. That’s the compounding effect the industry is leaving on the table when it treats intelligence as the whole game.

We know this flywheel is real because we’ve been running it on ourselves. Applied Intuition has spent a decade at the frontier of physical intelligence — perception, simulation, validation, vehicle software across automotive, trucking, mining, agriculture, and defense. Over the past year we built an agentic platform grounded in that infrastructure. It’s called Dana, and our engineers have built more than a thousand internal apps and agents on it. With Dana, development cycles have become roughly 20x faster, with higher output quality. Deployments went from once every few weeks to multiple times a day. Applications that took months to build now take days or hours. And, tellingly, applications that would never have justified months of effort now get built regularly. When the cost of building drops by an order of magnitude, the set of things worth building expands by more than an order of magnitude.

The intelligence made the engineering system possible; the engineering system made the intelligence matter. Neither alone gets you there.

What This Means for the Next Decade

If the industry keeps betting on intelligence alone, the next decade of physical AI looks like a slow one: dazzling demos, decade-long programs, and a widening gap between what machines can do in a lab and what’s actually operating in the world. The models will be extraordinary and the deployment curve will stay stubbornly flat, because every improvement will queue up behind engineering organizations that absorb change at last year’s speeds.

Pair frontier intelligence with agentic engineering systems and the curve bends. Physical AI starts reaching its full potential. Farms that stabilize output through labor and climate shocks. Mines with continuous, safer extraction. Freight networks that self-route around disruption. Defense systems that hold under degraded conditions. Multi-hundred-billion-dollar markets converging on the same stack, with the learning loop at the center of all of them.

Software ate the world by making it cheap to build applications for the digital economy. Physical AI will do the same for the physical one, but only if building intelligent machines becomes as fast and iterative as building software, without compromising the discipline safety-critical systems demand. That takes the best models and a reinvention of how we engineer, and the second half is the one almost nobody is building.

Twenty years from now, we won’t remember which company had the best world model in 2027. We’ll remember which company figured out how to continuously turn intelligence into deployed systems. That’s the problem we’ve been working on.

Peter Ludwig is the co-founder and CTO of Applied Intuition, a leader in physical AI and the company behind Dana, an agentic platform for building intelligent machines. Since co-founding the company in 2017, he has helped grow Applied Intuition into one of the world’s leading AI infrastructure companies, serving the automotive, defense, trucking, construction, mining, and robotics industries.

Subscribe now

This newsletter is provided for informational purposes only, and should not be relied upon as legal, business, investment, or tax advice. Furthermore, this content is not investment advice, nor is it intended for use by any investors or prospective investors in any a16z funds. This newsletter may link to other websites or contain other information obtained from third-party sources - a16z has not independently verified nor makes any representations about the current or enduring accuracy of such information. If this content includes third-party advertisements, a16z has not reviewed such advertisements and does not endorse any advertising content or related companies contained therein. Any investments or portfolio companies mentioned, referred to, or described are not representative of all investments in vehicles managed by a16z; visit https://a16z.com/investment-list/ for a full list of investments. Other important information can be found at a16z.com/disclosures. You’re receiving this newsletter since you opted in earlier; if you would like to opt out of future newsletters you may unsubscribe immediately.

America | Tech | Opinion | Culture | Charts

That logo you’re chasing is costing you your market.

Founders with great AI products are spending months courting their first Fortune 100 customers, burning through funding rounds, and tying up their teams on deals that will never convert. They chase these logos not because the deals are big, but because a marquee name on a sales deck supposedly makes every deal after it easier. So founders focus all their attention on them, and often even offer steep discounts (or pay customers to use them!) to land them. The founders don’t care that the revenue barely registers. The logo is the whole point. Meanwhile the buyers who actually need the software, and would pay full price for it, have never heard the company’s name.

It’s not that there’s no logic to this. AI is new, so founders assume they have to educate the market, and they default to the Lighthouse strategy: win a few marquee customers, build social proof, and reassure the buyer who’s afraid of making the wrong call. But for a lot of AI companies that instinct is exactly backwards. Their buyers already understand the problem and aren’t afraid of getting it wrong. They just need the math to work, and every week spent chasing prestigious logos is a week a competitor spends selling in Ohio. Those companies would be better off pursuing the Landgrab strategy: win on math, move fast, and sign the largest number of customers possible, logo be damned.

There are two GTM playbooks for AI companies selling into the enterprise: the Lighthouse and the Landgrab. The difference between them isn’t about product quality or team strength. It’s about what you’re selling and who you’re selling to.

The Lighthouse

Category creation is when AI enables work that couldn’t be done before. This kind of solution has no precedent inside the buyer’s organization, no incumbent to displace, and no existing mental model in the buyer’s head. You’re not replacing Salesforce but inventing something new, and that often means you’re asking buyers to take a leap. Social proof is everything so you need lighthouse customers whose adoption signals that this category is real and you’re the one to bet on.

Harvey built AI for legal professionals, a category that now feels entrenched but only a few years ago had no players, let alone customers. Law firms were buying research tools from Thomson Reuters and LexisNexis, which surface information for an associate to interpret. Meanwhile, Harvey does all that work itself: drafting, research, and due diligence across thousands of documents. Lawyers are trained to avoid risk, so no firm wanted to go first, but when Allen & Overy signed in late 2022, followed by Paul Weiss in early 2023, every peer took notice. Today Harvey has hundreds of millions in ARR and an $11 billion valuation, but it first needed A&O and Paul Weiss to tell the legal industry AI-powered legal work was real.

Hebbia ran the same playbook in financial services, building an AI intelligence platform for firms whose teams spend 60+ hour weeks poring over high-stakes data rooms. In this world of highly confidential deals and guarded reputations, no fund wanted to go first either. Hebbia broke through with the world’s largest private equity firms, hedge funds, and consultancies as early customers and then expanded to more than 40% of the largest asset managers by AUM, including KKR and BlackRock. Just as with Harvey, marquee names went first and the market followed.

Capturing lighthouses is a small-team, founder-led, and high-touch effort. Applied in the right markets, it lands large deals: six-figure ACVs at least, often seven figures early on. Sales cycles can run three to six months (or longer) because you’re navigating POCs, custom work, and buyers wary of costly mistakes. The team that closes the deal is often the same team that delivers the product, which is expensive and unscalable by design. Going first can come with outsized rewards for buyers, so the best lighthouse sellers make the leap feel like a chance to win big instead of a chance to be wrong.

The Landgrab

The game flips entirely when the buyer already knows the problem and a mistake won’t cost them their job. They understand what you’re selling, and the pitch is simple: “I replace Y at a lower cost or with a better outcome.” You don’t need Stripe’s CTO vouching for you to get a meeting. You get the meeting by showing a VP of Support their current spend and saying “we cut this in half.” Social proof still helps you close faster here, but the market already believes in what you’re selling.

In these markets, sacrificing speed is deadly because founders are not only competing with other startups, but also with incumbents building AI. Sure, Zendesk adding an AI copilot is fundamentally different from Decagon replacing the entire support function with agents. But the incumbent already owns the customer, and the more AI it adds each quarter (whether it’s make or buy), the harder it becomes to pull that customer away. That is why speed is the whole game. In the words of our colleague Alex Rampell: you need to get distribution before the incumbent gets innovation. The way you do that is with a landgrab.

Take Stuut, which automates accounts receivable: collections, payments, cash application, and dispute resolution. SAP and Oracle ship AR modules in their ERP suites, and players like HighRadius have sold point solutions since the early 2000s. Yet teams still lose countless hours chasing invoices. Stuut finally solves this: customers increase cash flow by 40%, cut manual tasks by 70%, and lower their collection period by 37%. The company went wide early, leaning into the lower middle market over Fortune 100 logos, and now serves manufacturers, distributors, and logistics companies across Michigan, Ohio, Texas, and beyond, deploying in under a week against the 6 to 18 months of a traditional rollout.

Decagon used a similar motion to win in customer support. The founders ran roughly a hundred customer conversations in a month before building their product, then sold on rapid deployment and immediate ROI, scaling from 0 to 8 figures in ARR in 18 months. In 2025 alone, the company signed more than 100 new enterprise customers across travel, finance, health, and retail, tripling its valuation to $4.5 billion in under six months.

Landgrab selling is demo-driven and relies on a larger team. The product needs to be standardized enough that a customer can onboard fast and see value quickly. Implementation is driven by forward-deployed teams specialized in delivery, rather than discovery. The unit economics have to work at volume because volume is the whole strategy.

Picking Your Strategy

Enterprise sales is fundamentally about risk and reward. Behind every deal is a person who has to sign their name to the decision, and who wants the same thing all of us want: for the thing they approved to work, and to still have their job next year. Instead of evaluating your product in the abstract, they’re deciding how much personal exposure it creates, and what evidence would make that exposure bearable.

That’s why two questions help you draw the map and decide which strategy you should pick.

How exposed is the buyer who signs? Not every mistake costs the same. In customer support or AR automation, a faulty reply or misstated invoice annoys a customer and gets fixed - the buyer’s downside is a bad quarter, not a bad career. A buyer’s exposure climbs with three things: whether the industry is regulated enough that a vendor mistake becomes the buyer’s compliance problem, whether you’re replacing a system of record or adding a tool alongside one, and whether the output faces the outside world (a filed document or customer-facing answer vs. an internal draft someone reviews). In law and financial services all three run hot, and one fabricated figure can misprice a position or sink a deal. For those buyers, the ROI math is beside the point. They’re managing personal downside no discount can offset.

Does social proof travel? In some markets, reputation carries well, while in others it does not. In financial services and law, firms watch each other obsessively, status is legible, and landing two marquee firms moves the whole market: whoever went first has effectively done the risk assessment for everyone behind them. In AR at mid-market firms, the controller in Des Moines does not care that a household-name brand uses your product and is never going to hear about it anyway. Buyers there don’t watch each other closely, so each sale starts from zero and a marquee logo buys you less. Concentrated, status-driven markets carry proof while fragmented ones make you earn every deal on the math.

Put these two questions together and you have the map. When exposure is high and proof travels, you’re in lighthouse territory: a few credible buyers going first unlock the whole market, the way it played out in legal work and financial research. When mistakes are recoverable and proof travels less, you’re in landgrab territory: math closes the deal, and coverage wins the market, which is what is happening in customer support and AR automation. The other two corners matter less to enterprise sellers but are worth naming. When proof travels but isn’t required, you may not initially need a large sales team; the product spreads itself, for instance engineer to engineer, the way dev tools do, and category creation happens bottom-up. Finally, when the buyer needs proof but the logos never reach them, you’re in a hard market and you likely haven’t heard of the companies stuck there.

You’ll notice other patterns that seem to predict the motion: SMB skews landgrab, risk-averse industries skew lighthouse, additive tools move faster than system-of-record replacements. Trace any of these back far enough and you land on exposure or proof. The controller at a 200-person distributor is a landgrab buyer because her market is fragmented and her deal is small. The industry code on the account never entered into it.

Markets don’t always fall cleanly into one bucket, and the two questions above lead to the clearest conclusion at the extremes. For the ambiguous cases, a few other factors are also worth weighing. For instance, an existing budget and napkin ROI usually confirm you’re in landgrab territory, but only after the exposure question clears. A buyer can have the budget, see the math, and still refuse to move until someone credible goes first. When the two point in opposite directions, exposure wins every time.

Sales cycles can provide another gut check to identify which territory you’re in. If they run longer than 60 days, if you’re doing custom work to prove the concept, if the buyer asks “is this safe?” before “what does it cost?”, your buyers need proof, and you need to run the lighthouse. The more common mistake right now is the opposite: assuming buyers need proof just because the technology is AI and therefore feels new. But the controller whose worst case is a misstated invoice isn’t managing career risk. Show up with a marquee logo instead of a number and you’ve answered a question they never asked.

The Traps

Lighthouse Traps

Becoming a hostage to the logo. Everyone wants the same marquee logos, and you end up in a brutal fight for the same 500 accounts that every other AI startup is pitching while the logos extract concessions because they know you’re desperate. Lighthouse customers are a means to an end. Get a few trusted names, then run the table. The vast majority of revenue sits in companies no one’s heard of.

Prestige without payback. The wrong customer won’t collaborate on repeatable software (you become a consulting shop), won’t pay recurring (unsustainable economics), or won’t pay high enough ACVs (the math breaks). Worst case: a prestigious logo that teaches you nothing replicable and underpays, and you’ve traded time for a vanity metric.

Pilot purgatory. Big logos love pilots. Six-month POCs that never convert, burning your best people on deals that were never real. The fix is time-boxed pilots with clear milestones and auto-converting contracts.

A lighthouse for one. You over-rotate on your lighthouse customer’s requests and build a product perfect for them and useless for everyone else. You win the lighthouse but no other ships follow the light in.

Landgrab Traps

Dying of indigestion. When you can sell to anyone the discipline is saying no. Some customers are poison: hard to onboard, hard to drive results for, low ACV. Without qualification discipline and a deal desk, you wake up with 200 customers and 50 underwater.

Grabbing land you can’t hold. You’re selling at volume, and if you scale coverage before the product is ready you create detractors at scale. 50 unhappy customers means churn. 500 is a reputation problem.

Mistaking the valley for the market. Canvassing every bus stop in a major metro isn’t grabbing land. The real opportunity is figuring out how to show your ROI to the 50,000 companies outside of normal networks.

Sequencing: Lighthouse to Landgrab

The best companies don’t stay in one mode forever. They sequence from lighthouse to landgrab deliberately, getting a bellwether in one vertical, dominating that vertical, and then finding adjacent verticals that look similar. At Affirm (a story worth its own post), the breakthrough was Casper. Once they had one mattress company they got every mattress company, and then they went to exercise equipment, and then to things that look like exercise equipment but aren’t. Mattresses and Pelotons have nothing in common except that they’re big-ticket items people want to pay for over time, and Affirm saw that before the market did.

Eventually, new categories turn into recognized ones because the lighthouse customers define them. But you have to earn the transition. Going wide before the category is established burns cash and credibility. The signal that you’ve earned it is buyers approaching you with allocated budgets and asking for a demo rather than asking who went first. When that happens, the lighthouse worked, and it’s time to run the landgrab.

Closing

Every AI founder believes they’re inventing the future. And they are! But your buyer doesn’t purchase the future; they purchase either proof or math. If they need proof, go win the logo that gives it to them. If they need math, get on a plane and show it to them before your competitor or the incumbent does.

The founders who get this wrong won’t fail because they built the wrong product or picked the wrong strategy off a menu. They’ll fail because they never asked which game they were in, and in a market moving this fast, you only get to ask it once.

Thank you to , , , Spencer Wiedeman, Frank Golden, and for their thoughts on this post.

Subscribe now

This newsletter is provided for informational purposes only, and should not be relied upon as legal, business, investment, or tax advice. Furthermore, this content is not investment advice, nor is it intended for use by any investors or prospective investors in any a16z funds. This newsletter may link to other websites or contain other information obtained from third-party sources - a16z has not independently verified nor makes any representations about the current or enduring accuracy of such information. If this content includes third-party advertisements, a16z has not reviewed such advertisements and does not endorse any advertising content or related companies contained therein. Any investments or portfolio companies mentioned, referred to, or described are not representative of all investments in vehicles managed by a16z; visit https://a16z.com/investment-list/ for a full list of investments. Other important information can be found at a16z.com/disclosures. You’re receiving this newsletter since you opted in earlier; if you would like to opt out of future newsletters you may unsubscribe immediately.