America | Tech | Opinion | Culture | Charts

Welcome to a special Vacation Week issue of chart deep dives, where we go into a specific topic with data from a portfolio company. Today, we have a fun issue, thanks to our friends at Deel who shared some data with us on… how people take vacation.

All of the data below are from companies using Deel, so they are going to skew towards startups, tech companies, and remote-first firms (relative to average population data, anyway.) So, rather than “this is how everyone takes vacation”, this is going to more accurately represent, “how people in our industry (and other white collar vocations) take vacation around the world”, which frankly is just as interesting a comparison if not more.

Unsurprisingly, there’s a wide range of how much vacation days are offered to the median employee, and an equally wide range of, “do people actually take all of the vacation they’re owed?” North America clusters more stingily than Europe does, but there’s still a lot of variation within regions, particularly among people who claim 100% of the days off that they’re offered:

The one caveat we can think of here, which would be hard to untangle with the data we have here, is, “Do norms vary between countries on how much people formally take the vacation days off, as opposed to informally (and not officially docking the days with HR)?”

Over the course of the year, the Germans hold the OOO crown. But not consistently! Swedes and Italians trade places in July / August. The British are (proportionately) at their desks in peak summer; the Dutch over Christmas.

True to form, European out of office messages are doing a lot of work over the course of the summer.

And also the rest of the year, while we’re at it.

Interestingly, there’s a wide range of how long people take their summer breaks - and, while Europe clusters differently from the Americas on the usual pattern, Brazil breaks the trend. There’s a reason for this, it turns out! Brazil has some interesting laws about how employees take vacations. Brazilians get 30 days of vacation allowances (including bank holidays), but they have to take those holidays in blocks; a maximum of 3 per year, apparently. You learn something every day!

Still, most “vacations” are a day off, not a long holiday.

Another regional variation: vacation days in India are booked last minute. Whether that reflects a genuine last minute trip decision, or simply last minute disclosure to HR, we can’t say.

Long weekends seem to have special allure in the summertime. It makes sense that vacation days skew towards Mondays and Fridays, and also that the effect gets magnified in the warmer months. Still, this is a smaller discrepancy than we would’ve guessed!

A popular alternative, the “Summer Friday”, is more officially popular abroad than it is here. Unofficially, though, is another question. Our guess is that Americans are doing it off the books, rather than formally taking the half-days off.

Looking at last year as an example, it’s no surprise that the single most common day to take off is a “buy one get one free” situation.

Birthdays are also a popular day to take off, but more so for big ones. Turning 31? Not worth celebrating, apparently.

Also, Sweden, what’s going on here?

And then there’s the dirty little secret of vacation days… sick days. Which, of course, are only used for recovering from illness; we would never imply anything otherwise.

Of course, we’ve spent all this time on summer holidays, but the real empty office weeks are at Christmastime.

The War on Christmas never won the enterprise, apparently.

So, no matter how hard our Swedish, French and German friends work to make those vacation days count, summer break is still getting mogged by Christmas break. (Everywhere but… Armenia?)

Wherever and however you’re taking any time off this summer, enjoy it!

Have a great summer, from the folks at Deel and a16z. If you want more charts like these, you can subscribe to Deel’s substack here:

Subscribe to Deel

Aside from something at the end of the week to commemorate America’s 250th, we’ll be off until July 13th. See you then.

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

Moby Dick, by Herman Melville

Topic: Obsession, Revenge, Seafaring, Whale Anatomy, How to negotiate startup equity, American Dynamism

Length: 625 pages

Where to read it: On the beach, listening to crashing waves

Pair with: July 4th celebrations

Why you should read it: It’s the Great American Novel.

Really why: It’s so fun to read this book. It’s one of the great startup books of all time, and the whole point of reading it is for the obsessively detailed, tactile knowledge. (For example, you’ll read a few pages on how to coil a rope properly, so that there isn’t any latent tension in it that could violently unkink and throw someone overboard!) Also, this is one of those books that’s so referenced in popular culture (the “White Whale” as a metaphor for obsession, for example) that you’re basically required to read it once.

Best quote:

Reread rating: Probably one and done, but you must do it once

Anything else: Your reward for making it most of the way through the book is getting to meet one of the great bit characters in American literature, Captain Boomer

The Magic Mountain, by Thomas Mann

Topic: The passage of time, the meaning of the Western Canon, what you will do when you face Death

Length: 706 pages

Where to read it: At altitude

Pair with: A good vintage of Gruaud Larose

Why: The single most helpful book I’ve read for sorting out the Western Canon of Ideas in your head. Won the Nobel Prize for Literature

Really why: I enjoyed this book so much I gave an Interintellect salon on it once. The premise of the book is, this guy Hans Castorp goes to visit his brother in a sanatorium in Davos, and ends up staying there for 7 years, with all of these people who are caricatures of all of the European countries and their idiosyncrasies and problems on the eve of WW1. In particular, there are two Italian guys (named Settembrini and Naphta) who, over the course of the book, argue their way through the entire Western Canon, and the reader gets an incredible birds eye view as they fight over every big timeless issue. You will finish this book as a more educated person than when you started.

Quote:

Hot take: You’ll think about current American electoral politics in a new light after reading this.

Reread rating: Reread once a decade

Anything else: Get the John E. Woods translation

The Count of Monte Cristo, by Alexandre Dumas

Topic: Adventure, Historical Fiction, Revenge

Length: ~1,276 pages (or ~52 hours) AKA, not for the faint of heart.

Where to read it: On a long plane ride or in short stints

Why: I know no one will believe me (see: page count) but if you want adventure this summer, it is a shockingly fun and suspenseful read that keeps you turning the page!

Really why: “As for you, you meant evil against me, but God meant it for good.” (Genesis 50:20) This verse summarizes the whole of the book to me and has fundamentally changed my view of life. The story of Edmond Dantes is one of a man on the cusp of gaining everything, wrongfully accused, thrown in prison, and thus losing everything. When he emerges, he completely remakes himself to enact his revenge plot. Throughout the novel - and Dantes’ life - there is profound loss and despair, yet the persistent providential hand that seems to turn each evil thing into something unexpectedly good and beautiful. I would give this to anyone lost in life, unmotivated, despairing, or simply stuck. The intelligence and pure dedication with which Dantes remakes himself is inspiring and slightly terrifying. The providential hand that somehow redeems the darkest of situations is even more so.

Best quote:

Hot take: Unlike every movie or TV adaption, I would not change the ending at all.

Reread rating: Realistically, every few years because of the length. But I was ready to reread it as soon as I finished it.

Anything else: Adaptations are often a good watch in their own right, but don’t quite capture the heart of the story.

Expanded Cinema, by Gene Youngblood

Topic: Media theory

Length: 444 pages

Where to read it: Tompkins Square Park

Pair with: A screening of 2001: A Space Odyssey or the shorts of Stan Brakhage, Andy Warhol, Carolee Schneemann, etc.

Why: In this book, Youngblood asks what happens to human consciousness when images, networks, computers, video, and environments become part of how we think? There are a couple books I return to when trying to understand how new kinds of technologies create new kinds of people (Democracy in America by Alexis de Tocqueville is another example of this kind of book) and this is probably the main one.

Really why: I picked this up from one of those outdoor used booksellers on the corner of 7th street and 2nd ave in the East Village, sometime around 2018/2019. At the time, I wanted to make movies. At the time, I was also working as an equities analyst at an investment fund in Manhattan. Obviously some asymmetry there. Anyway, this book really trained me to look at art/new mediums as training regimes that get people to see the world differently. Youngblood was writing about narrative cinema, television, and experimental film (and beyond just the works themselves…the underlying mechanical and newly-created digital processes as well). But I think a lot of his observations can be applied to any new technology that introduces humanity to new forms of collective experience. When I encountered the book, I was thinking a lot about crypto and insurgent trading apps like Robinhood. Now, AI is what I think about when I revisit the book.

Best quote:

Hot take: Better than anything Marshall McLuhan wrote.

Reread rating: Read at the steepest point in the S-curve.

Anything else: You might be tempted to skip over the longer sections where he recounts the plots of experimental films you probably haven’t seen. Resist the urge! Find the movies online and watch them.

The Boys in the Boat, by Daniel James Brown

Topic: Narrative nonfiction / sports history. Nine working-class kids, an Olympic dream, and 1936 Berlin.

Length: ~400 pages (~14-hour audiobook)

Where to read it: On a dock with your feet in the water.

Pair with: A workout. Chaser book: Into Thin Air, John Krakauer.

Why: Not every summer has an Olympics but we can always use Olympic stories.

Really why: We’ve lost the amateur era of most things. The Boys in the Boat is its time capsule. Nine working-class kids with no scholarships, no sponsors, and no audience rowed themselves to Berlin on the eve of one of history’s darkest periods. The book also hands you a word, “swing,” for the moment a crew syncs so completely the boat goes weightless, doing the unglamorous work that actually moves the shell. I’d read it again just for George Pocock, the book’s quiet philosopher, and his reminder that you cannot fake the work because the water always knows. The fastest crew is the one where no one needs the credit. For anyone building a team, or anyone who has started measuring their life in metrics and forgotten why they began.

Best quote:

Hot take: George Pocock the boatbuilder is the best character in the book. The 2023 movie missed most of what makes the book great.

Reread rating: Every few years, with Pocock’s epigraphs on annual rotation.

Anything else: Audiobook is top tier. It was one of Edward Herrmann’s last recordings before he died, and his pacing is unhurried in exactly the right way.

Predictably Irrational, by Dan Ariely

Topic: Behavioral economics. Our weird little human operating systems.

Length: 384 pages (7.5-hour audiobook)

Where to read it: Anywhere you’d people watch (so you can watch passersby be irrational).

Pair with: A large coffee you bought because it didn’t seem that much more expensive than the medium. Chaser book: Thinking, Fast and Slow, by Daniel Kahneman.

Why: Our irrationality is consistent enough to predict. Once you see the patterns, you cannot unsee them in your own choices.

Really why: You’ll catch yourself reaching for the middle option on a pricing page, or rearranging the fridge so the junk food hides behind the fruit. You’ll watch the word ‘free’ talk you into something you never wanted, the first number in a negotiation quietly set the whole range, and a pricier glass of wine somehow taste better. You’ll keep three tabs open long after one would do, guard a mediocre chair you happen to own like an heirloom, and stop short of offering a host cash for a great dinner because some part of you knows it would offend. Once you start spotting these, they show up everywhere: in your inbox, your calendar, your spending, and most of all in the tidy stories you tell yourself about why you chose what you chose.

Best quote:

Hot take: It doesn’t really matter that a bunch of the experiments didn’t replicate. We can all have our N of 1 experiences to prove them “true”.

Reread rating: Let it stare at you from your book shelf to remind you of your oddities.

Anything else: Read the endnotes and listen at 1.25x.

Range: Why Generalists Triumph in a Specialized World, by David Epstein

Topic: The case for breadth and late specialization.

Length: 352 pages (10.5-hour audiobook)

Where to read it: The flight home.

Pair with: Hobbymaxxing and side quests. Chaser book: Outliers.

Why: People who sample widely and specialize late outperform the ones who picked a single lane at age six. You want 10,000 hours but maybe not all of one thing.

Really why: You learn something from everything you do, and almost none of it is wasted. In AI terms, range makes your personal context window bigger: a specialist works from a narrow prompt, while a generalist retrieves across everything they have ever done. The best solutions are often analogies from someone’s unrelated hobby. Let the detours you once apologized for become the reason you can connect two things nobody else can. When life’s rules of engagement keep moving, range is the hedge and curiosity is its own form of strategy rather than a distraction. For anyone who feels too scattered to be good at one thing, or anyone catching themself trying to optimize the wandering and wondering out of their life.

Best quote:

Hot take: The worship of early specialists is mostly recall and survivorship bias. The most durable people have unrelated skills that keep colliding into each other.

Reread rating: Every few years, or any summer you start to feel behind.

Anything else: Short on time? The first chapter and the “match quality” section are 80/20 the rest of the book. Natural follow-up: Epstein’s earlier book, The Sports Gene.

Dinner with Joseph Johnson by Daisy Hay

Topic: Biography / cultural history

Length: 536 pages

Where to read it: When you find yourself wondering how Ben Franklin or William Blake would have thought about AI and technology.

Pair with: Henry Fuseli’s The Nightmare — the painting hung over Johnson’s actual dining table.

Why: Before founder dinners, VC fellowships, and curated salons, there was Johnson’s closed circle guest list dinner table, where the most formative thinkers argued the next century into being. If you’re in SF because you believe great things happen by getting the right people together in rooms, this is the 18th century parallel to that.

Really why: The case for Dinner with Joseph Johnson is the case for scenius — Brian Eno’s word for the fact that genius usually isn’t a single person, but a scene. Once a week in late-18th-century London, above a bookshop in St. Paul’s Churchyard, Joseph Johnson set a famously bad table and an iconic guest list: Mary Wollstonecraft, William Blake, Joseph Priestley, Benjamin Franklin, Wordsworth, Coleridge. He wasn’t the smartest person in the room, but as a publisher, he was a convener who found these people, fed them, put their work into print, and kept it in circulation when the government was actively trying to shut them up (he went to prison for it). What came out of that table and the thinkers he “backed” included the founding text of modern feminism, the radical liberalism that fed the next century of politics, the late-Enlightenment science tipping into Romanticism. These happened because the right people were in the right city at the right moment, arguing at the same dinner. If you live in SF and believe the real edge of being here are the people and conversations you get early access to — this is the original version of that.

Best quote:

Hot take: The “lone genius” is a myth invented mostly by people who weren’t invited to the dinner. Every Great Man had a Joseph Johnson/early mentor and believer — we just don’t write books about the hosts.

Reread rating: One and done, but pass it on to the next generation.

Anything else: Starts slow and front-loads with names, dates, history, politics - dense but necessary context for setting the scene of the 1790s. Don’t try to memorize all the ensemble members at first mention – trust that the people you need to know will resurface. The relationships between the characters become richer the longer you read.

Becoming Steve Jobs, by Brent Schlender & Rick Tetzeli

Topic: Biography

Length: 480 pages

Where to read it: Wherever you are when you feel like giving up.

Pair with: One Too Many Mornings by Bob Dylan (Jobs’s favorite song)

Why: Must-read for founders. It’s the story of how adversity, far more than genius, is what made Steve Jobs great.

Really why: Becoming Steve Jobs is a must-read for founders. The book centers on the 12 “wilderness years” between Jobs’s firing from Apple and his triumphant return. For the better part of a decade, both NeXT and Pixar were failing while Jobs poured in more and more of his own money. It looked like he would squander the whole fortune from his first stint at Apple and refused to quit. Eventually Toy Story, the first fully computer-animated feature film, made Pixar’s 1995 IPO one of the most successful of the decade, and Apple acquired NeXT, whose operating system became the foundation of macOS. Those wilderness years turned out to be the most pivotal of Jobs’s life, forging him into the leader capable of taking Apple to unimaginable heights. Adversity, far more than genius, is what made him great.

Best quote:

Hot take: Better than Isaacson’s book

Reread rating: Every few years (or whenever it feels like your company is going to die)

The Venerable Bede, Historia Ecclesiastica Gentis Anglorum

Topic: Early Medieval Poasting

Length: ~400 pages

Where to read it: Northumbrian Monastery

Pair with: Book-water, a potent anti-venom–good for snake bites.

Why: An Anglo-Saxon de Tocqueville, Bede undertook a saints and miracles beat, with colorful, rigorous and detailed accounts of the mystical and folkloric that illuminated key features of the nascent Gentis Anglorum (while never wasting an opportunity to dunk on the Britons, or other godless heathen).

Really why: The Dark Ages were pretty dark—full of war lords, plague, and slaughter. Still, the Venerable Bede was a genuinely curious observer, devout theologian, and capable polymath (and sometimes poet) who described (in vivid and colorful detail) a unified English People (the “Gens Anglorum”) before the English People would recognize themselves, as such.

Best quote:

Hot take: Contrary to contemporary revisionism, the “Dark Ages” is a totally valid descriptor of the early Medieval period.

Reread rating: Read once in college, and once more recently.

Anything else: Easy to skip around? The Venerable Bede wrote chronologically (which was something of an innovation), but you don’t have to read it chronologically.

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 few weeks ago, we announced the a16z Design Engineer Fellowship—an 8-week cohort of AI-native design leaders at the forefront of how software will be designed and built in the future.

We received thousands of applications from talented people around the world. Today, we’re excited to share more about the 73 fellows we’ve selected to join the inaugural cohort who are leading design at companies like OpenAI, Cursor, Stripe, Ramp, Vercel, Pinterest, and Partiful.

Many fellows wrote the original playbooks in product design like - Joshua To from Meta, Ian Silber from OpenAI (prev. Instagram), and Yuliya Gorlovetsky from Stripe.

Some are currently building category defining products, zero to scale - like Dann Petty at QuiverAI, Ryo Lu at Cursor and Nicholas Garro at Vercel.

Others are design leaders at tech powerhouses like - Ammaar Reshi at Google, Maheen Sohail at Meta, and David Hoang at Atlassian.

Across the board, these fellows are the ones writing and rewriting how great products get made, building design systems that scale, and raising the standard for what a design team can ship.

We hope that you’ll join us in celebrating everything this group has already done, and everything they’re about to do next.

Congratulations and welcome, a16z Design Engineer Fellows! To stay in the loop on applications for the next cohort, fill out this form.

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

Cycles, different but the same

If you compare this cycle to the previous one, in some ways it’s exactly the same, and in other ways, it’s a complete inversion.

It’s the same because tech is plainly the cycle-winner for both the post-GFC period (2010-20) and the post-pandemic period (2020-present). The rest of the field, however, has flipped—the last cycle’s winners are now losers, and vice versa:

• Healthcare, consumer products and media all offered double-digit returns post-GFC, but are now doing ~3-6%,

• while energy, raw materials, construction and financials all went from low single-digits to mid-to-high double-digits.

The laggards became the leaders, while the leaders became laggards.

Tech is the exception, as a repeat cycle-winner, but there’s some nuance. Hardware is the real standout of the current cycle (and did pretty well, last time too), but software has followed the broader pattern of inversion.

Taking a step back, there’s a pretty obvious pattern here that we’ve alluded to before: markets have shifted their attention from capital-light, consumer-oriented sectors, and now cast their loving eyes on the capital-heavy “real” economy, driven in large part by the AI infra buildout.

It’s a rotation from bits to atoms.

“Asset-heavy” companies have flipped-the-field, after a decade of trailing behind the “asset-light” variety.

Of course, if this cycle is like the last cycle, then the general idea is that all of this capital-heavy infrastructure eventually flows through to the software/app-layer. Back in the post-GFC era, chipmakers (and cloud-builders) dominated early, but eventually gave way to all the apps, marketplaces, and enterprise softwares that flourished on the chip-powered phones and PCs, and server-powered clouds. In other words, the rotation to atoms was temporary and cyclical, rather than a more enduring structural shift.

That certainly could happen this time around, as well—in fact, it would probably be pretty disappointing if the AI infra buildout did not eventually flow through to the capital-light layer (and, of course, both could eventually thrive in tandem). But, even so, further upstream of the public markets, there’s some indication that the atoms-revolution may have some endurance of its own—and, it’s not strictly-speaking an AI infrastructure thing, either.

The premium to “real world” tech is showing up in the private markets, not just in AI infrastructure, but in robotics, as well:

Measured by the value of top 100 private companies (by category), Robotics (and physical AI) wasn’t even on the chart in 2016, but a decade later, it has now eclipsed fintech and payments as the second largest category.

If you follow where the VC dollars are flowing, the surge in interest in robotics shows up there, too:

According to Pitchbook, Q1 was a record for Robotics and Physical AI by both deal value and deal count, with ~$16B invested across just under 500 deals.

For context, the rush of robotics investment is ~2x larger by count and ~4.5x larger by value than the preceding stretch from ‘21-’25.

The point being that the rotation to atoms (in the private markets, at least) isn’t just about the chips and inference that will ultimately power the next generation of software—hardware, as a product in its own right, is on the rise.

And it makes sense too. Even better software has enormous potential, but robotics pushes tech into a set of real-world “tasks” that software by itself cannot touch. To the extent that AI is an unlock for the software that powers hardware, it expands the surface area for demand, in unprecedented ways—not unlike the way that electricity ultimately enabled machines to do work that humanity had barely contemplated (before the “horsepower” became so readily available).

For now, the highest profile new frontier for robotics has been defense—and it, of course, helps that defense budgets are expanding the world over. But if things go as planned, the rotation to asset-heavy may be deeper, wider and longer than any previous modern tech-cycles.

AI enabled or AI “enabled”?

In the early innings of the LLM wave, management consultants were considered a likely AI-winner, at least in the near term. The logic was pretty straightforward: companies will want to use AI, and they will hire consultants to figure out how. Accenture, especially, was considered well-positioned because not only could it provide advice and a roadmap, it could do the actual end-to-end implementation, as well, aka “managed services.”

Without getting into the specifics of why, it’s nonetheless true that whatever optimism the market may have had for Accenture, it appears to have evaporated:

Accenture’s free cashflow multiple climbed as high as 30x in early 2025, but it has come all the way down to ~6x, or about a third of its longer-term average.

Decide for yourselves why precisely the market soured on Accenture so quickly, but one thing that’s become increasingly clear is that when it comes to the broader undertaking of “adopting AI,” there’s more to the story than simply adopting AI. Not all implementations are value-generative in the same way, and getting it right (or more right) may involve some subtlety in the development and ideation phase, at least according to some recent research.

In a study involving 515 “high growth” startups, researchers zeroed-in what (in their view) it means to be truly “AI native.” More specifically, they wanted to know what it takes to move from “AI improves a task,” to “AI improves the firm,” and the results are pretty striking.

It turns out that the key is what the researchers call a “mapping” problem.

When firms in the study were given information about how other firms have reorganized production around AI (“treatment firms”), it kicked off a fundamentally different sort of discovery process. Rather than simply trying to replicate existing processes “but this time with AI,” treatment firms began further upstream, wrapping AI into the business outcome, leading to different processes entirely.

The researchers offered product-development, as an example:

In this case, AI didn’t replicate preexisting steps in the process—it rewired the process around what AI was capable of, albeit towards the same basic business outcome.

That’s, of course, just one example, but across the board, the productive impact on “treatment firms,” was substantial. Treatment firms had:

• ~44% more use cases for AI:

• ~2x the revenue for the top vigintile (and 50% more at the top decile):

• ~40% less capital consumed (with the gap widening at the furthest ends of the distribution):

To summarize, when high growth startups really undertook the task of “adopting AI,” they found more use-cases, generated more revenue, and consumed less capital than those that didn’t.

It’s a fairly striking result that should both allay some fears about the “AI ROI question,” but also shed some light on why that ROI hasn’t (yet) materialized at the firm level, at least as widely as some would like. According to the researchers, the implications are (a) the productivity lift from AI at the firm-level is, in fact, transformative; but (b) the real unlock is at the discovery phase, i.e. “discovering where and how to deploy AI is a key bottleneck in realizing the gains,” and that’s not as simple as just “adopting AI.”

In that sense, having a “discovery bottleneck” means that AI is following a path not unlike previous tech-driven productivity leaps forward.

When electrification first took hold, for example, many manufacturers simply replaced the steam engine with a large electric motor while preserving the old architecture of overhead line shafts and belts. Factories were basically the same, but “this time, with motor.” The larger gains, however, only started to arrive when manufacturers realized they could install smaller motors at each individual machine (and scrap the whole shafts n’ belt rigging almost entirely)—factories were eventually completely redesigned around embedded power (rather than the other way around). The rest, of course, was one of the greatest productivity leaps forward in history.

AI startups running lean

On the subject of AI, startups, and academic research, we’ve got another one from the same researchers: it turns out that AI startups, do in fact, run lean—at least according to this study of four years’ worth of YC batches.

The researchers looked at YC Batches W20-F24 (with first rounds closed between 2020-2024) and linked them to Revelio employment data by employee count, function, and seniority. They wanted to know if AI startups hired and/or organized themselves differently than non-AI startups.

What they found was:

• AI startups start smaller and run smaller:

  

• The distribution of smaller (by headcount) startups skews heavily in favor of AI startups:

  

• AI startups tend to be less hierarchical, with AI startups dominating the largest share of little-to-no-hierarchy firms:

  

The implications speak for themselves, although there is likely more devil in the details, but you get the idea: if your premise is that AI will enable firms to do more with less, then there’s some more fuel for your fire.

Elsewhere, Stripe Economics has also weighed-in (again), on the “solopreneur” phase of AI enablement. With lots of caveats around how “solopreneurs” were identified in their data, Stripe offers some more support for the idea that, yes, AI is unlocking more entrepreneurship and firm-formation (albeit of the low-hiring variety), and that solopreneurs are having a good deal of success.

Consider the share of solopreneurs by income threshold:

Not only is the share of $100K+ solopreneurs rising, but the share of $5M+ and $10M+ earners really began to inflect upwards in ‘23 and ‘24.

In the words of Stripe Economics:

We find that there has been a substantial increase in the number of solopreneurs earning over $100,000 in our index, but an even larger increase in the number earning at higher income thresholds, with a clear acceleration since 2023. More than twice as many solopreneurs earned over $1 million in 2025 than in 2023, and close to three times as many crossed $5 million and $10 million.

Perhaps even more interestingly, the share of solopreneurs earning above these income thresholds has also doubled in the last two years, suggesting that—rather than the surge in business applications reflecting low-quality experimentation with a few lucky standouts—the cohorts of new solopreneur businesses might actually be of higher quality than in the past.

Again, with all the caveats around how solopreneurs were identified (in this case, by recourse to Stripe’s solopreneur-centric tools) and/or how these businesses may have evolved headcount over time (unbeknownst to Stripe), the data suggests that the AI-powered small business era continues to rise.

Tech makes grocery more productive, but also less “productive,” at the same time

a16z Charts happens to have a thing for grocery-related charts, and if it ain’t broke, don’t fix it, so here it goes.

A funny thing about grocery stores is that, unlike the broader category of retail trade, grocery did not get all that much more productive over the past 30 years:

Or rather, while retail has seen more or less steady productivity growth since 1990, grocery stores first declined, then had some recovery, then flat-lined, and then started getting more productive again, notwithstanding the recent swoon—but still, nothing like the retail productivity rager.

It’s interesting because it’s partly a story about technology (and its relationship to productivity), and it’s partly a story about how we measure productivity, which is more or less output divided by labor-hours (and that’s an imperfect measure, at best).

You see, with grocery (and retail), the great invention (after the cash register) was electric scanners. They were first introduced sometime in the 70s, but by the 90s they were basically everywhere. Scanners did two things, primarily: (1) they allowed for a massive increase in the range of inventory; and (2) they contributed to the increasingly sophisticated data-collection efforts by retailers and grocers to both understand what their customers wanted to buy, and also how much inventory to have on-hand.

In the 90s, both grocery stores and retailers began to get much bigger, benefiting from tech-driven economies of scale that were good for consumers, but more or less spelled the end of Ma and Pa.

At that point, however, the fates of retailers and grocers began to diverge. While retailers vastly expanded their inventory, they did so without adding all that many new workers, instead focusing more on off-the-shelf and prepackaged goods that could be stocked and monitored with far fewer people than before. Grocers, on the other hand, decided to expand beyond groceries into specialty services, like florists, bakers, deli counters and the like.

Of course, as the share of specialty services grew, so too did the demand for specialty labor. And, as per the chart above, even though grocery stores got more productive, in the sense of vastly expanding the range of goods and services offered and doing so at relatively lower prices, they did not get more “productive,” in the sense of output/labor-hours. That’s also why retail “productivity” ran way ahead of grocery “productivity,” while wages for both grew at basically the same pace (which would be odd, if retail workers were actually that much more productive than grocery workers).

It wasn’t until groceries took a page out of the broader retail and general merchandise playbook that grocery productivity started improving again:

Around 2000, the share of non-food-at-home products began to grow substantially—higher margin prepackaged foods, snacks and general merchandise increased almost 5x in a decade. The other thing that grocery did is that it offloaded more of the work of stocking, displays, etc. to the vendors, as something like an in-kind charge for the privilege of shelf-space. It’s a neat “productivity” hack when the labor hours don’t go away, but they become someone else’s problem.

From a “productivity” standpoint, the shift unlocked more throughput without more labor-hours, and voila, a productivity renaissance for grocery.

While labor’s share of grocery income steadily rose until ~2002 (and labor’s share of retail income fell), they have both been falling fairly steadily since, at least until recently.

The declining “labor share of income” is basically the flipside of “productivity”—more output with fewer workers will lead to a declining labor share of income (not taking into account the appreciation in 401ks due to all that profitability, of course).

What’s interesting, though (and back to tech and productivity), is that the latest wave in shopping innovation—ecomm and delivery—seems to coincide with diverging “productivity” fates for grocery and retail, yet again. While ecomm has been a boon to retailers, who can now sell things without leasing a square foot of store, edelivery probably just means the same or more people wandering through the grocery store, picking out items. Curbside pickup might even be moderately more labor-intensive than regular shopping.

Whether that’s causal or coincidence, what is true is that, post-pandemic, grocery productivity has gone south again (and the labor share of income has started rising again), while retail gets leaner and meaner. Same technology, same productivity lift, and yet, very different “productivity” stories.

Good thing for grocery, though, is that there’s always (high-margin) money in the advertising banana stand.

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

William Whittenbury is an astronautical engineer who supported space missions including the Artemis Program and the James Webb Telescope.

On April 10, 2026, four astronauts splashed down in the Pacific Ocean off San Diego, completing the first crewed lunar voyage in 54 years. The flight took nearly 10 days. Rebuilding NASA’s ability to fly it and return to the moon took five decades. Why it took this long is, in many ways, as fascinating as the mission itself.

NASA’s Artemis II mission is just the beginning of many planned missions for the program, which seeks to establish a sustainable, long-term human presence on our nearest neighbor for science, exploration, resource development, and training for deep space missions to Mars and beyond. The mission was a crewed test flight that looped around the Moon in a great figure-eight trajectory and brought them home. This mission did not land on the lunar surface, but along the way, it made history, reaching the farthest from Earth that any human being has ever traveled.

Why Did It Take 54 Years?

One of the most common questions about the Artemis program is why NASA hasn’t returned to the moon since the iconic Apollo missions of the 1960s. The short answer is funding and politics. The longer one starts with how much more NASA wanted to accomplish on the Moon – and beyond.

At the height of the Apollo Program, NASA had grand plans for continued large-scale missions under the Apollo Applications program, which sought to use the Apollo spacecraft, Lunar Module, and Saturn V rocket for a variety of expanded activities. One element of Apollo Applications that did come to fruition was Skylab, America’s first space station, which hosted crews in 1973 and 1974. A larger version of Apollo Applications was the Space Transportation System Integrated Program Plan, which envisioned a low Earth orbit (LEO) space station, a reusable space shuttle, a permanent lunar base, reusable nuclear-powered space freighters, and crewed missions to Mars by the 1980s. Unfortunately, as public attention drifted to Vietnam and unrest at home following the tremendous success of Apollo 11, the Nixon Administration shut down the Saturn V production line and funded just one piece of that grand plan — the Space Shuttle.

NASA hoped that the Shuttle would become a “space truck” that would facilitate the other elements of the plan later, but budgetary realities and military involvement compromised the design. The Shuttle included the winged orbiter, which housed the crew, cargo, and the three main engines; a disposable external fuel tank, which was dropped during each mission and needed to be replaced; and two solid rocket boosters that provided the majority of thrust at liftoff and would be recovered after parachuting into the ocean.

The Shuttle broke new barriers as the world’s first reusable spacecraft and kept America in the human spaceflight business from 1981 to 2011. It opened access to a much larger and more diverse cadre of astronauts and inspired millions. This was an amazing service record but wasn’t without tragedy, notably including the Challenger and Columbia disasters in 1986 and 2003. The Shuttle also never lived up to its promises of rapid turnaround and low-cost access to space, with each launch costing over a billion dollars by the time it retired. The Shuttle’s signature achievement was providing the heavy lift and crew launch capability needed to assemble the International Space Station (ISS), which has maintained a continuous human presence in orbit since 2000 but is now nearing the end of its operational life. However, it couldn’t leave low Earth orbit, leaving NASA incapable of transporting people back to the Moon or beyond for thirty years.

The Shuttle and ISS absorbed a huge portion of NASA’s limited budget, leaving few resources to develop next-generation technologies or think about going back to the Moon. This began to change in 2004 when the Bush Administration launched Project Constellation to develop a new family of launch vehicles called Ares out of repurposed Shuttle hardware repackaged to enable moon missions. Two vehicles in the family were slated. Ares I was a “single-stick” rocket with one solid rocket booster derived from those on the Shuttle and a new upper stage. The much larger Ares V would provide super-heavy cargo lift capacity. A new crew capsule, Orion, which looked very similar to the Apollo spacecraft but was larger and incorporated more modern technology, was designed for the program. Ares I was intended to launch Orion to either LEO or to a rendezvous with an Earth Departure Stage (EDS) launched by Ares V, which would take the capsule to the moon. A new lander called Altair would be a larger version of the Apollo Lunar Module. This was a different architecture from Apollo, which was able to launch the entire Apollo spacecraft, including the Lunar Module, on a single Saturn V rocket. The overall Ares architecture was intended to replace the Shuttle in its capacity as the crew vehicle for the ISS in LEO as well as enable missions to the Moon, and eventually Mars (Ares was named for the Greek version of Mars, the god of war in classical mythology).

However, Constellation soon ran into trouble. The Ares I-X test vehicle first flew in 2009, but the design experienced serious issues with pogo oscillation – a phenomenon that carried health risks for the crew and affected the propulsion system of the rocket. Because the gigantic Ares V rocket used a different tank diameter than the Shuttle’s external tank, it would require new tooling and production facilities. By 2010, the program was behind schedule and seriously over budget, at a time when the government’s spending was constrained by the global financial crisis. Following the recommendations of a panel of industry experts called the Augustine Commission, the Obama Administration canceled the Constellation program in 2010. This left NASA in a precarious position — not only was the moon back out of reach, but there was now no replacement for the Space Shuttle, which was set to retire in 2011 once the International Space Station’s assembly process was finished.

What followed was a humbling decade for American human spaceflight. From the Shuttle’s retirement in 2011 until 2020, the United States had no way to launch its own astronauts into orbit. Instead, NASA bought seats aboard Russian Soyuz spacecraft, paying Roscosmos a fare that climbed past $90 million per seat by the end of the arrangement. For the country that had won the original Space Race, it was a galling place to be.

Against this backdrop, Congress intervened in a bipartisan effort to save America’s space launch capabilities. The Orion capsule program, which was making good progress, was salvaged from the ashes of the Constellation project, and NASA was ordered by the US Senate to begin developing a new heavy-lift launch vehicle drawing on the lessons of the Ares program but without the limitations discovered during the project. This new vehicle was imaginatively named the ‘Space Launch System,’ or SLS. The hardware, though, was only half the problem. Each cancelled program also drained NASA of hard-won engineering expertise — the kind of institutional muscle that takes decades to build and only years to lose. Recently, that erosion is what NASA Administrator Jared Isaacman targeted when, at the a16z American Dynamism Summit in March 2026, he announced NASA Force, a program to recruit top aerospace, software, and systems talent from industry into the agency on roughly two-year terms.

The Hardware of Artemis II: SLS and Orion

SLS was designed to make maximum use of existing Shuttle hardware and production facilities, to cut costs, reduce certification time (certifying hardware for human spaceflight is time-consuming and expensive), and keep people employed on the critical Space Shuttle production lines. SLS’s design incorporated two Shuttle solid rocket boosters, but with an extra segment included in the middle (boosters are modular to facilitate rail transport from the production facility in Utah), four RS-25 engines identical to the Space Shuttle main engines, and a core stage based on the Shuttle External Tank, with the same diameter (allowing common tooling) and the same iconic orange thermal protection foam. The upper stage of the rocket, called the Interim Cryogenic Propulsion Stage (ICPS), was derived from the upper stage of the Delta IV Heavy rocket then in use by NASA and the US Air Force.

A large coalition of legacy aerospace companies was assembled to produce the full stack: Boeing is the prime contractor for the core stage, Northrop Grumman produces the solid rocket boosters (following its acquisition of Orbital ATK in 2018), Aerojet Rocketdyne (now part of L3Harris) produces the main engines, United Launch Alliance (a joint venture of Boeing and Lockheed Martin) produces the ICPS, Airbus Defense and Space produces the service module, and Lockheed Martin produces the Orion spacecraft. The Orion capsule itself carries four crew and offers roughly 60% more habitable volume than the Apollo command module, paired with a European-built service module that supplies power, propulsion, and life support, putting an international partnership into the hardware itself. All told, more than 1,100 contracting companies, spread across all 50 states, have a hand in the program. That distribution is on purpose, as spreading the work nationwide builds the broad congressional constituency that has kept SLS funded through multiple administrations — a fact of the program’s policies as much as its engineering.

Standing at 322 feet tall, SLS is taller than the Statue of Liberty, although it is slightly shorter than the Apollo-era Saturn V. Combined, the four main engines and two solid rocket boosters produce 8.8 million pounds of thrust at liftoff, making SLS the most powerful operational rocket in history (the Soviet N1 and SpaceX’s Starship have/had higher thrust ratings, but the N1 never successfully launched and Starship has yet to enter operational service).

SLS development proceeded at a relatively slow pace due to NASA’s constrained budget environment. NASA has received public criticism for the cost and delays associated with the program – each launch carries a price tag of roughly $4 billion, and development ran to roughly $27 billion. Set against the Saturn V, the numbers look different. Adjusted for inflation, the SLS cost less than half as much to develop as the Saturn V while operating under tighter annual budgets. Currently, NASA’s whole annual budget (encompassing Artemis as well as all other projects) accounts for about 0.3% of federal spending, contrasting to 4% of the federal budget in 1965, the peak year for Apollo development.

Orion flew its first test flight, called Exploration Test Flight-1, in 2014 on a Delta IV Heavy booster. The Space Shuttle flew its last flight in 2011, with the remaining orbiters retired to museums in Los Angeles, New York City, Washington, DC, and Cape Canaveral. Under NASA’s new scheme, Orion would no longer be responsible for crew transport to the ISS; instead, the Commercial Crew program was created to issue development contracts to private companies to develop new vehicles for LEO. This yielded the SpaceX Crew Dragon, which entered service in 2020 and is now the mainstay of America’s human spaceflight fleet, and the Boeing Starliner, which has faced a torturous development with many issues, most notably the time it “stranded” astronauts Butch Wilmore and Sunita Williams on the International Space Station for 286 days from 2024-2025.

After years of delays, the first Space Launch System core stage was rolled out of the Michoud Assembly Facility (MAF) in New Orleans in 2020 and began a long series of tests called the Green Run campaign at NASA’s Stennis Space Center in Mississippi. After the successful completion of these tests, the stage was transported by barge to Kennedy Space Center in Florida, where it was mated with the solid rocket boosters, the upper stage, and the Orion spacecraft in the giant Vehicle Assembly Building (VAB), one of the world’s largest buildings. After several fits and starts – including a few scrubbed launches and two hurricanes – the SLS finally roared to life on the Artemis I mission on November 16, 2022. Artemis I was an uncrewed test of the Orion spacecraft that sent it into a distant retrograde orbit around the Moon. This was the first time a spacecraft designed to carry humans had left orbit in 50 years. SLS performed its mission flawlessly, unusual for the debut of a new launch vehicle, and a payoff for years of careful design and testing by NASA personnel and the 1,100-plus companies that built it. Orion also performed well, although examination of its heat shield following reentry revealed concerning divots, forcing a full review of the heat shield’s design and the vehicle’s reentry trajectory.

The Crew of Artemis II

Following the success of Artemis I, NASA began preparations for Artemis II, the first crewed flight of the program. A diverse crew was selected from NASA’s cadre of astronauts.

Commander Reid Wiseman served as a fighter pilot for the US Navy in test and combat roles in Iraq and Afghanistan before joining NASA. He flew to space on ISS Expedition 40/41 and served as the chief of the Astronaut Office. Pilot Victor Glover served in the US Navy in Iraq before joining NASA in 2013, serving in ISS Expedition 64/65. He was the first person of color to travel to the moon. Mission Specialist Christina Hammock Koch worked in physics and other science roles with NASA Goddard, Johns Hopkins Applied Physics Laboratory, the US Antarctic Program, and NOAA before joining NASA in 2013. She completed the longest-duration spaceflight by a woman (328 days) on ISS Expedition 59/60/61, and was the first woman to travel to the moon. Mission specialist Jeremy Hansen served in the Royal Canadian Air Force as a CF-18 fighter pilot and served as a NEEMO aquanaut. He was the first non-US person to travel to the moon, on his first spaceflight.

NASA astronauts traditionally choose the names of their spacecraft; this crew decided to name the Artemis II Orion Integrity.

The Mission and Its Moments

Artemis II ultimately faced two years of delays from its initial projected launch date as NASA sought to address the heat shield concerns from Artemis I, which were ultimately addressed via a heat shield redesign for future missions and an adjusted trajectory for Artemis II. Artemis II launched on April 1, 2026 on the second SLS rocket, this one painted with special “America 250” livery to recognize the country’s sestercentennial. SLS performed so well, with over 99% orbital insertion accuracy, that two planned burns in the mission were not necessary. After initial checkouts in Earth orbit, including a simulated rendezvous maneuver, Integrity conducted the trans-lunar injection burn, becoming the first crewed spacecraft to leave Earth orbit in 54 years.

A photo of Earth captured by the crew during trans-lunar injection. The planet is backlit, with the sun behind it, rendering the atmosphere and auroras visible. (NASA)

Artemis II’s overall mission profile was very similar to Apollo 8: a circumlunar flight to test out the spacecraft with humans aboard, essentially in a “figure 8” around the moon. However, the similarities didn’t end there. Apollo 8 famously flew at the end of 1968, during one of the most tumultuous years in modern American history – including the Vietnam War abroad, high-profile assassinations, and unrest at home. On Christmas Eve, its crew beamed back a message and a view of the Earth that briefly seemed to steady a reeling world. A private citizen later sent the astronauts a telegram that simply said: “You just saved 1968.”

Artemis II carried an uncanny echo of that moment. It, too, flew during a major holiday – Easter – and into a world unsettled by war in the Middle East and a highly polarized political environment at home. On Easter morning, Mission Pilot Victor Glover shared the following message:

“I think these observances are important, as we are so far from Earth and looking back at the beauty of creation. I think for me, one of the really important personal perspectives that I have up here is I can really see Earth as one thing.

And you know, when I read the Bible, and I look at all of the amazing things that were done for us, who were created, it’s…you have this amazing place, this spaceship. You guys are talking to us because we’re in a spaceship really far from Earth. But you’re on a spaceship called Earth that was created to give us a place to live in the universe, in the cosmos.

Maybe the distance we are from you makes you think what we’re doing is special. But we’re the same distance from you, and I’m trying to tell you—just trust me—you are special. In all of this emptiness—this is a whole bunch of nothing, this thing we call the universe—you have this oasis, this beautiful place that we get to exist together.

I think as we go into Easter Sunday, thinking about all the cultures all around the world, whether you celebrate it or not, whether you believe in God or not, this is an opportunity for us to remember where we are, who we are, and that we are the same thing, and that we got to get through this together.”

On April 6, Integrity officially broke Apollo 13’s record for the farthest distance traveled from Earth. The crew maxed out at 252,756 miles from Earth, far enough out to see features on the far side of the Moon that no human eye had ever taken in before. Shortly after breaking the record, the crew shared a highly emotional moment when they named a crater “Carroll” after Commander Reid Wiseman’s late wife, an NICU nurse who passed away in 2020 after a courageous battle with cancer. The crater, located near the boundary between near and far sides of the moon, is visible from Earth as a bright spot at certain times. Their two teenage daughters, who gave Reid a friendship bracelet for the mission, watched from Mission Control.

Adding even more emotion, NASA played a surprise recording of a message left by legendary Apollo 8 and 13 Commander Jim Lovell for the crew shortly before he passed away last year:

“Hello, Artemis II! This is Apollo astronaut Jim Lovell. Welcome to my old neighborhood! When Frank Borman, Bill Anders, and I orbited the moon on Apollo 8, we got humanity’s first up-close look at the Moon and got a view of the home planet that inspired and united people around the world. I’m proud to pass that torch on to you – as you swing around the Moon and lay the groundwork for missions to Mars…for the benefit of all. It’s a historic day, and I know how busy you’ll be. But don’t forget to enjoy the view. So, Reid, Victor, Christina, and Jeremy, and all the great teams supporting you – good luck and Godspeed from all of us here on the good Earth.”

After passing around the back side of the moon, Artemis II transmitted some amazing images of Earth and the moon eclipsing the sun.

Public response to the mission was overwhelmingly positive. Applications by kids to space camp doubled, and commentators cheered the crew for their clear demonstration of competent leadership, caring emotional support, and international cooperation. Artemis II truly gathered the best of us and put it on display for the world to see.

Artemis II splashed down in the Pacific Ocean on April 10, 2026 and was picked up by the San Antonio-class landing platform dock USS John P. Murtha.

What Comes Next?

Artemis II was just the beginning; a test flight. NASA’s next challenge is actually landing people on the moon for the first time since 1972. SpaceX and Blue Origin have been contracted to develop two very different lunar lander designs under the auspices of the Human Landing System (HLS) program. SpaceX’s giant vehicle is based on its Starship launch vehicle/spacecraft, while Blue Origin’s Blue Moon is a much more traditional design. NASA wants to test both vehicles in Low Earth Orbit on the Artemis III mission, which is slated to launch by the end of 2027. Houston-based Axiom Space is developing next-generation spacesuits for walking on the lunar surface, which are also expected to be tested on Artemis III. NASA wants to launch the actual landing mission, Artemis IV, by the end of 2028. This is an aggressive schedule given delays with the lander development and problems with the two launch vehicles, Starship and New Glenn, that are required to launch them. NASA is also planning upgrades to SLS, including a new upper stage derived from United Launch Alliance’s Centaur that will debut on Artemis V and increase the rocket’s payload capacity.

NASA increasingly intends to transfer responsibility for low Earth orbit to commercial operators so that the agency can concentrate its limited budget on deep space, where the commercial case is still thin. That frontier of deep space comes with formidable engineering obstacles that a circumlunar flight like Artemis II never had to fully confront.

• Radiation is chief among them: beyond Earth’s protective magnetosphere, crews on long voyages such as the months-long transit to Mars face a steady dose of cosmic rays and the threat of solar particle storms. Effective passive shielding without crippling mass penalties remains a significant challenge, and active shielding remains in the realm of science fiction.

• Propulsion is another: chemical rockets are slow and inefficient for the distances Mars demands, which has revived serious interest in nuclear thermal and nuclear electric propulsion, including a DARPA-led demonstration program aimed at proving a flight-ready nuclear thermal engine.

• The quieter supply-chain problem lurking behind deep-space ambitions is a shortage of plutonium-238, the fuel for the radioisotope thermoelectric generators (RTGs) that power probes and could help sustain equipment in places where sunlight is scarce, such as the permanently shadowed craters of the lunar south pole or during the 14-day long lunar night.

At a recent event called Ignition, NASA Administrator Jared Isaacman laid out a sweeping vision for NASA’s lunar future, including the establishment of a permanent lunar base starting with the Artemis IV mission.

NASA is now mobilizing the whole US space industry as well as international partners to make this dream a reality. The lunar base, located near the lunar south pole, will enable long-term scientific research with an eye towards eventual Mars missions, and is consciously modeled on the way humanity already operates research stations in another forbidding environment: Antarctica. In Antarctica, rotating crews sustain a permanent scientific presence in a place no one calls home. The base will also focus on pilot programs for in-situ resource utilization, notably the creation of fuel and drinkable water from lunar ice and mining of high-value materials from the lunar surface such as helium-3, which has applications for nuclear fusion and quantum computing.

These missions also come during a time of heightened competition with China, which has announced its intent to send a crew to the moon by 2030 and is making aggressive progress with its own vehicles. Two geopolitical blocs have emerged surrounding lunar exploration: the US-led Artemis Accords, an agreement surrounding responsible lunar exploration and development that has been signed by 67 countries, and the International Lunar Research Station, an alternative group led by China and Russia. The geopolitical stakes are just as high as they were in the 1960s, if not higher, and the US must maintain its space leadership to ensure that democratic norms prevail on the high frontier.

Artemis II was an engineering milestone. It was also a reminder, at a fractious moment in history, after decades of stagnation, of what humans can accomplish when they work together and look up to the stars. As a peaceful, scientific exploration initiative featuring international cooperation and inspirational leadership, Artemis II is America’s best tool in this contest ahead.

Subscribe now

This material is solely for educational purposes and is not investment advice or an offer of investment advisory services. This material should not be used as the basis for an investment decision. a16z is an investor in SpaceX and Cursor through its managed funds, and thus has a financial interest in the company’s performance and future prospects. In particular, a16z benefits if the company grows in value; and a16z funds will receive any customary dividend payments in connection with their status as a shareholder of the company. However, a16z is not being compensated by SpaceX or Cursor for this material.

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

Every technology cycle changes data center architecture and the AI revolution is no exception. Most of the news today is about compute, GPU racks that consume hundreds of Kilowatts of power and require liquid cooling. But an even bigger change may be happening in Networking.

Classic cloud data centers are designed for fast connectivity between servers, to storage and to the internet. They typically use a multi-layer fat-tree network of switches to achieve high bandwidth and resilience. The servers are connected to the network via interface cards on the PCIe bus.

For modern AI data centers, this is no longer sufficient. In addition to this classic front-end network, there is now a second back-end network that directly connects GPUs with each other. This architectural change gave rise to NVIDIA becoming a major networking vendor as well as the proliferation of protocols like NVLINK and Infiniband.

Unsurprisingly, managing multiple networks with multiple protocols is highly complex. Configuration now happens sometimes directly on the switch, sometimes on end points, sometimes on DPUs and sometimes via fabric controllers. Tenant isolation or VPCs (Virtual Private Clouds) need to be created and enforced at networking hardware level, then connected to each other or to the internet via routers and network address translators. And all of this has to happen at massive scale, with a very high degree of reliability. Because if the network is down, the data center with its expensive GPUs stops working.

That’s why we’re excited to invest in Netris. The company has built a powerful and robust network management solution for GPU clouds or anyone running a large GPU data center.

When we spoke with Netris customers, perhaps the single most impressive thing was the scale at which Netris was operating across tens of thousands of GPUs.If you build a GPU data center for close to a billion dollars and entrust its functioning to a software solution from a start-up, that is a strong sign of confidence. Netris has been building network management software for over 8 years.

We were also impressed by the architecture and depth of their solution. At Big Switch Networks a decade ago, we found managing most physical switches extremely difficult as switches lack full-fleshed real-time APIs to manage state. This lack of APIs was a major driver for software defined networking (SDN) that shifted this functionality into the end hosts. As it turns out, times have changed and switches have become more open. Netris today installs agents on the switches that allow them to do much more complex configuration with a much higher degree of reliability. And configuring the basic network is just the foundation. Netris also offers network automation, multi-tenancy and edge services like routing, load-balancing and NAT. Netris further adds many day 0 and day 1 functions, such as the simulation of networks before the rollout and diagnostics of the connectivity post deployment.

But most importantly, we were impressed by the team. They pair a scrappy start-up culture with exceptionally strong engineering and the persistence to iterate until the product could power some of the most advanced data centers of the planet. We are excited to partner with Alex Saroyan, Arsen Arakelyan and Tigran Martirosyan and take networking into the age of AI.

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 structure of the modern AI industry has been shaped largely by scaling laws. Large, general-purpose models have conclusively outperformed smaller, hand-crafted models. Training these large models requires tens of billions of dollars and hundreds of thousands of GPUs, so talent and resources have consolidated in a small number of big labs. And the advantage held by the labs is only compounding as the frontier advances.

This structure has generated phenomenal progress, which we hope and expect will continue. It makes sense for the big labs to continue to develop the core, horizontal capabilities of frontier AI models.

But we also believe the full potential of AI will not be realized until the technology is placed in the hands of builders. The array of problems that language models can address is simply too vast, and too impactful, for a handful of companies to tackle them all. The data and domain expertise that lives outside the labs really does matter - as the labs’ extensive RL efforts demonstrate - and not just as a prompt to send to an API. The most direct path to maturity and massive impact for the AI industry is to let engineers and researchers outside the labs to do real AI work, i.e. to push the frontier in their own domains of expertise.

We’ve seen a great example of this through Cursor, where they have grown from relying on third-party models, to building their own Composer models on top of open source, and now pre-training frontier models at SpaceX. The performance and economic viability of the product has improved in each phase, and the modeling work so far compounds in the same way as the labs’ centralized efforts. Not every organization needs to go all the way to pre-training, of course, but most would see very real benefits from being able to run experiments and update model weights.

Satya’s recent post makes the case very clearly why great technology always has, and always will, need a true ecosystem: developers, developers, developers.

To make this vision a reality in AI, we need two things: (1) A family of frontier open source models available for anyone to extend. This gap is currently plugged by the Chinese models, though that’s likely not the long term solution. (2) A lab-grade research platform to help normal engineers do frontier AI work. This is where Mirendil comes in.

Mirendil is building a system that can help anyone do AI work: they train frontier models that are expert at AI R&D and build the product around it. The system is meant to loop over research and engineering problems more autonomously, and make more and more progress without human intervention. It's like a coding agent that has been built specifically for these kinds of AI tasks and controls its own GPUs. While Mirendil's product is likely to be used by engineers and AI researchers to start, the goal is to support less technical users, so that scientists and other domain experts can run their own experiments.

Call it vibe research. If it works, it could change the way the AI ecosystem is structured, and support experts across many fields: materials science, drug discovery, legal assistants, enterprise software, coding agents, etc.

This is not a simple product to build. It’s a systems problem, in that the backend infrastructure, agent harness, evals, post-training, and even pre-training strategy all need to be designed together. And the training data needs to cover the full loop of AI research, from proposing experiments, to writing and running code, interpreting results, debugging failures, improving kernels, managing compute, comparing checkpoints, and deciding what to try next.

Most of the major AI labs have built some version of this platform for internal use. But they are not necessarily designed for external users, and the labs also don’t have a strong economic incentive to release the platforms for their competitors to use. The recent Fable 5 release, for example, blocked advanced AI research alongside safety risks like bioweapons and cyberattacks. There’s also a real cultural shift required when research teams have the mission to replace themselves through automation.

So, we think there really should be an independent company focused on this problem. And we believe Mirendil has the strongest team to tackle it. Behnam Neyshabur has worked on AI for science for 7+ years at Google and Anthropic. Harsh Mehta built the first version of the Anthropic autoresearch platform, initially as a team of one, and has seen what’s necessary to scale the platform broadly within the company. Shayan Salehian has been a core member of the ML engineering team at xAI dating back all the way to the original Twitter. Finally, Tara Rezaei is a 23-year-old MIT graduate and olympiad medalist with a relentless go-getter attitude.

We’re very excited to announce today that we’re leading the seed round for Mirendil. They are working on one of the biggest possible problems in AI, and they are one of the few teams who have experience and strong priors about how to make the end-to-end system work.

We think the next phase of AI is going to include a broader ecosystem of engineers, researchers, scientists, and domain experts. They will need the tools to do real model work themselves. And Mirendil is building that platform.

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

Hi, I’m Josh, and I’m thrilled to be here. Today I’m joining A16Z as a partner to help find and support the next generation of great consumer companies. In particular, I look forward to helping founders and teams navigate the journey from an early spark to a durable platform. The AI wave is in full bloom and already transforming the world of work; now I think it’s an amazing time to build new AI-first consumer apps and services that people will use throughout their daily lives.

When I turned 50 last fall, I realized how lucky I’ve been to be part of some incredible products and company journeys, along with some special life experiences too. Helping build Facebook, LinkedIn, Twitter and Robinhood were all journeys that started from discovering a new spark of consumer behavior, which is like opening a door to a brand new world of possibility that isn’t built yet.

My career has been all about the world-building that follows that spark. Early users show you what’s in the world; you get to make it come to life. Figure out the user patterns, the feature hierarchies, the growth loops, and the development practices that defined social, and defined mobile, for billions of people. My role in these stories wasn’t to blaze the very first trail discovering new streams and valleys (although I’ve had the privilege to work side-by-side with some great founders who did), nor was it to “operate” or scale the world once built out. I was right in the middle, the messy middle, laying down the rails, and mapping out the towns, and setting the foundations for a durable product, and a durable company. To me, this is hands-down the most fun part of the journey.

It can be tough; it can look silly. As Chris Dixon has said, most great products start out looking like a toy. In the early social days, people saw Twitter as a dumb site where people posted what they had for breakfast, but that has become one of the most important information networks in the world. In the mobile era, people derided Robinhood as a simple stock trading app no one would ever use with their ‘real money’ but that’s now become a full-fledged financial platform for tens of millions. Or take Musical.ly, which was just a way for teenagers to post music videos until it became Tiktok, providing entertainment and more to billions. These were all places I worked at or was on the board through these major shifts. Worlds take time.

In recent years, there was a period where it seemed like consumer was pretty dead. And that was really just a way of saying, “the worlds that consumers wanted to inhabit were all pretty built out”. Instagram works very well as a scaled-up social network; the iPhone works very well as your remote control device for the world. But two important things have changed, that have completely opened the world-building doors again.

First, obviously, is that AI has completely upended the assumptions about what ordinary users can do; we all know this. The biggest consumer breakout so far, ChatGPT, introduced the world to the experience of open-ended conversations and relationships with software; and we’re only at the very beginning of that growth opportunity. But meanwhile, just as importantly, we have a new wave of users, especially the newest cohort of gen-alpha consumers, that really want to make things. They are native world-builders themselves. They came up playing Roblox and Minecraft, they have no preconceived limitations about what an app is, or what they can do with it. They expect to be able to customize or build anything they want for themselves and with their friends. And now, suddenly, we’re there.

The models are finally ready. Costs of inference are getting optimized with open models, and even on-device models. And, when I first experienced OpenClaw earlier this year, I had the epiphany that it isn’t the models that matter, but the harnesses, loops, and context which will lead to so many new opportunities ahead. Just as we’re learning in enterprise, we can finally invent new products that allow users to do things more naturally, using simple language to express their needs.

So what does this all mean for consumer founders and operators? The setup is there; it’s an incredibly exciting time to be building in consumer again. Just like in every major tech wave before, from dot com through social and mobile, I expect that billions of people will soon discover new ways to explore the world, get things done, learn, shop, manage their money, arrange their travel, connect and communicate with friends and family, and much more. We’re ready for all of this, and more, to get reinvented around new interfaces and a new compact between software and the user, that is far more generous and accommodating for users’ peculiar needs, or creative spark.

This means there is a lot of work to do in building out these new worlds, laying down the new rails, mapping its new towns. I can’t wait to meet amazing teams with big ideas, and early sparks of traction. And I love the ones building something that many people will deride as just a toy, but the team knows they have a bigger vision they can’t wait to bring to the world, and bring the world to their vision. I’ve been on so many of these journeys myself. And now, I’m eager to help a lot more teams create something that lasts, and navigate the journey from a spark of something working to making something durable.

It’s time to build!

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

Every business has one decision that everything else bends around. In home services, that decision is dispatch: which technician goes to which job, in what order, at what time. Get it right and revenue, technician utilization, and customer experience all move up together. Get it wrong and no amount of marketing or lead-gen can save the day. Dispatch is the brain of the business.

So it is striking that, for the last three years, almost everyone building AI for the trades skipped it. They rushed to the edges instead - voice agents, chat widgets, follow-up bots - clustered at the top of the funnel where leads come in. It is the most visible part of the business and the easiest place to bolt on a model. The result was a stack of disconnected point solutions, each owning a sliver of the customer, none of them talking to each other, and none of them touching the decision that actually determines whether an operator makes money.

Our thesis is simple: in vertical AI, you do not win by adding another tool at the edge. You win by owning the hardest, most central decision in the business and building outward from it. Probook is doing exactly that in home services, and today we are thrilled to announce that a16z is leading Probook’s $34 million Series A, bringing the company’s total funding to $40 million after a seed round led by Sequoia.

Probook built dispatch first, the part most vendors avoid because it is genuinely hard. Routing the closest available technician is easy; routing the right one is not. Technicians have different skills, conversion rates, and areas of expertise, and the gap between a good match and a bad one shows up directly in revenue. Only after owning that decision layer did Probook expand into the rest of the front office: intake, data scrubbing, customer messaging, and outbound.

Because everything shares one context layer, the whole system works as one. Every customer stays on a single text thread with a single number, from the first call to the technician at the front door. Every inbound lead is answered with full information. Every booking is cleaned before it is assigned. Humans handle the exceptions; the software handles the rest. This is not a copilot that hands work back to an already-overloaded front office. It does the work, and operators add points to their EBITDA as a result.

Probook is hard to replicate because of where it came from. CEO George Eliadis grew up in the industry, pressure washing alongside his dad in upstate New York across six summers in the truck. He knows what it is to be up on a ladder washing a house and miss the call that would have booked the next job. Years later, he spent a summer inside TR Miller, a $40 million HVAC, plumbing, and electrical shop in Illinois that became Probook’s first customer, and saw the same problems play out at scale.

George and his co-founders, CTO Lewis Zhang and Ben Cervantez, have paired that hard-won domain empathy with outlier technical depth. It shows in how they build. Probook deploys with customers in person, configures the platform shoulder to shoulder with their front-line teams, and stays on the hook for the outcomes it sells. In a category full of software that is sold and forgotten, that posture has made Probook a rare vendor that operators actually trust to run their business.

Probook already runs across hundreds of locations nationwide, from independently owned shops to private-equity-backed platforms, including TurnPoint Services, Master Trades Group, Del-Air, Peterman Brothers, and Sila Services. Summers Plumbing, Heating & Cooling, with 14 locations and 260 technicians on the platform, booked 2,542 jobs in its first month on Probook with zero human intervention. Del-Air more than doubled productivity from 10 to 22 techs per dispatcher.

We could not be more excited to partner with George, Lewis, Ben, and the entire Probook team. If you operate a home services business, see Probook for yourself at probook.ai, and if you want to help build it, they’re hiring.

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 I’m launching Renaissance Geek – the company behind Impeccable – with backing from a16z led by Anish Acharya, and a new partnership with GitHub. Here’s the bet I’m making. - Paul Bakaus

It’s a weird time to try to create excellent digital work.

It is suddenly much easier to produce a first draft of almost anything: code, interfaces, copy, images, internal tools, prototypes, decks, product ideas. The blank page is less blank. Work that used to take days now takes minutes.

AI has unquestionably raised the floor.

And yet.

Why does it feel like producing something excellent is getting harder?

We’re being told to let AI execute on our behalf, by trading the time-intensive, manual creative labor for a prompt input. To go all in on loopmaxxing. We might have to sacrifice precision, but hey, at least we’re shipping faster, right? Well, that’s a false dichotomy.

There’s a reason we haven’t yet seen AI video’s Hayao Miyazaki, AI music’s Paul McCartney, AI writing’s Haruki Murakami, or AI design’s Tobias van Schneider.

“Lower the floor, raise the ceiling.” John Maeda reminded me of that phrase recently, and it goes back to the early days of computing at MIT. Lower the barrier so more people can make things; raise the ceiling so experts can go further. John used the phrase for this exact moment, and he generously named Impeccable as one of the tools doing the lowering.

He’s right, and I’ll keep pushing that barrier down, until reaching a solid baseline takes no effort at all. In fact, I’ve invested most of my adult life in pursuit of what I call anti-gatekeeping: jQuery UI allowed a generation of web designers to ship rich web UIs for the first time. Chrome DevTools became a web designer’s trusted tool to preview what their sites would look like on mobile. And yes, Impeccable brings competent design to everyone who wants it.

But excellent work is not one-shotted. It comes from clear intent, relentless iteration, judgment, and craft. It is now very easy to make something. It is still very hard to make something truly great, let alone genre-defining.

Mindlessly shipping more stuff faster is not a strategy, and plenty of companies are about to learn this the expensive way. The creative AI tools we build can’t just focus on lowering the barrier to entry and raising the output floor, they must also enable creatives to insert themselves into the agentic loop for the last mile. That last 20% that pushes the work out-of-distribution, out of mediocrity, beyond the average.

It’s what I call amplified craft, and that’s the work I’m most excited about.

Today I’m announcing Renaissance Geek, the company behind Impeccable, with backing from a16z, led by Anish Acharya.

Impeccable started as an open-source design skill for AI coding agents. It gives builders and agents a shared design vocabulary, persistent project context, deterministic quality checks, and a live browser loop for visual iteration. It runs in the real codebase, where the work will actually ship.

It spread faster than I can cleanly measure. It’s reached over 40,000 GitHub stars and 160,000 installs through skills.sh alone, which reflects only a fraction of the real numbers (Impeccable is commonly installed via npx impeccable install or Claude’s native marketplace).

But Impeccable is also the first expression of a broader bet: AI does its best creative work when a human is actively part of the loop.

Taste Can Be Amplified, but Not Lab-Grown

There is a tempting idea floating around right now: maybe we can solve taste.

Train better models. Build better taste benchmarks. Capture the patterns. Make the model produce tasteful work by default.

I understand the appeal. I also think it misses the point.

One of my favorite new pastimes is categorizing and detecting design-slop. Impeccable is one of the most popular ways to get rid of obvious slop in this category. But the absence of slop doesn’t result in taste.

Slop is also, as it turns out, a moving target.

Here is what it looked like in 2022:

We tried to fix it: We told the models to avoid patterns like purple gradients and glowing buttons, then pushed them into a different part of their latent space.

Which gave us this: 2026 slop.

Instrument Serif italic headlines on warm beige are not bad design. In the right hands they can work. But once the pattern is everywhere, it stops signaling taste, starts signaling the absence of a decision, and worse: it simply doesn’t perform well because it doesn’t stick out. Today’s antidote to slop becomes tomorrow’s slop the moment everyone reaches for it.

So no, you can’t bottle taste. It’s personal, contextual, and always moving; try to mass-produce it and you get algorithmic Uniqlo. What you can do is catch the tells that have nothing to do with taste, and Impeccable does: weak hierarchy, lazy spacing, broken contrast, the compositions every model reaches for when no one is steering.

Stripping slop, though, is the side quest. The main quest is amplifying human intent. The best creative tools hand you sharper instruments and leave the seeing to you; a camera doesn’t tell you what to shoot, it just makes more shots possible.

Design Is Moving Into Production

The design/engineering handoff is breaking. The world it assumed, where a design freezes and then engineers translate it into code, barely exists anymore. Code changes every day now, and an agent can spin up ten new interface states before lunch. A frozen mockup can’t keep up with a product that won’t hold still.

I’ve watched this happen all year, in startups and Fortune 500s alike. PMs are shipping real code. Designers are becoming design engineers. The work is migrating to where the product actually lives.

Through all of it, engineers got a new tool every week. Designers mostly got left out. What did show up for them was built for PMs and engineers, or it was incumbents building richer connectors to an existing design canvas.

Figma still matters, to be clear; it’s a wonderful place to explore and think, and none of this means designers have to learn to code (though I firmly believe knowing CSS and HTML well makes you a better web designer). The shift is simpler than that. More of the real design work now happens against the live product, in the browser, where it ships. And for good reason: it leads to faster iteration without process friction, while others are still waiting for handoffs.

That gap is why I built Impeccable. It teaches agents to speak design, so a designer, engineer, or founder can steer in plain language: critique this layout, fix the type, recolor it, watch it change live in the browser. And whatever you create with Impeccable is real production code, ready to ship.

The Interface Is Still Primitive

The human-agent interface is still early.

Right now, the dominant way to work with agents is chat. Chat is powerful. I use it constantly. But chat is also passive. It gives you a blank box and asks you to imagine everything.

Great creative software works nothing like that: A video editor gives you a timeline, a design tool gives you a canvas with layers and guides, and music software gives you tracks you can see and scrub. None of these just take orders; they show you the work and let you feel it change as you push on it.

Most AI tools still behave like a brilliant intern sitting behind a curtain. You can ask for almost anything, but the interface rarely helps you discover what to ask. That has to change. If agents are going to do creative work with us, they need richer loops: ways to critique, compare, remember, and adjust in real time.

Impeccable’s shared design vocabulary and Live Mode are early experiments in a more active human-agent interface. You can point at the work, steer it, compare variants, and accept changes back into source. There is much more to explore, and Renaissance Geek will invest deeply in the interfaces and workflows that amplify creative work.

Impeccable is our first tool for this new age, but it decidedly won’t be the only one.

The Era Of The Renaissance Geek

For a long time, being a generalist was treated as a liability. Pick a lane. Specialize. Become the designer, the engineer, the PM, the marketer, the operator. Teams were built around handoffs between people with cleanly separated roles.

AI is making that model feel less inevitable.

The people who seem most dangerous right now are T-shaped generalists with high taste, high intent, high agency, and deep curiosity. They may start as designers, engineers, founders, writers, researchers, or operators. But they are increasingly living closer to code, because code is where ideas become real.

That is the Renaissance Geek: A modern-day ‘renaissance man’ who can move between disciplines because AI gives them reach, and whose ever-evolving taste driven by their genuine interest in art, culture, society and technology gives them direction.

These are the people I care the most about: the ones who want the work better, not just faster, who still sweat the last 10 percent and can feel when a product has no point of view.

AI will make those people much more powerful.

What We Are Building

Impeccable is Renaissance Geek’s first product that amplifies craft, and we’re focused on making it a daily, trusted toolkit for solopreneurs and large enterprises alike.

To get Impeccable into the hands of even more builders, I’m also excited to share today that Renaissance Geek is partnering with GitHub.

The team at GitHub cares about the same things we do: raising the ceiling on quality and design. That’s why Impeccable comes pre-bundled with the new GitHub Copilot app, one of the most widely used agent harnesses in the world. Every builder who opens the app gets a design and quality layer from the start.

Backed by a16z in a round led by Anish Acharya, Renaissance Geek will keep building the tools for that last mile: the ones that help humans steer AI toward work worth shipping. Anish and I see the same gap in how today’s models and coding harnesses still fail us on creative work, and he’s written up why a16z backed the company. I’m incredibly lucky to have him along for the ride.

Remember the Renaissance Geek? I’m trying to hire a few. First, a chief of staff to turn this momentum into a real operating cadence, plus design advocates and design engineers who care about the last 10 percent more than is strictly healthy. If that’s you, the open roles are here.

The next wave of AI should not be about pushing humans further out of the loop. That story is boring and dystopian. The more interesting future is delightfully strange and exciting: tiny teams with enormous range, designers moving in code, engineers suddenly learning about vertical rhythm, PMs prototyping real products, agents that become active collaborators, and craftspeople using machines to make work that would have been impossible alone.

AI raised the floor.

Time to raise the ceiling.

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

Orbits are now battlefields — and it’s time we admitted it.

Russia is jamming GPS satellite receivers across most of Eastern Europe. Iran has launched hundreds of ballistic missiles through outer space. China has openly contemplated destroying Starlink satellites. Plainly: our adversaries are waging wars in space.

We can ignore this fact and naively hope that China and Russia will abide by treaties for the peaceful use of outer space. If we do, we will cede the orbital high ground to our adversaries — and lose secured access to the space-based communications, reconnaissance, targeting, positioning, navigation, and timing services that shape the modern battlefield.

Or we can choose to treat space as the warfighting domain that it is — and figure out how to win a space war.

Both authors of this report have seen first-hand how America’s adversaries are waging war in space, which is why we thought this subject worthy of spilling this much ink.

Christian spent seven years as a satellite manufacturer and operator, working on both commercial and military missions and closely tracking America’s adversaries in space. Alex Oliver was a military officer for 16 years. Both of us are now at Andreessen Horowitz as partners on the American Dynamism team

So we remember when a Chinese satellite threw another satellite out of geostationary orbit, the many times they buzzed US satellites, and the aggressive lawfare they waged within the International Telecommunication Union. We watched Russia slowly and surely fall behind in the modern space race, only to make headlines with a rumored nuclear weapon on orbit.

But if our time in the space industry has taught us anything, it is to be wary of conflating the way things are done with the way things ought to be done. Over the past 20 years, the space industry has been upended by companies like SpaceX that care little about precedent — Elon Musk famously fired the original Starlink executive team for their traditionalism and associated slowness — and given SpaceX’s success, the space world must surely have learned the difference between a physical law and an industry convention.

In the first part of this report, we will therefore take a page from the Book of Elon and focus not on what is being deployed into space today, but instead on a first-principles understanding of what space is: meaning, how outer space works as a physical domain, and what those physical laws imply for the prospects of fighting and winning a war in space.

Then we will take a more practical look at the ways, means, and ends that are necessary to win a space war. We’ll discuss the importance of maximizing upmass, proliferating military space architectures, winning the cold war in space, and wielding such fearsome offensive capabilities that our adversaries don’t test our resolve.

In order to understand those strategic considerations, we’ll need a firm theoretical footing — so first we will establish three facts about the physical reality of outer space:

• Space is close to earth, but it’s hard to get there.

• Orbits are predictable, but tracking objects in space is hard.

• Space is physically massive, but operationally small.

PART ONE: FIRST PRINCIPLES OF SPACE WARFIGHTING

#1: Space is close to earth, but it’s hard to get there.

The first principles

You are standing on top of a massive mountain, above the atmosphere, with a baseball in hand.

You throw the ball and it lands 100 feet in front of you. You try again, shooting the ball out of a cannon, and it lands 1,000 feet in front of you. Then you go for broke and attach the ball to a rocket — and the ball never lands. You have launched it “into orbit,” meaning so fast that by the time gravity has pulled it down to “ground-level,” it has already whizzed beyond the horizon.

It turns out that “sufficiently fast”, meaning the specific amount of change in velocity (“delta-v”) needed to get an object into orbit around Earth, is around 9.4 km/s.

That’s not easy. As Christian’s favorite astronaut, Don Pettit, wrote (emphasis and link ours):

All our rockets are governed by Tsiolkovsky’s rocket equation.

The rocket equation contains three variables. Given any two of these, the third becomes cast in stone. Hope, wishing, or tantrums cannot alter this result. Although a momentum balance, these variables can be cast as energies. They are the energy expenditure against gravity (often called delta V or the change in rocket velocity), the energy available in your rocket propellant (often called exhaust velocity or specific impulse), and the propellant mass fraction (how much propellant you need compared to the total rocket mass)...

If the radius of our planet were larger, there could be a point at which an Earth escaping rocket could not be built. Let us assume that building a rocket at 96% propellant (4% rocket), currently the limit for just the Shuttle External Tank, is the practical limit for launch vehicle engineering. Let us also choose hydrogen-oxygen, the most energetic chemical propellant known and currently capable of use in a human rated rocket engine. By plugging these numbers into the rocket equation, we can transform the calculated escape velocity into its equivalent planetary radius. That radius would be about 9680 kilometers (Earth is 6670 km). If our planet was 50% larger in diameter, we would not be able to venture into space, at least using rockets for transport.

When you build a rocket, you load it with as much fuel as possible (max propellant mass fraction), you make the best engine and choose the best fuel you can (max specific impulse), and the equation solves to give you a certain amount of delta-v. That delta-v allows you to perform a certain set of maneuvers in space — like escaping from the surface of a planet, rocketing between stars, changing the angle of your orbit, and so on.

Delta-v is counterintuitive. It takes 9.4 km/s of delta-v to get from the surface of Earth into a circular orbit 350 km up. If you then spent another 9.4 km/s of delta-v, where do you think you’d end up? The answer is the surface of Mars. A satellite on Mars is therefore just as close to Low Earth Orbit as a satellite on Earth, even though the latter is 140 million miles further away by straight-line distance. As sci-fi legend Robert Heinlein wrote, “If you can get your ship into orbit, you’re halfway to anywhere.”

Developing an intuition for delta-v takes a while, but for now, you just need to know two rules of thumb: (a) straight-line distance between two points means very little in space, and (b) it’s hard to get to orbit.

The implications for space warfighting

First, upmass is everything. It’s hard to get mass into orbit, so if you can do it better than your adversaries, you have a massive advantage.

Only 14 organizations in the entire world successfully launched multiple orbital-class rockets in 2025: four from the US, six from China, and four from the rest of the world combined.

As such, securing your access to orbit is a Tier 1 priority: designing more rockets and bigger rockets, scaling rocket production, improving rocket component supply chains, developing more launch sites, developing technology to launch rockets on the move, and so on. Hopefully you start doing these things a decade before a war breaks out.

In a space war, you also want to deny your enemies the ability to launch rockets. This is a more controversial priority, as many of the rockets listed above are commercial, not government-owned — but there is unfortunately precedent for attacking commercial space systems in wartime. Launch infrastructure is, with unfeeling military logic, a bottleneck and therefore a choke point. There are only about 40 active rocket launch pads in the entire world, and only so many factories building new rockets. That’s why Christian made this prediction:

If you lose your ability to launch, you lose your ability to replenish your space fleet and run the risk of being completely shut out of the space domain. This is why General Chance Saltzman called the freedom of movement to (and in) space “the formative purpose of the Space Force,” and why denying your adversaries that freedom is the most straightforward way to win a space war.

Second, upmass limitations affect satellite design. Because getting mass into orbit is hard, each gram is precious and engineers must constantly fight against increasing the mass of their satellites. This is no easy task. A satellite has to survive the violent, loud environment of a rocket launch, and then must survive in the vacuum of space while being bombarded with radiation, with no maintenance, for years. Satellites often have limited physical redundancy (which limits their fault tolerance), limited propellant (which limits their delta-v), and only enough shielding to resist radiation, if they have any shielding at all.

Practically, this means that satellites tend to be fragile, immobile objects — or, in the words of former vice chairman of the Joint Chiefs of Staff, General John E. Hyten, “big, fat, juicy targets.” Satellites generally cannot survive direct strikes of any kind.

Another offshoot of this mass-sensitivity, perhaps counterintuitively but certainly true historically, is that it’s tempting to make each individual satellite as large and as “exquisite” (read: decked out with the best sensors/radios) as possible. This was the dominant strategy pre-Falcon 9, when the U.S. was only launching a few dozen rockets per year, and when those satellites therefore had to be able to do their jobs independently.

This is not a physical law, but rather a constraint to design around. Would you rather have one bulletproof satellite, or ten normal ones? Do your satellites need to survive a nuclear blast, like the AEHF satellites, or would you rather have attritable assets scattered across many different orbits? You have to choose.

Third, space-to-ground weapons are inevitable. In one of Christian’s favorite sci-fi books, Robert Heinlein’s The Moon is a Harsh Mistress, a rogue lunar colony bombards the Earth by flinging lunar rocks down the gravity well — where they ultimately impact Earth going around 11 km/s, delivering as much kinetic energy as a small nuclear weapon. Such weapons are still in the realm of science fiction, but might not remain that way for long.

The prospect of keeping weapons perched on a cliff above one’s enemies will likely prove too tantalizing for all space powers to avoid — so the more important question is whether we expect those weapons to take the form of traditional munitions, rods from God, or something more exotic still. All of these options, however, require so much delta-v to get off of Earth’s surface that they will likely only make sense once they can be made on orbit. Say, by a rogue lunar colony.

#2: Orbits are predictable, but tracking objects in space is hard.

The first principles

It’s hard to launch anything into space without the entire world noticing, tracking, and therefore watching as it traces predictable orbits around the Earth.

Rockets are violent, loud, bright events that take place from predictable locations on Earth, and are therefore extremely noticeable. Starship launches are occasionally mistaken for earthquakes; Falcon 9 launches at twilight create massive alien jellyfish in the sky. Even beyond the seismic, auditory, and visual signals, rocket launches get so hot that they flash bright in infrared — a fact that global superpowers learned early in the Space Race, and exploited by launching infrared-sensing early-warning satellites. Still today, global powers maintain IR satellites given their importance to defense: the U.S.’s Space Based Infrared System has cost America more than $20 billion to operate and maintain, and its replacement, Next Generation Overhead Persistent Infrared, was expected to cost $14.4 billion from 2021-2026.

Once in space, satellites travel along regular and repeating orbits; the International Space Station (ISS), for instance, completes a full orbit roughly every 90 minutes, and systematically passes over the entire Earth.

The ISS, therefore, is visible to about 90% of the world’s population at least once every three days. Amateur astronomers can use the well-known orbit of the ISS to point their telescopes, hoping to catch the Station as it transits (passes in front of) the Moon or the Sun. This same principle — the predictability of orbit — holds true for every satellite on orbit today, and means that tracking the physical locations of most satellites is trivially easy.

It’s also easy to pick up the radiofrequency (RF) signals the ISS constantly transmits — you can listen to astronauts in your backyard with a $15 handheld radio. The ISS intentionally sprays RF indiscriminately over the ground, but the same observability extends even to other satellites. Most spacecraft carry two radios: one for the “payload” signals (comms traffic or images taken on orbit), and another for “telemetry, tracking, and command” (TT&C; signals used to drive the satellite and get its vitals down to mission control).The goal of a TT&C antenna is not to maximize throughput, but to be able to command the satellite no matter what’s happening to it — even if it’s a very bad day and the satellite is tumbling uncontrollably — so TT&C radios are omnidirectional. That makes satellites quite visible on orbit, letting anyone on Earth track one if they’re motivated enough.

Less physically, there are strong international norms around disclosing rocket launches and on-orbit satellite maneuvers before they happen. The UN maintains a register of all space objects, and the Space Force’s 18th Space Defense Squadron maintains space-track.org to share what it knows about every object it’s tracking.

So, normally, it’s easy to keep track of satellites — but space war would certainly be an abnormal time. If someone wants a satellite or especially a satellite payload (read: sensor or effector) to remain hidden, they can easily make it so.

It’s exceptionally difficult to track every satellite constantly; normally, you would follow a satellite long enough (using radar/RF if you can, or a telescope if you can’t) to determine its orbit, then leave it be while you tend to other satellites. But if a satellite maneuvers (or releases a subsatellite payload) while you’re looking away, it can take a long time to “re-acquire” the satellite. It’s also far, far harder to track satellites the further they get away from Earth — especially if they’re in cislunar space or beyond. Speaking from experience, determining where a satellite is on orbit is hard even when you fully control the satellite — the first hours after separation from the rocket, when you’re trying to locate the satellite that you built and launched, are among the most stressful of any satellite operator’s career.

So, satellite orbits are normally predictable. Except when our adversaries don’t want them to be.

The implications for space warfighting

First, attribution in space is extremely hard. If you’re in a plane flying over Kuwait, and you get shot down by ground-based air defenses, Kuwait probably did it. If you’re in space, and your satellite stops working, it’s hard to know why.

For example, a Starlink satellite recently broke into multiple pieces. It now seems that the breakup was due to an “internal energetic source,” but to confirm that (and rule out foul play) took a detailed investigation — which took time and effort, because satellites are always flying blind. Imagine you are SpaceX and your satellite suddenly stops sending you telemetry (read: data on its vital signs). You don’t know why. You cycle through a set of procedures designed to get back in touch with the spacecraft (e.g., spin up a backup ground antenna, command resets to the satellite in the blind, and so on), and only after getting deep into the investigation does “maybe the satellite exploded” become a serious theory.

In a war, perhaps you’d know about this possibility and have a series of telescopes ready to assist with visual inspection early into any anomaly investigation. Even if you did, however, it would be very difficult to know (and harder to prove) who hurt your satellite doing what. Certainly the clues to solve that attribution won’t come from the satellite that was lost, and probably any investigation will take precious time that you don’t have to spare — especially if it’s not one satellite, but dozens or hundreds, that suddenly stop transmitting.

Delays in combat are deadly, and due to the lack of attribution, space has a thicker fog of war than any other domain.

Second, the “dynamic space operations” dream has obvious appeal. If you can move your satellites at will, you can not only relocate them — say, to mass satellites in a particular position at a particular point in time — but also make them considerably harder to track. Which means that your adversaries need to spend, and will spend, more resources making sure they know where they are at all times.

China, for instance, has demonstrated large maneuvers with their Shijian-21 satellite and some suspicious operations with Shiyan 12-01 and Shiyan 12-02 (more on those later). Russia has some “interestingly” maneuverable satellites in Resurs-P3 and Luch Olymp K-2. And the United States has Geosynchronous Space Situational Awareness Program (GSSAP) which explicitly requires maneuverability, and the RG-XX program that will replace it.

Space planners would love all of their satellites to be maneuverable, but it’s physically challenging. You have to carry your fuel with you throughout your lifetime on orbit, and there’s no free lunch — today, you choose between chemical propulsion (fast acceleration, bad gas mileage) and electric propulsion (slow acceleration, good gas mileage). Thrusters that can move quickly and efficiently are exciting but still in the realm of science fiction.

Third, winning an advantage in Space Domain Awareness requires constant one-upsmanship. All three space superpowers intend never to get surprised on orbit — and all three want to have technologies that surprise each other.

The exact tit-for-tat of this game gets classified quickly, but the game’s existence is a matter of public fact. America deploys satellite-sensing satellites to look at the satellites of our adversaries. Then our adversaries create satellite-sensing satellites to look at our satellite-sensing satellites. And so on. The Washington Post reported on an incident in which USA 270 and Shiyan 12-01 / Shiyan 12-02 played out a game of cat-and-mouse on orbit — but it’s still unclear to us, having read everything public on the encounter, who was the cat and who was the mouse.

Winning an intelligence advantage on orbit normally requires new technologies that are withheld from the public eye (see: secrecy around national spaceplane programs). But a more modern strategy has emerged recently: scale. The more you put on orbit, the more your adversaries have to track.

Today, the United States has roughly an order of magnitude more satellites in orbit than any other country — a massive advantage — but China’s space program is nothing to laugh at. China’s rockets occasionally fall on rural villages, sure, but their recklessness is evidence of their maniacal resolve to rival the United States’ capabilities on orbit. They are launching many different kinds of payloads: from spaceplanes, to communications satellites, to their own Space Station. You can laugh at China’s failures, but do so at your own peril. Failures are how you learn in the space industry. Just ask SpaceX.

#3: Space is physically massive, but operationally small.

The first principles

Yes, we know, space is huge. Low Earth Orbit is about 260 billion cubic kilometers1, and there’s the rest of the known universe beyond it. Still — we want to advance the argument that the peculiarities of orbital dynamics make outer space effectively small.

For example, take geostationary orbit (GEO), where satellites take precisely one day to orbit the Earth and therefore seem to hover, stationary, in the sky. GEO is extremely far from Earth; you can fit 2.8 Earths in between the surface of Earth and geostationary satellites.

And yet, the right way to think about GEO is not as a massive, 3-dimensional volume of space, but as a single, 1-dimensional line.

You can only stay stationary relative to Earth if you are exactly 35,786 km away from Earth and exactly at the equator. If you end up one km closer to Earth, you’ll orbit faster than the surface rotates and drift east; if you end up one km further away from Earth, you’ll orbit slower than the surface rotates and drift west. So, with only one acceptable latitude and only one acceptable altitude, the only dimension left is longitude. Granted, that dimension is still large — 264,924 km long — which gives an average distance of 720 km between each active geostationary satellite. But the point stands: space is operationally small despite being physically large.

Space’s smallness is also a result of Kepler’s first law, which means that every circular orbit is centered on the Earth’s center of mass. Most satellites today are in circular orbits, which means in practice that satellites orbiting at a particular altitude must be tightly coordinated.

The SpaceX Starlink constellation is a great example, and a sight to behold. The public is now familiar with the Starlink satellite “trains” that glint at dawn and dusk, and space enthusiasts may even have played around with one of the beautiful Starlink visualizers, of which there are many.

But when we look at the constellation, we see something like this:

Hell of a chart — but it’s illustrating something quite simple. The cyan numbers at top say that the satellites in this “Group” are 475 km above Earth at 53 degrees “inclination,” meaning tilted up 53 degrees relative to the equator. Each red thing is a Starlink satellite, the “planes” of the X axis separate the different trains of satellites, and the “phase” of the Y axis shows in what position an individual satellite sits in its train. Even if we lost you there, observe one thing that should be obvious from a scan of the chart — because the vertical lines of satellites are non-overlapping, they will not run into each other.

But now imagine that another satellite operator wanted to use that same altitude, 475 km, but at a different inclination. It would be chaos! The two trains would intersect twice, once on each side of the Earth, and would therefore be at constant risk of collision. Avoiding collision would take an unrealistic level of coordination, and so in practice, Low Earth Orbit constellations tend to operate at different altitudes. This, again, makes a large, 3D slice of space look 1-dimensional: LEO’s altitudes are GEO’s longitudes.2

Unfortunately, however, even this 1-dimensional picture understates the problem. These rival constellations must co-exist not only in physical space, but also in radiofrequency space.

Coordinating radiofrequency spectrum is a topic on which entire dissertations have been written and for which entire government agencies have been created. It’s also a topic Christian is personally extremely interested in, having led the regulatory team at Astranis for a number of years, and having worked on terrestrial spectrum auctions in his early career as a consultant.

In the same way that each Starlink satellite needs its own swimlane to avoid hitting other satellites, each packet of RF energy emitted from a satellite needs its own spectrum to avoid interfering with other packets. Anyone older than 30 will remember this from using the car radio on a road trip: as long as channels operate at different frequencies (e.g., 98.5 and 97.1 FM), they can both operate in the same area, but if you are driving across town boundaries and each one has a station broadcasting at 98.5, they will interfere with each other.

Satellites primarily operate in high-frequency bands, like S, C, Ku, and Ka. The former three are the legacy bands. Ka band is a more modern, higher-frequency band,3 and satellite operators are even pushing into still-higher frequencies. The point of going higher frequency is that you can push down more data per second; the risk is that it’s generally harder to build the radio equipment and your signal is more sensitive to things like clouds. The higher frequency bands are generally more open with fewer operators broadcasting within them, and the legacy bands are more heavily contested — and spectrum licenses to operate in those legacy frequencies can be extremely valuable. Amazon, for instance, is trying to buy a satellite company for $11.5 billion, presumably more for their spectrum than their satellites.

We will leave our spectrum overview there. Christian could talk more about how the FCC recently modernized its spectrum sharing rules, wax poetic about the possibility for lasers to save us, or share war stories about how Astranis snatched geostationary orbital slots from incumbents who were asleep at the wheel — but for now, it’s enough to just know that spectrum availability also makes space seem small.

The implications for space warfighting

First, access points to space can be controlled. We talked earlier about how there are only a few dozen launch sites in the world, and if you can control these sites, whether from orbit or with conventional, terrestrial forces, you can control the entire space domain.

Spectrum is a similar bottleneck. If you know what frequency someone is operating on, you can control, or at least deny, their signals. Radiofrequency jammers are extremely common in modern wars — the frontline of the Russia-Ukraine conflict has been called the “most signal-dense combat zone in human history” due to drones and the electronic warfare systems being used to combat them. Russia has also jammed and spoofed GPS satellite signals around the conflict zone, creating issues so prevalent as to have become commonplace for commercial airline pilots.

Ground stations can also be a controllable access point to space. In order to receive signals from, or transmit commands to, a satellite, you need an antenna on the ground. These antennas have historically been 9-, 12-, or even 24-meter (diameter) behemoths, and because parabolic antennas shine brightly with a distinctive “glint” when viewed through Synthetic Aperture Radar, they are difficult to hide as potential military targets. Modern ground stations, like those of a16z portfolio company Northwood Space, are generally smaller, flat-panel, phased-array antennas that use many coordinated modules to achieve the same performance as a single, gigantic antenna — which can go a long way toward solving this problem, but which have not yet been deployed broadly.

Second, single entities can have outsized control over the entire domain. SpaceX is the obvious example. Starlink has control over its chosen altitudes in Low Earth Orbit, which “only” took them seven years and around $12 billion of outside capital to achieve. (The Space Force budget request for 2026 alone is $71 billion, for comparison.)

It would be easy to imagine another company achieving similar control over other altitudes in Low Earth Orbit, over other special orbits in LEO like the sunsynchronous orbits, over choice longitudes in geostationary orbit, or even over stable gravitational anchors: planets, moons, and Lagrange points. There just aren’t that many unique places in space around Earth, so a sufficiently motivated actor — say, a geopolitical rival — can act fast to take control of a large percentage of them.

Third, one small domain with global overwatch is extremely valuable. Space has been called the “ultimate high ground” for decades, and for good reason — from one place, you can see the entire world.

This vantage point is extremely valuable. Modern warfare is deeply reliant on space not only for the obvious services like communications and GPS, but also for overwatch: global missile warning, munitions targeting, battle damage assessments, analyzing troop movements, and far more. It is powerful to have your assets deployed to a location from which they can see and cause effects anywhere in the world. All of this is to say nothing of the incredible commercial services that rely on the space domain, of course. Space is worth preserving.

Fourth, and by far most importantly, space is a fragile domain. As in any small domain, attempts to take out a single object in space might have cascading downstream impacts. To illustrate the risk, let’s do some spherical cow-grade simulation and assume a satellite is struck by a munition that explodes in its precise center of mass.

Instantaneously, two things happen: the single object becomes many smaller objects, and the single orbit becomes many diverging orbits. On average, these orbits will still have the same orbital velocity as the original satellite, and given another peculiarity of orbital mechanics, they will all end up returning back to where they originally exploded. As we know, introducing thousands of random objects in random orbits into a dense orbital shell is dangerous. If one satellite explodes, other satellites will have to diverge from their planned orbits and (God forbid) may themselves get hit by debris, and eventually cause a runaway effect.

The picture is even more dire if the satellite was struck by a nuclear weapon. This is true both for obvious reasons — bigger blast, intense x-rays, and more — and less obvious reasons, like that the radiation from the blast would get trapped by Earth’s magnetic field and cause satellites to fail, potentially extremely quickly and at massive scale.

While we think the odds of accidentally triggering a Kesslerian doomsday scenario during normal operations are wildly overblown, a malicious actor could probably find a way to hit the “reset” button on a dense orbital shell, and potentially on all of near-Earth space.

As America continues to deploy more assets into space, life on Earth — both civilian and military — grows increasingly dependent on the space domain. And by the brutal logic of war, this makes space a better target for America’s adversaries. As Colin S. Gray said:

“It is a rule in strategy, one derived empirically from the evidence of two and a half millennia, that anything of great strategic importance to one belligerent, for that reason has to be worth attacking by others. And the greater the importance, the greater has to be the incentive to damage, disable, capture, or destroy it.”

Simply: the more powerful one nation becomes in their control of outer space, the greater the temptation for other nations to hit the reset button.

The way to win a space war is to achieve normal wartime objectives — achieve freedom of maneuver; secure lines of communication; deter and prevent coercion — and to deny those objectives to your adversaries, while preserving the space environment. You can’t create massive fields of debris or render the environment otherwise unusable, and you also have to prevent a sufficiently aggrieved adversary from doing the same.

In a space war, America must fight on the side of order. And in the rest of this report, we’ll discuss how we can fight that war — and win it.

PART TWO: MEANS, WAYS, AND ENDS

Russia just mobilized a fleet of satellites to surround a commercial Earth-imaging satellite: ICEYE-X36, a Finnish satellite that sells imagery to Ukraine.

Over the course of a week this May, five Russian satellites fired their thrusters and burned a remarkable amount of fuel4 to match the ICEYE satellite’s orbital inclination. In plain English, this means that the satellites are now extremely close to each other, with a closest pass of just 500 meters — all while the satellites orbit 550 kilometers above Earth’s surface.

It’s hard to interpret Russia’s behavior as anything less than a threat, with the not-so-subtle subtext that the Kremlin is not pleased with ICEYE supporting the Ukrainian Ministry of Defense. And when Russia feels threatened, they have historically shown their willingness to attack

commercial satellites — it is official Russian policy that commercially-owned infrastructure that aids in military efforts may be “legitimate target[s] for a retaliatory strike.”

All five Russian satellites reached their target orbits as of May 21, so this situation is fast-developing — but it is just as accurate to note that this act of aggression on orbit is many years, and indeed decades, in the making.

As we discussed earlier, the United States, Russia, and China are actively at war in orbit. Russia is intimidating commercial satellites and threatening to launch a nuclear weapon; China is dead-set on seizing the water ice on the lunar south pole; both nations interfere with U.S. military satellites “every single day.”

The undeniable militarization of space motivated our deep dive into the physics of outer space. As we learned, those physical laws constrain our development of warfighting technologies for the orbital domain, and must guide any future, theoretical development project for space.

Now, we undertake a more practical analysis of the present-day technologies and strategies of space warfighting. In other words, how to win a space war.

This means expanding our desired end state of winning the space war, which will require means, ways, and ends: means are the capabilities required; ways are operating logic; and the ends are the conditions that define the end state.

To begin, therefore, we will start with the means — three operational goals that highlight the capabilities the United States needs to successfully wage a space war. These are maximizing upmass to orbit, scaling commercial satellite manufacturing, and creating a resilient military space architecture.

These means enable the ways we intend to win the war: to proliferate capabilities and effects across orbits, and to protect the domain from enemy aggression.

And, in turn, these means and ways should bring about our desired ends. We must ensure freedom of maneuver, both from Earth to orbit and in orbit; we must secure lines of communication for our space and other fighting forces; and we must deter and prevent coercion of allied commercial and military assets.

If we do, we will win the space war.

Means: Maximize upmass to orbit.

If you want to win a space war, you need to get a lot of mass into space — and that means developing launch infrastructure: rockets, component supply chains, pads, testing infrastructure, and more.

Current state

Every part of the rocket supply chain is hard, but the rockets themselves remain undeniably the hardest. To date, humanity has only invented one launch vehicle capable of delivering mass to orbit repeatedly, reliably, and cost-effectively: SpaceX’s Falcon 9 rocket.

Last year Falcon 9 launched more than every other rocket on Earth, combined, and its secret ingredient is reusability: if you don’t throw away the airplane after every flight, you dramatically reduce the cost of orbital launch. The precise scale of this cost advantage is not public knowledge, even following SpaceX’s S-1 filing, but in 2013, Elon Musk said that the first stage of a rocket is 75% of a rocket’s cost — so reusing it up to 30 times is a huge deal.

SpaceX landed its first booster in 2015, and today, more than 10 years later, they are still the only entity in the world, company or country, to land orbital-class boosters repeatedly at scale.

Blue Origin, the American rocket company owned by Jeff Bezos, is tantalizingly close to matching SpaceX’s achievement. Blue today has a reusable suborbital rocket, New Shepard, which flies paying customers like Katy Perry to the Kármán Line and back, and a heavy-lift orbital rocket, New Glenn, which can carry 45,000 kg to Low Earth Orbit (roughly 2x of Falcon 9, and ½ of Starship). In November 2025, Blue successfully landed a New Glenn booster after carrying a NASA payload to orbit, and launched/landed that same booster again in April 2026 — but unfortunately, the second stage of the rocket (which sits on top of the booster and carries the payload to orbit) failed, making it difficult for Blue to claim complete victory. And in May 2026, just as Blue seemed to be gaining momentum and preparing to launch 100 New Glenn launches per year, they suffered the “worst disaster in the history of [the company]” when a New Glenn rocket exploded during a test-fire and destroyed the pad on which Blue’s orbital (and, notably, lunar) rocket launches rely.

Rocket Lab, a New Zealand-founded and US-domiciled rocket company, is also close: they are already launching at volume today — with 46 consecutive successful launches over a period of 2.5 years — but the Electron rocket is small (just 320 kg to LEO) and not reusable. Rocket Lab is working on a bigger rocket, Neutron, which would rival Falcon 9’s payload capacity to orbit, but it has not yet launched and recently had a setback that delayed the schedule to Q4 2026. (A timeline that does not yet account for Berger’s Law.)

Other exciting attempts to develop reusable orbital launchers in America include those of Stoke Space, Relativity Space, Firefly, Astra, and Cowboy.

China today has by far the largest and most active space launch industry outside of the United States, with multiple independent projects underway. The state project to develop the Long March rocket family lives under many different organizations (CASC as the parent; SAST and CALT as nested subsidiaries); the five major “non-public” companies — known as the “space dragons” — are CAS Space, LandSpace, Galactic Energy, iSpace, and Space Pioneer. No Chinese entity has successfully recovered an orbital rocket booster to date, but these efforts are well-funded, moving extremely quickly, and notably unhindered by environmental reviews as are common in liberal democracies.

There are also a number of non-reusable rockets still on the market, invariably national projects massively subsidized by their respective governments. The American SLS and Vulcan, the Russian Soyuz rockets, Chinese Smart Dragon-3, the French Ariane, the Indian PSLV and GSLV, the Japanese H3 — these rockets would struggle mightily in a commercial competition against any reusable rocket, but have been kept alive for political reasons.

Importance to space warfighting

In 2025, the United States launched about 9.3 tons to orbit per day, China 1.8 tons, and Russia 0.4 tons.5 We have a commanding lead, but that lead is not safe.

Just ten years ago, the United States also held a commanding lead in upmass, but with 10x less launch capacity: just 0.9 tons to orbit per day against China’s 0.4 and Russia’s 0.4. Russia stagnated, China 4.5x-ed, and we 10x-ed. Standing still is not an option if we want to maintain our launch advantage. Starship, New Glenn, Neutron, Nova, Eclipse — we need them all, and more.

Holding a launch advantage is powerful in a space war. If you can put the most mass into orbit, you can “spend” that mass however you want — on better shielding from debris or radiation, on more propellant for your vehicles, on more experiments to prove out new technologies, on massively proliferated constellations of assets, or even on simple redundancy.

Upmass capability also allows you to replenish on-orbit capabilities quickly in the event of disaster. Winning a replenishment sprint is not exactly the same thing as winning a decade-long upmass marathon, but both require more launchers, more factories to build those launchers, and more spaceports.

Spaceports in particular are often neglected. SpaceX has started on the path towards building additional launchpads, which is incredibly difficult in America. China has built launchpads in the center of the country — a strategy which has downsides but also makes those assets easier to defend in wartime. The FAA in the United States has so far proven unwilling to consider spaceports being established with overland flight paths, despite precedent from the Space Shuttle era.

This means in practice that all of our spaceports are on the water, and therefore relatively easy to attack with air, sea, or subsea vehicles in a time of conflict. At the very least, the United States needs an inland launchpad for redundancy and national security missions, but we see no strong argument against having a fleet of launch sites across the sparsely populated American West.

Means: Scale commercial satellite manufacturing.

If winning a space war requires mass on orbit, we will need to fill up our rockets with useful satellites, which means making massive factories that build spacecraft by the hundreds or thousands.

Current state

Most objects in orbit today are Starlink satellites. SpaceX has launched more than 12,000 Starlink satellites to date, of which 8,600 are active.

This is a remarkable achievement. SpaceX is only a few hundred satellites away from having launched as many satellites as the rest of the world combined, all time, despite having given the world a 61-year head start. SpaceX uses these satellites to provide fast, low-latency, global connectivity, and the constellation is still growing — Elon tweeted that he wants to launch over 10,000 Starlink satellites per year in the near future.

As with launch, however, there is a wide gulf between SpaceX and the field. The rest of the top ten biggest constellations have less than 1,600 satellites active on orbit, combined.

These constellations (and the rest of the world’s on orbit satellites) divide into three applications relatively cleanly: communications, imagery, and defense.6

Communications

The most obvious feature of space — that it’s high up, and objects there can see a wide swath of the Earth’s surface — is hugely beneficial for radiofrequency communications, which generally require a direct line-of-sight between transmitter and receiver. A 30-meter-tall cell tower can only transmit and receive out to the horizon, roughly 20 km away; a satellite in Low Earth Orbit might be 550 km up, and therefore be able to see out to a horizon more than 2,500 km away.

It should be no surprise, then, that communications has always been the dominant industry in space, whether for broadcast, broadband, or, soon, direct-to-cell services.

The world’s first (and perhaps still the coolest) communications satellite was launched in 1960, less than three years after Sputnik. It was known as “Echo 1” and measured a massive 30 meters in diameter — but it was little more than a mylar balloon, acting as a mirror for communications signals in Low Earth Orbit. NASA engineers beamed a message up from California, bounced it off the balloon’s metallic surface, and successfully received the reflected message at Bell Labs in New Jersey.

Subsequent generations of communications satellites got more technologically advanced as they moved into higher orbits — either in medium-earth orbit (where the GPS satellites operate today), or in geostationary orbit (where a single satellite can hover “stationary” in the sky). Up high, you need to deploy fewer satellites for consistent coverage, which was a huge boon for the first truly successful space businesses, like DirecTV. In 2001, DirecTV generated about $10.3 billion of revenue (in 2026 dollars) with just five satellites on orbit, which you can compare to SpaceX generating $11 billion in revenue with Starlink in 2025 with an average of 6,000 satellites in orbit.

LEO has always had appeal — it’s closer to Earth, which shields satellites from radiation and reduces their communications latency with the ground — but before Starlink, satellite constellations only had one principal difference from high-orbit, independent satellites: they were incredible incinerators of cash. Iridium raised $5 billion, Globalstar raised $4 billion, ICO Global raised $2.6 billion, Teledesic raised $1 billion, and all of them failed. Even 20 years later, OneWeb went bankrupt after raising $3.4 billion in its struggle to deploy a LEO broadband constellation. It turned out that having a reusable, orbital-class launcher was a necessary precondition for a successful satellite megaconstellation.

Imaging

Imaging satellites, in contrast to communications satellites, have been dominated by defense use cases since their creation. Nations want surveillance over their adversaries — and, perhaps obviously, you can see a whole lot from space while being very hard to shoot down.

In the early days, imaging satellites like Landsat 1 were fielded not for commercial purposes, but for their strategic importance.

Landsat was followed by GOES, the NOAA-operated Geostationary Operational Environmental Satellite that launched in 1975 to capture weather imagery over the Western Hemisphere.

Even with innovation driven and funded by defense applications, the commercial market in those early days was tiny — perhaps on the order of $100 million total global sales in 2000, one-fiftieth the size of DirecTV alone at the time.

In the 2010s, small satellites breathed some life into the commercial industry; shoebox-sized sensors, launched by reusable rockets, suddenly made the business case possible to close. Skybox Imaging was the first player to take a real shot on goal, raising $91 million and launching just one satellite before getting acquired by Google in 2014 for $500 million. Planet Labs was the big winner of this cohort, going public in 2021 at a $2.8 billion valuation. Today, the company images the entire surface of the Earth every day at 3-meter-per-pixel resolution.

Even so, satellite imaging is still dominated by defense customers: the rule of thumb in the US industry is that 50% of Earth observation revenue comes from the US government, 25% from allied governments, and just 25% from commercial buyers.

Importance to space warfighting

Commercial satellite systems have both direct and indirect uses in space warfighting.

Some space systems are “dual use” — in which the same system is used by government and commercial buyers alike. The intelligence community buys commercial satellite imagery; SpaceX launches satellites for the NRO using normal Falcon 9 rockets; and so on. The more robust the domestic space economy of a nation, the more opportunities it has to use commercial systems for direct military services.

Similarly, nations with leading private space enterprises can work with those companies to build systems for exclusive government use. The best example of such a system is Starshield, an independent constellation of Starlink-like satellites that are built by SpaceX and operated by the U.S. Government. This also occurs at the component/sub-system level, of course: companies like Apex have stood up factories that can build dedicated satellites for government and commercial customers alike, and companies like K2, Endurosat, and Impulse Space are similarly ready to scale up to match government demand. This is the modern equivalent of Ford’s Willow Run bomber plant during the Second World War — if you have production capacity and manufacturing know-how in your country, you can build for a war effort.

Deploying commercial satellites at scale also provides indirect benefits to a space warfighting effort. As noted above, plain old scale — launching more stuff into space — makes it harder for your adversaries to track exactly what you’re doing on orbit. This benefit also has a dark side, of course, in that commercial assets might be suspected to be military assets in disguise, or treated as military assets despite having dual-use missions. The latter seems to be the case with the ICEYE-X36 satellite, as described earlier; the former is our suspicion for effectively every Chinese rocket and satellite program, given the scale of the CCP’s military-civil fusion.

The lines are blurry between military and commercial space operationally, and even legally. Under Article 1 of the Outer Space Treaty, no nation can claim territory in orbit, which means no nation can lawfully keep any other nation away from its satellites. Russia surrounding ICEYE at 500 meters, or tailing an NRO satellite, is provocative and reckless — but under current law it might not be illegal. France accused Russia of an identical maneuver in 2018. Nations are also still responsible, and financially liable, for any satellite that launches from their country — commercial satellites included — so there is technically no such thing as a purely “private” asset in orbit. And, finally, the principle of proportionality of armed conflict forbids any strike whose expected collateral damage is excessive relative to the military advantage gained, but that might apply to nearly every kinetic attack in space. If a single hit can spawn a cascading debris cloud that lingers for years, perhaps no attack in space, even one made in self-defense, can be proportionate.

In practical terms, however, these treaties historically mean little once conflicts emerge. If you care about maintaining international norms on orbit, you have to prevent a conflict from going hot in the first place — because once shots are fired on orbit, nobody’s going to think twice about violating the Outer Space Treaty as an act of defense.

One hopes that commercial activity in space acts will act as a deterrent to hot conflict. If a nation shoots something in orbit, they are putting at risk their own military and civilian satellites — and the astronauts onboard the International Space Station and China’s Tiangong. Even for Russia and China, that is not an easy amount of collateral damage to sign up for, so the proliferation of commercial and civil activity in space — even from China and Russia! — may deter space-to-space offense.

But relying on that deterrent effect is risky. What matters most is not the absolute amount of dollars in hardware a nation has on orbit, but the relative value of orbit to that nation versus its adversaries. Russia, for instance, has just over 300 satellites on orbit today, compared to about 1,200 from China and about 11,500 from the United States. Iran has 14. You’d therefore expect that Russia or Iran would have a lot more to gain from hitting the orbital reset button than China, which only makes Russia’s purported nuclear threat all the more credible.

Means: Disaggregate and proliferate military satellite systems.

Current state

In 1957, Sputnik ignited the Space Race, a manic techno-geopolitical competition between the world’s two rival superpowers. History has now whitewashed those tense, belligerent years, but the real reason Sputnik terrified the United States into spending $300 billion (2025 USD) on the Apollo program was that it proved the USSR had better intercontinental ballistic missiles than we did. If the Soviets could launch a basketball-sized satellite into space, they could land a thermonuclear warhead in downtown Manhattan.

This fact alone has caused space and national security always to be deeply intertwined. From Sputnik’s launch through the fall of the USSR, the Americans and Soviets frantically developed space-based nuclear weapons delivery systems, reconnaissance satellites, and orbital military outposts — the USSR even launched a literal space cannon.

Today, decades later, not much has changed. Russia is jamming GPS satellite receivers across most of Eastern Europe, Iran has launched hundreds of ballistic missiles on trajectories through outer space, and China has openly contemplated destroying Starlink satellites. All three nations are jockeying for control of strategic orbits and water deposits, trying to take pictures of each other’s spy satellites, and demonstrating anti-satellite weapons, both non-kinetic and kinetic.

Every space-faring nation is now building its own independent military infrastructure on orbit. For navigation, the United States has GPS, Russia has GLONASS, China has BeiDou, the EU has Galileo, India has NAVIC, and Japan has QZSS. For weather imaging, the United States, China, Russia, India, Japan, and South Korea have their own on-orbit assets. The nations also have their own geostationary data relay satellites, missile warning satellites, lunar exploration plans, and, of course, launch vehicles. Even the International Space Station, which was once a symbol of unity for all mankind, will soon be deorbited. China has permanently crewed its replacement, Tiangong, and Russia is working on a space station of their own as well.

That there exists a premium on sovereignty in today’s world should not be a surprise. Geopolitics is real again, and nations know that overreliance on international, shared space assets can be punished — like when Starlink allegedly turned off over Ukraine in 2022. In building this sovereign infrastructure, nations have tended to treat each individual satellite as a grand national project, which has therefore turned these satellites into exquisite assets — absolute behemoths that are so expensive, nations can only afford to field a few of them. America has only ten workhorse satcom satellites, ten more missile warning satellites, and 32 GPS satellites on orbit today.

In response, many nations around the world have demonstrated the ability to shoot satellites out of the sky.

On September 13, 1985, Air Force test pilot Doug Pearson launched a missile from an F-15A and shot the Solwind P78-1 satellite out of the sky. The missile, an ASM-135, struck the satellite at an altitude of about 525 km with a closing speed of 6.7 km/s (or 15,000 miles per hour, in freedom units). The attack took just under three minutes from missile release to impact.

The Secure World Foundation maintains a comprehensive catalog of kinetic anti-satellite missile tests, and has logged 33 by the United States (ending in 2008), 26 from Russia (most recently in 2021), 14 from China (most recently in 2023), and two from India (both in 2019).

The ability to shoot satellites out of orbit is now well-documented, and the U.S. Defense Intelligence Agency believes that China now has the ability to strike space targets in higher orbits, possibly even extending all the way out to GEO.

Importance to space warfighting

Because today’s military space assets tend to be exquisite, but few in number, degrading or destroying even a few satellites would make prominent and predictable holes in global coverage that an adversary could exploit.

Recently, the United States and other nations have started to invest in making their space systems more resilient, which — per Space Force General Chance Saltzman — can and must take many forms:

• Disaggregation — today, satellites tend to be loaded up with sensors completely unrelated to their primary missions. GPS satellites, for instance, also have Nuclear Detonation Detection sensors on board. This sounds wise in theory, but in practice means that losing a single asset can impact multiple missions — and often means that programs are delayed when extra sensors are delayed in delivery, necessitate more testing, or at least require additional time to integrate and test. It’s better to have dedicated, smaller satellites — one for each mission — that you can build and launch quickly.

• Diversification — As we will discuss later, space itself is fragile, which makes overreliance on any particular orbit unwise. Even if LEO was rendered unusable, we would still need missile warning, comms, and other missions to continue; building hybrid, multi-orbit architectures can make our on-orbit capabilities dramatically more resilient.

• Distribution and Proliferation — Simply launching more satellites, with missions split across many smaller assets rather than single larger assets, meaningfully changes the engagement math to the benefit of the defender. Each smaller asset is both harder and less impactful to destroy, which makes an attacker less likely to bother attacking in the first place.

• Maneuverability — Stationary satellites are vulnerable, so the Space Force really wants to be able to “maneuver without regret.” This complicates targeting, makes anti-satellite missions more costly for potential attackers, and can proactively help mass forces in particular orbits when they are needed. This likely requires developing new propulsion technology; modern electric propulsion is too low-thrust, while storable chemical propulsion gives high thrust but low specific impulse, so it burns through propellant too quickly for frequent and long-duration maneuvers.

• Protection — Active and passive measures are necessary to defend existing assets, as we will discuss in detail later.

Systems like SDA’S PWSA and RG-XX are evidence that America is moving in the right direction, but fielding these new capabilities will take time and discipline. You not only have to create the right programs with the right requirements and to overcome the organizational inertia to change nothing at all, but you also have to resist the allure of higher power, more sensors, more capabilities, more delta-v, and the like. It’s very hard to argue in favor of a less capable system, but often that is the right choice: adding capabilities adds cost and complexity, for one, but most importantly, it adds risk to the system and risk to the schedule. Proliferation is still the right solution: we need more satellites doing more independent missions to make the math favor defenders, not attackers.

That said, exquisite satellites will always have their place, and we need to defend them. There are no publicly acknowledged “protect and defend” satellites on orbit today, but it is public information that the USSF has a Delta-level organization responsible for orbital warfare — and building new technologies to serve that Delta will be important over the years to come.

Finally, it is worth noting here that a distributed, proliferated space architecture is one in which autonomy is of paramount importance. Satellites will need to be able to react autonomously to identified threats, whether from intentional attacks or inert debris — so we expect to see technologies like “dogfighting” become more prevalent in the near future, and are highly supportive of efforts to modernize space operations.

Ways: Dominate the cold war

Perhaps the simplest way to understand what’s happening in orbit today is to model it as a Cold War.

In a cold war, each combatant wants to jockey for advantage without making the conflict turn hot. This can include either avoiding acts of war, engaging in acts of war under covert cover, or selectively revealing the existence of exquisite technologies that deter adversaries from wanting to start fights. In a space-based cold war, covert action is either exceptionally difficult or exceptionally easy: it’s next to impossible to disguise an anti-satellite missile attack or a maneuver to bring one satellite close to another; it’s similarly difficult to find the origin of a cyberattack, a satellite-blinding laser, or plain-old RF interference. As a result, modern space war is full of grey-zone tactics that are either hard to attribute or hard to classify as outright acts of war.

Electronic warfare and radio-frequency-enabled cyber operations are grey-zone tactics already central to modern combat. The front lines between Russia and Ukraine have been called the “most signal-dense combat zone in human history,” with both sides fielding extremely powerful GPS jammers and other EW weapons. In the United States raid to capture Nicolas Maduro, we rolled out the nearly-forgotten EA-18G Growler, a jet that carries large EM sensing and jamming/spoofing payloads. Similarly, the United States experimented with anti-satellite lasers as early as 1997, and China wants the world to think it’s doing the same today: China repeatedly opened a retractable roof to reveal laser gimbals precisely when they knew foreign satellites would be passing overhead.

Lasers are capable of blinding (“dazzling”) satellite sensors without destroying the satellite itself; hacking is capable of disabling ground equipment or satellites without in-space explosions; jamming and spoofing are similarly in a grey zone, temporary effects that deny but do not destroy space assets. These tactics seem to be the primary way in which space war is being waged today; they are meant to achieve military goals without escalation and without generating debris.

We should not forget, however, that satellite systems extend far beyond space vehicles — including ground infrastructure, secure access to spectrum, cybersecurity, regulatory clearance, and more — and these often-neglected components are critically important both to fielding new capabilities and to winning the cold war.

As a case in point, consider Raytheon’s GPS Next-Generation Operational Control System (“OCX”). The goal of the program was to make a ground system that could support a jam-resistant waveform (“M-Code”) to fight through the ever-present jamming of a space cold war. OCX was kicked off in 2010 and was expected to be completed for $3.7 billion by 2016. Sixteen years and $6.3 billion later, the program was cancelled because Raytheon was unable to deliver. The great irony — and Raytheon’s shame — is that the M-Code waveform is successfully on the new GPS satellites, but it can’t yet fully operate because of OCX’s failure to deliver ground software.

This is far from an isolated incident: satellites are sexier than support systems, so everything beyond or downstream of the space vehicle often ends up neglected, understaffed, underbudgeted, and burdened with accumulated technical debt. We could share stories of cybersecurity failures (like the 2022 hack of Viasat’s satellite network) or testing failures (like that of the Soviet lunar program), but a far less commonly discussed issue are the cold war tactics currently being used in the halls of the International Telecommunication Union.

The ITU is the United Nations agency responsible for global radio spectrum coordination. Its massive set of Radio Regulations spans thousands of pages and is maintained much like ancient religious texts, with lifelong scholars who dedicate their lives to ensuring this or that word changes or does not change. The Radio Regs simultaneously have immense influence — as there are many nations around the world that follow them to the letter — and no teeth. China and Russia ignore the rules and have been accused of registering “paper satellites,” meaning fake satellite filings, with the ITU to block other, law-abiding countries from deploying real satellites to key orbits. China, in particular, seems to have realized that if they control the ITU politically, they can both act with impunity and shape the rules to make life difficult for other countries. China held the ITU presidency for years, and is set to host the quadrennial ITU World Radiocommunication Conference in China in 2027.

We need to fight this cold war more effectively, and that means dominance in the domains listed above — and preventing the war from going hot.

Ways: Defend the domain.

On July 9, 1962, the United States detonated a thermonuclear warhead in Low Earth Orbit.

In a fraction of a second, 10²⁹ electrons were released into the magnetosphere at near-light speed — generating an electromagnetic pulse powerful enough to blow out street lights and knock telephone systems offline in Hawaii, nearly 900 miles away. The detonation ultimately created a new, artificial radiation belt that persisted for years and caused the premature deaths of more than one-third of all satellites operational at the time. An unreleased paper from UC Berkeley suggests that this is actually an underestimate of the destruction that could result from the use of a modern nuclear weapon in space today — the preliminary number that we read is that a single nuke could destroy 93% of all satellites, across all orbits.

The United States and the Soviet Union banned atmospheric and exoatmospheric nuclear testing in the Nuclear Test Ban Treaty of 1963, but — recklessly and unbelievably — Russia now seems ready to ignore the Treaty and deploy a nuclear weapon on orbit. In February 2024, the White House publicly confirmed House Intelligence Chairman Mike Turner’s warning that Russia is developing a nuclear anti-satellite capability. NSC spokesman John Kirby acknowledged the program; Assistant Secretary of State Mallory Stewart later told a CSIS audience that U.S. intelligence has tracked Russia’s pursuit of this capability for years and now assesses its progress with new precision.

This is a massive problem for obvious reasons, and the risk is not contained to Russia alone — nine nations have nuclear weapons today, all of them have the ability to loft those weapons to orbital altitudes, and some of them have very limited space infrastructure, which means no risk of friendly fire. The potential impact of a nuclear detonation on orbit is hard to overstate; for a rough comparison, imagine if it were possible to light the entire ocean on fire, thereby destroying every ship in the sea. There are plenty of reasons why nations wouldn’t do such a thing — just as there are reasons they don’t use nuclear weapons in conventional wars today — but nuclear weapons still hang like a Sword of Damocles over commercial and military operations in space.

Conventional munitions and anti-satellite weapons also have the potential to cause significant damage on orbit. As we discussed above, when a single object in space explodes, it becomes many smaller objects, each with a new and diverging orbit. On average, these new orbits will still have the same orbital velocity as the original satellite, so they will not fall out of the sky, but instead will stay in a dispersed cloud at roughly the same altitude as the original blast. Introducing thousands of random orbits into a dense orbital shell is obviously dangerous. If one satellite explodes, other satellites will have to diverge from their planned orbits at best and may get hit by debris at worst, which could cause a runaway effect.

Simply, space is fragile. This is perhaps the single most salient fact that distinguishes space from other warfighting domains. It is scary, but true, to say that many nations around the world have the ability to hit “reset” on humanity’s exploration of near-Earth orbits.

This is why we chose to elevate the protection of the space domain to a “way” within our framework. We must defend the orbital domain to win a space war — and to do so will require strategies and tactics that look completely alien to other, terrestrial forms of combat.

Gen. Chance Saltzman, then the highest-ranking officer in the U.S. Space Force, framed the problem simply in his 2023 “Competitive Endurance” white paper:

“Historically, domain control on land, at sea, and in the air has been achieved by the threat or application of overwhelming destructive military force. Applying this same approach to the space domain would create hazardous debris that jeopardizes the ability of the Joint Force to exploit the very space capabilities these drastic measures seek to protect. It would also disrupt the activities of civil, commercial, and scientific space operations worldwide. Consequently, unlike in other domains, our concept for domain control in space cannot rely on overwhelming destructive force.”

The way we see it, if anyone successfully fires an anti-satellite missile, everyone loses.

Once things go hot, all hell has broken loose, you will likely see a conflict bridge to the ground — with spaceports and ground stations at most immediate risk — which will therefore make it extremely difficult to stop anywhere short of orbital reset. Even if you win every “dogfight” and shoot down every enemy satellite, you have created an exceptional amount of debris and you are still vulnerable to ground-based counterattacks.

This means that the capabilities you need to defend the domain are two-fold.

First, you need defensive technologies that can intercept anti-satellite missiles or otherwise render them ineffective. This could come from an active denial system like Golden Dome, or left of launch operations, propulsion powerful enough to dodge missiles on orbit, or dedicated protect-and-defend satellites — but it has to come from somewhere.

And second, you need offensive space capabilities so capable that they convince every other nation they are woefully overmatched in conflict.7 We should endeavor as a nation never to use these capabilities, but they should both exist and be publicly acknowledged. The world must know that you cannot wage a space war against the United States and win, which means developing technologies capable of taking out many adversarial satellites in one fell swoop — without generating significant and cascading debris.

These offensive technologies will take many forms depending on the mission they are intended to deny, degrade, or disrupt. Directed energy weapons could be used to destroy electro-optical, infrared, radar, or other sensors, or to create precise structural damage to functionally destroy a satellite. Concentrated RF energy can damage radio receivers for communications satellites. And any number of reversible and irreversible effects are possible if you can maneuver your offensive satellite close to a target: robotic manipulation, deployment of chemicals, and low collateral-effect kinetic weapons are all technically feasible.

The trouble, of course, will be making the attacker-defender math work. If you need one attacker for every defender, not only would an offensive attack be costly, but also it would be extremely hard to disguise — you’d have to match your orbits to those of your adversaries, which takes time. The even harder math to close is the propulsion budget. At the end of the day, an orbital “dogfight” will be won by whoever has more delta-v. Better spacecraft control and autonomy can help, but more often than not at the margins between matched systems — and because we’re still in the early innings of maneuvering without regret, big potential technological advantages are still out there to be seized.

ENDS: How to win a space war

The days of space as an idyllic, peaceful reserve for science are decidedly over.

Russia tested an anti-satellite missile in 2021, hacked a commercial satellite in 2022, jammed GPS throughout the Ukraine conflict, tailed an NRO satellite with a potential “counterspace weapon” in 2024, and is intimidating a commercial imaging satellite in 2026. China hauled a dead satellite out of geostationary orbit in 2022, practiced “dogfighting” maneuvers in 2024, and deployed an electronic warfare satellite to orbit in 2025.

In 2021, Gen. David Thompson said that Russian and Chinese systems were interfering with military satellites “every single day.” And in 2025, Gen. Stephen Whiting told the Senate Armed Services Committee that China is building counterspace capabilities at “breathtaking pace“ to deny U.S. and allied access on demand.

Space is now a warfighting domain, and the only remaining question is how to win a space war.

Most imminently, America can focus on developing the means of winning the war. A nation that can launch more mass, more often, from more places, and reconstitute faster after losses cannot be easily denied access to the domain. Commercial satellite manufacturing is the way we fill those rockets with useful capacity, build wartime depth, offer valuable services to consumers, and make the lives of our adversaries more difficult on orbit. Proliferation, disaggregation, diversification, maneuverability, and protection all change the basic math of attack and defense for our military systems. They make it harder for an enemy to know what to hit, and less likely that any single successful attack will create a decisive military effect.

The ways to win the war matter just as much. Dominating the space cold war — across electronic warfare, cyber, spectrum, ground systems, regulation, and more — is how we fight every day below the threshold of open conflict. Defending the domain is how we prevent that conflict from crossing the threshold where everyone loses.

Together, those means and ways lead to the ends we desire. Freedom of maneuver means America and its allies can get to orbit, move in orbit, reinforce in orbit, and continue operating even in the face of threats. Secure lines of communication mean that the joint force can see, talk, navigate, target, warn, and coordinate, even when an adversary most wants to blind and isolate them. Deterrence means that Russia, China, Iran, and any other hostile power looks at American and allied space systems — commercial and military alike — and concludes that coercion will fail and attack will invite consequences they cannot accept.

This is the central lesson of modern space warfighting: fragility and weakness invite coercion; resilience and strength deter it. If America can launch at unmatched scale, manufacture satellites at unmatched volume, distribute capability across thousands of assets and multiple orbits, fight and win the cold war, and credibly defend the domain without destroying it, then we will win.

The stakes are as high as they sound. Russia has shown a willingness to threaten commercial satellites and flirt with weapons that could imperil the entire orbital environment. China is racing to build sovereign space power and to shape the rules, or ignore them, when doing so serves its interests. The United States is the only nation capable of protecting space for all mankind. If space is to remain usable for commerce, science, exploration, and the defense of the free world, America must know how to win a space war — and must be strong enough that no adversary doubts that it would.

Special thanks to our friends in the space industry who helped improve our thinking and shape this piece: Adam Cohen, Andrew Reddie, Charlie Horowitz, Mike Grace, Ian Chun, Payam Banazadeh, Rishay Jain, Nato Saichek, and Zak Kirstein.

Subscribe now

This material is solely for educational purposes and is not investment advice or an offer of investment advisory services. This material should not be used as the basis for an investment decision. a16z is an investor in SpaceX and Cursor through its managed funds, and thus has a financial interest in the company’s performance and future prospects. In particular, a16z benefits if the company grows in value; and a16z funds will receive any customary dividend payments in connection with their status as a shareholder of the company. However, a16z is not being compensated by SpaceX or Cursor for this material.

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

Healthcare runs on the office landline.

Behind every doctor is a room you never see: staff in headsets, scheduling patient visits while sitting on hold with insurance companies, typing into the electronic health record (EHR) one field at a time. None of it is medicine, and all of it decides whether you get seen and whether your doctor gets paid. Confirming your insurance coverage before a visit can mean a 40-minute phone call. Get one field wrong and the appointment doesn’t clear, the claim goes unpaid, and you land in collections months later – for care you assumed was covered. Decades of patient portals later, the work still happens over the phone.

That’s the layer Prosper rebuilds.

Prosper is an AI-native platform that runs the voice-heavy work for clinic operations end-to-end: appointment scheduling, insurance eligibility and benefits verification, and patient billing. What sets it apart is the breadth of capabilities. Existing voice automation is built for simple operations with rigid rule trees that ultimately allow you to only automate a small portion of this work. Prosper is built for the combinatorial complexity of large specialty care groups: 10+ locations, 100+ appointment types, 1000+ insurance plan combinations. It handles patient, provider, payor, pharmacy, and pharma conversations in one system, writing structured data into the practice’s tools, navigating payor portals when there’s no API to call, and escalating work to human staff only when an AI agent can’t finish it. What comes out the other side, much sooner, is a financially cleared patient appointment.

Last generation voice agents were never truly end-to-end service providers. Each tool picked a lane and stopped there. But healthcare administrative staff don’t work like this – they’ll only trust a platform that can be their first line of defense for any phone call they handle. We have been looking for a team building this; we have been looking for Prosper.

In under two years, Prosper has gone from launch to a growing roster of leading specialty practices across the country, and signed channel partnerships with major EHR and RCM vendors that put it in front of hundreds of thousands of providers and millions of phone calls. Revenue is compounding fast, and the value shows up where it counts: a jump in scheduled and paid appointments, without further burdening clinic staff. In a market full of voice AI demos, Prosper is one of the few converting interest into business-critical, sticky deployments.

At the heart of all of this are co-founders Xavier and Josep, who have felt every pain point Prosper solves. Xavier grew up inside his family’s medical practice before going on to run large-scale call center operations at Handy (acquired by Angi, NASDAQ: ANGI). Josep is a former MIT CSAIL researcher who led product at CoverWallet (acquired by Aon, NYSE: AON), where he digitized distribution for some of the largest insurance lines in the U.S. They understand provider operations, call center logistics, and insurance company workflows from the inside—and they’ve assembled a deeply technical team with whom they are deploying Prosper into the world.

We led Prosper’s $30M Series A because fixing healthcare’s phone call crisis is how you rebuild the administrative core of healthcare globally. Every scheduling, eligibility, and billing conversation is a consequential decision point—care happens or it doesn’t, the provider gets paid or doesn’t—and Prosper is the essential infrastructure at the center of it all.

Go forth and Prosper.

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 a16z New Media Fellowship is back for round two! We’re excited to introduce a brand-new cohort of 65 generational storytellers, podcasters, X-poasters, and substackers at the frontier of new media.

The fellowship kicks off next week and will run through the summer. Here’s a glimpse into the group:

• Noah Smith, Economist and writer, Noahpinion

• Chris Painter, President of METR

• Aly Pavela, Head of Community Communications at Stripe

• Rohan Kumar, Head of Vertical Platforms at MrBeast

• Hannah Broadhurst, Head of Entertainment & Content Marketing at Amazon

• Danielle Meister Lurie, Head of Global Comms, Social Media, & Film at Zipline

• Thijs Simonian, Co-Director of Stanford TreeHacks

• Ian Curtis, Technical Staff Member at World Labs

• Madison Kanna, Software Engineer at Baseten

• Grace Werner, VP of Growth at the Foundation for American Innovation

• Katherine Dee, Writer of default.blog

• Austin Schumacher, prev VP of Marketing at UTA, VP of Brand at Coinbase, Head of Brand Marketing at Lyft

• Anu Atluru, Writer of Working Theorys

• Adel Wu, Growth at Reducto, prev SWE at Meta

• Christian Catalini, Founder of the MIT Cryptoeconomics Lab

We hope you’ll join us in celebrating this group for everything they’ve accomplished so far - and in following their journey ahead!

If you’re interested, you can also sign up here to stay in the loop about future fellowships and new media opportunities.

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 excited to announce a16z’s Talent Engineer Fellowship: an 8-week cohort for the people building at the forefront of recruiting. - Danco

A16Z New Media is one year old! And we’ve had several portfolio companies (and other venture firms) ask us how our team works, how we built it, and how they might build their own. Here is what we share with them:

If you’re ever pitching a16z, at some point we turn it around and do “the Bear Hug”, as our competitors call it: we pitch you on all of the resources and power and networks the firm will put behind you; maybe you’ll have dinner with Marc until 2 am, and close out the restaurant while he regales you with detailed knowledge of your domain. And then we’ll follow up the next morning with 20 tangible things we’ll help you with in our first week of working together.

“New Media” has become one of those critical things we advertise in the Reverse Pitch, as a way a16z helps you build your brand, reach key people, and win the battle for attention.

What do we actually do? It won’t capture everything, but at first estimate you can think of a16z New Media is “go-direct as a service” for startups. We have four primary motions:

• In-house creative (Video, design, editorial, from legends of the craft)

• Owned channels (Our X account, Substack, podcast feeds, and more)

• Full service from a dedicated team, from pitch to launch and beyond

• A big network of people, with a bit of structure, software and AI to help leverage it

Let’s go through them:

1. We have in-house creative, notably for video and for writing.

Ben & Marc are both genuine aficionados of media history, and they can talk to you for hours about how media has evolved over their careers, from a highly polished, careful exercise (basically, “there’s one job, and it’s don’t say anything that gets you in trouble”) into a more authentic, unfiltered practice where being interesting is what matters. When we talk about “New Media”, we mean this new world where to win you must go direct, and to go direct you must be interesting.

Launch Videos are a new (and we think, enduring) format for founders to make a statement at a critical moment, that’s built for this new media environment where you don’t have to ask permission to get on someone else’s show or op-ed page: you can just make your own statement, however you like, and share it widely.

“Tell your whole story, in a couple of minutes” is a great constraint, and we love the challenge. We have a video team in-house at a16z that produces them, on timelines that most contract agencies will tell you is impossible. We know how important speed is for founders! Not only because everything at a startup moves at a fast pace, but also because creative work is about catching lightning in a bottle, and really nailing the energy of a moment. When video production timelines stretch into months (which is usually the kind of lead time an agency asks for), you’re probably going to lose some of the magic.

Founders love it! They love the ambition, and the pace, and the final product. Here’s AI legend Dr. Fei-Fei Li, on our work with World Labs:

Meanwhile, we also have an editorial team that helps founders (and our investing partners) think through written ideas, capture them as essays, and serve as editors and thought partners. Many founders we work with are outstanding writers already, others want to work at the craft and improve. Either way, founders really appreciate having someone outside of their company they can talk to, and who can help them frame their company in a way they may not have considered before. This can be helpful for all stages of companies, even really big ones.

We love working on all kinds of essays. One founder might have a really big and clear idea that they want to share about how AI is changing work; another might be focused on nailing the written “why” for why their company exists. The right written piece, stamped with the endorsement of a firm like a16z, can make a real difference in the trajectory of a company, and in the “legible interesting-ness” of a founder and their product.

It’s important that founders are interesting! Because venture capital, and startups in general, are an exercise in preferential attachment. People have to eagerly want to attach themselves to you, your idea, your product, your team, your vision. Of all of the limitless opportunities out there, people have to preferentially attach to you. That’s a tall order right now, when limitless opportunity is all over the place. That’s why our in-house creative teams exist in the first place: to help you win the battle for preferential attachment and attention, by bringing out what’s already interesting about you.

Now, this in-house creative work is similar, at surface-level, to the kind of work an agency does, although there are world class agencies out there that we recommend and partner with. But there are some important differences though, and one is the kind of people we have on these teams. We honestly believe that we can hire much better people than your average agency. The job is more interesting, and the work is more incentive-aligned.

Founders really appreciate that they’re talking to people like Gaby Goldberg, Alex Danco, James Reina, or others they feel are “on their level”, in their ability to follow the plot, understand the subject matter, and deliver on their nuanced goals. We love hearing from founders, “I’ve read your Substack for years! I’d love to write something with you.”

2. We have owned channels, with our brand and reputation on them.

New Media is about being interesting and saying what you really mean, versus Old Media where you were mostly trying to stay on message and keep out of trouble. That’s because, in the old world, you were entirely dependent on other people’s channels to get your word out. There were a select set of newspapers or TV shows that could legitimize you, and if you ever made them mad, you were in serious trouble.

Even worse, the old line used to be, “Either you’re journalism, or you’re PR Flack.” As a founder playing the media game, you were either one or the other. There was no category in the middle, if your goal was, “I want to tell my story and be authentically understood.”

Now, “New Media” as an open playing field where anyone can tell their story is a great thing; but distribution is obviously still important in this new world:

• We have our X account, with a million followers

• We have a daily email newsletter with a quarter million subscribers

• We have the podcast, with a million downloads a month

• We have emerging (for us) channels like Instagram, where we have 160K followers and growing fast

There are two reasons why this is important. The first one is straightforward: when you work on an essay or a launch video or something else with us, we can guarantee you distribution. When you work with other agencies and investors, they can’t really promise you reach, in any real sense. (But some do anyway, and we’ve all seen launches with engagement metrics that don’t pass the smell test, if you know what we’re getting at. That isn’t a game we play at a16z. We don’t need to, and we’d never jeopardize your reputation that way.)

The second reason is more interesting, and has to do with what distribution even is, in a world where the important brands are people now, as opposed to companies.

There’s an important New Media principle that signal is what’s important, and people confer signal, not brands. Founder, partners, researchers, power users, and great X accounts are all characters who make an idea legible and credible. Our friend Anu wrote a piece on New Media the other day, “New Media is Insider Media”. We’ll quote a bit of it here:

“Insider Media is media for insiders, by insiders. Both creator and audience belong to the same in-group. The public gets to be in the room but is never the audience being catered to. The topic is less important than the audience whose opinion matters. It’s a closed-room broadcast where principals talk to principals while fans, aspirants, and adjacents watch, learn, and report out.

Insider Media is the opposite of journalistic media—outsider media—made for outsiders by outsiders. Outsiders will call Insider Media an echo chamber or bubble, but the proximity is very much the product. This tension is inevitable in any world where status compounds and access is currency.”

Anu is right on with her definition here, but there’s an important way in which we think about it a little differently, which is that “being in-group” is more positive-sum than people think. The whole point of tech is that it’s a growing pie, there are more interesting people and ideas and storylines each year than in the previous year.

Our goal with our channels, and all of the reach they can confer to you, is to help lots and lots of founders become high-signal, because we don’t think it’s a fixed pie. This is completely consistent with everything a16z has ever stood for: we’re about helpful-maxxing, not gatekeeping and withholding our distribution only for the very select. That’s why it’s so important that we own our own channels: we don’t have to “measure out” founders’ signal or access, we can simply share more, all the time, because no one can tell us no.

3. We have a team of high-touch service people who figure out how to best help founders, and then get it done, concierge-style.

We told you about the ‘reverse pitch’, where in fundraising meetings, we turn the chairs around and we pitch the founder. Founders really love this, because it actually firms up what working with a16z looks like. It’s a good introduction to how a16z is different from other firms: we don’t just have one marketer who writes you a one page plan that you subsequently ignore. We pitch you on a whole bunch of things that we are then on the hook to actually deliver!

After the close, our “portco services” team is how we keep that commitment. The team figures out what the founder is actually trying to accomplish and then stitches together the right capabilities across the firm. For a launch, that might mean securing a key video asset, getting a great piece of writing written on time, lining up press, and making sure all of it lands on schedule. Oftentimes, this includes traditional comms! The right story in the right publication can still matter a lot. But the press hit works best when it is part of the larger system, and undergirded by the founder’s own framing. Our comms & marketing partners at a16z are seasoned veterans at this, and they know how to get traditional media working for you without compromising any of the “go-direct”ness.

These launches aren’t just about looking good on the timeline; they drive measurable business impact. We’ve seen startups gain tens of millions of dollars in contract value, growing massive waitlists for their products, doubling their hiring pipeline, and other really tangible things.

For many of our technical founders, New Media is a new muscle. Most founders don’t fully believe the power of it until they see its impact for themselves on their business. Increasingly, we’re thinking about our portfolio services team’s role as helping founders see an “existence proof” of how New Media can help them book meetings with hundreds of top customers, or drive hundreds of inbound job applications, or catalyze their next funding round.

Once founders see this impact, they often can’t get enough. So we’ve seen enormous demand to hire “data storytellers”, “heads of socials”, and “technical long-form writers” after working with our team. We’re now scaling up our efforts to help portfolio company CEOs build their own internal New Media capabilities.

It’s tempting to read all of this and assume it only works for the kinds of founders who are already resonant: the Brian Armstrongs and Palmer Luckeys of the world who were already going to win the attention war. But we have an equally good time working with founders who have built something important but need help fighting the battle for attention; not because their work isn’t compelling but because being compelling on a timeline is its own hard-to-cultivate skill. It helps that we can employ all of the format advantages that New Media offers as a category: the “one hour podcast, one minute clip” format has proven effective at bringing out what’s interesting about people with something to say; even if they aren’t polished or on-message.

In this sense, “New Media” as we practice it isn’t really a type of content; it’s a type of packaging that’s designed to help you be interesting. For example, it might mean:

• The “longform conversation / essay + short form clips” barbell, or other formats designed to highlight compelling ideas with messy context rather than crisp on-message

• Guerilla-style amplification, like mastering the QT format to get an essay the algorithmic traction it needs to break through: doing whatever it takes to deliver the longform idea to the select audience that should hear it

• A well-crafted press campaign that can get you a media spot because you don’t need it, but they want you anyway

This is why having a portfolio services team that coordinates everything matters a lot. The point isn’t to “ship a bunch of content”, the point is to highlight what’s already interesting about the founder and their product, make it resonate on the timeline, and gain signal within the group of people they care about. After all, our goal isn’t for founders to have one great launch day and that’s it - it’s for them to build great brands that compound over time. Founders and their marketing teams can learn on their own how to make great content (just by being interesting!); where we teach them a useful skill is how to consistently package it, so that their brand will compound and strengthen for the long haul.

4. We have a giant network of people, who know and trust each other, and a little sprinkle of software and AI, that helps us put this network to work for our founders

The battle for people’s attention and the quest to be interesting isn’t only fought on the timeline; it’s also fought in person, in small groups and in perfect introductions. Preferential attachment for the important things is ultimately something that happens between people who trust each other. Which brings us to our fourth team:

We have a pretty great dataset of “who is everybody in tech.” We’re a big venture firm, we take a lot of meetings, we’ve meticulously held onto everything we’ve learned over years of being in the business. And now, finally, we can really do something with that potential.

The most obvious way that we can help founders is to help them hire. And hiring isn’t just about blasting out job posts; you need a network of people who trust each other, can surface the right signals, and care about who they put forth. (But the blocking and tackling of job posting still matters. Our Build Newsletter, which features lovingly curated job postings from the portfolio, is a founder favorite.)

You also may have seen our New Media Fellowship, which launched last year as the first of many talent & hiring fellowships run by this team.

Fellowships help you meet really high-quality people who share your craft. We bring together the best operators in fields like new media, design / growth / forward-deployed engineers, and more, and host dinners, run groupchats, and sometimes invite 90+ people to Carmel for a couple of days to learn from each other and share wisdom. One thing you might notice about the fellowships themselves is that they all containerize roles that never really existed before until a few years ago. As our tools change, our jobs change too.

We have three goals with these fellowships: One, to help founders hire great talent, and level up their own leadership bench. (That one’s obviously useful, and everybody asks for it!) Two, to build relationships early with people who might start their own companies, so that when it comes time for them to pitch us, we have a rich set of experiences with them.

The third one is our favorite, and actually the most meaningful thing for the community long-term: it’s to help people practicing these new crafts meet one another, and learn how their peers are practicing the craft and move through their careers. We feel like this is a really valuable service we can provide, if we’re helping thousands of people make meaningful connections, and maybe find that perfect job with a portfolio company.

To pull this off (and much more), we host and facilitate a lot of events. Our events team is genuinely great, and they pull off everything from small curated get-togethers to big statement bashes flawlessly. There’s a magic to doing them well, and doing them often.

There’s a lot more coming soon on this front, including similar communities for executives, emerging investors, and something very cool on the “help make hiring better for everybody” front that we’ll announce soon.

So, what is New Media?

The funny thing about New Media is, it isn’t actually new! The world has always had interesting people who are expansive but not polished; fascinating but frequently off-message, whatever. There have always been niche media formats where these people can shine. What’s different now is that the world changed, in many ways that reward being interesting over everything else. So now there are ambitious people, and very profitable business models, pursuing New Media as a deliberate offensive business strategy.

Over the past couple of years, many of the startup marketing and comms job postings have suddenly become “New Media branded”, not in the sense that any of the tasks are new (the jobs themselves are pretty familiar to anyone who’s ever worked at an agency, run a social media account, or been a staff writer somewhere.)

Are these new jobs? Well, yes and no. They’re not new jobs in the sense that, if you look at what all of these New Media folks are doing, it’s pretty classic marketing and comms work:

What’s different, again, is the packaging and the branding of the overall effort. It’s meant to attract interesting people who are subject matter experts in something, internet-native, and eager for a lot of responsibility. Many people on the New Media team were doing something else before—they were founders, investors, product managers, or just happened to love posting, blogging, or making videos; but didn’t know there was a role that was ambitious enough for them to deploy these talents.

The roles can feel hard to describe until you’ve seen them in action. We actually see this happen all the time: once we work with a portfolio company on a project, especially when we embed someone there, they immediately try to hire someone to do that kind of work once the project wraps up, now that they have a clearer idea of the job and of its value. (And we’re happy to help them find great people, through our network and particularly through the New Media Fellowship!)

When a16z was founded in 2009, its big bet was that a venture firm should be built like a network (or like CAA) rather than a collection of solo investors (basically Firm > Fund), and that the operating team’s job was to construct platforms that, in a Harvard Business School study framing of the firm, harnessed network effects to deliver compounding value for everyone on them. The marketing team courted journalists and cultivated relationships with PR agencies; the talent team counseled engineers. The wager was that goodwill, relationships, and reputation accrue, and that the firm patient enough to accumulate them would eventually operate from a position that would be difficult to replicate Thirteen years after the publication of that HBS study, that bet has decidedly paid off.

A16Z New Media is the newest expression of this strategy. When you help founders directly with the work of being interesting constantly over time, that work compounds in value: in founder goodwill and portfolio outcomes, in follower growth and channel reach, and in tacit knowledge of how to launch a company really well. To make this work, you need quality and quantity. Writing one great essay every once in a while, or producing one charming video, isn’t going to cut it. You need to be relentless if you want a media asset to actually compound, not just make one nice launch video. To paraphrase Elon, “Anyone can make a nice car once. I’ll be impressed when you can make a million cars.”

Fortunately, we have one of the most incentive-aligned business models in the world to help founders with New Media. If it’s helpful for founders, we’re going to do it, and we’re going to do it a lot. That’s what a16z has always stood for, and New Media is one of the things we’re most excited to pitch you, when it comes time for that Bear Hug.

Subscribe now

This material is solely for educational purposes and is not investment advice or an offer of investment advisory services. This material should not be used as the basis for an investment decision. a16z is an investor in SpaceX and Cursor through its managed funds, and thus has a financial interest in the company’s performance and future prospects. In particular, a16z benefits if the company grows in value; and a16z funds will receive any customary dividend payments in connection with their status as a shareholder of the company. However, a16z is not being compensated by SpaceX or Cursor for this material.

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 launching the a16z Talent Engineer Fellowship: an 8-week cohort for the people building the next operating system for recruiting.

Here’s the problem. AI has made every part of recruiting noisier at exactly the moment companies need sharper signal. Resumes are easier to polish. Inbound is harder to parse. Outbound is easier to spam. And the best candidates, more in demand than ever, have mostly stopped looking.

The old recruiting stack still matters: better sourcing, better process, better relationships, better closing. But it no longer differentiates, because everyone now has the same tools. The advantage is shifting to teams that build their own.

Sales already went through this. When Clay and its cousins arrived, a new role emerged — the GTM engineer — and the companies that figured it out early ran circles around everyone still grinding through manual SDR work. Recruiting is at the same inflection point right now.

We call the role that comes out the other side the Talent Engineer.

What Is a Talent Engineer?

A talent engineer is an engineer who recruits — not just a recruiter who codes.

The distinction matters. This isn’t a recruiter using AI tools, it’s not a recruiting-ops person keeping the ATS clean. It’s the person who gives companies like Cursor, OpenAI, xAI, Ramp their unfair advantage in finding the best people fastest, building systems to identify/attract top talent.

You’ve probably met the proto version. Every good startup has one: the engineer who never had the title, may not have reported into recruiting, shipped real code — and somehow ran the technical screens, interviewed half the company, and got everyone’s best referrals. They were doing talent engineering before there was a name for it.

What they build: sourcing agents, market maps, relationship graphs, candidate-intelligence workflows, referral systems that actually work, AI-assisted technical evaluations, hiring-manager context engines, talent-brand surfaces, and the weird internal tools that make the whole machine smarter.

What they understand: that recruiting, done right, isn’t an assessment problem. It’s a many-to-many matchmaking problem — the right person, the right role, the right moment — and matchmaking is an engineering problem.

If this sounds like you, apply here by July 24. Applications are rolling, and the fellowship begins in late August.

Apply Now

How It Works

• 8 weeks, starting late August.

• Time commitment is flexible — everything is opt-in

What You’ll Get

• A tight network of ~40 AI-native talent leaders, recruiting systems-builders, and talent-infrastructure founders

• An in-person dinner series in SF, expanding to more cities by demand

• Demos, live working sessions, and honest feedback on what you’re building

• Off-the-record conversation about what’s actually working in sourcing, candidate intelligence, evaluation, referrals, and closing

• Access to the a16z ecosystem — including founders trying to hire for roles that don’t have names yet

Who You’ll Be Joining

You’ll break bread with founding fellows and industry leaders like:

• Viet Nguyen - Founder @ Talent Collective, prev. Dir of Recruiting @ Vercel

• Dan McCarthy - Senior Talent Engineer @ Zapier

• Jared Long - Recruiting Ops Lead @ Cursor, prev. Head of Recruiting Systems, Operations, and Analytics @ OpenAI

• Jeffrey Li - CTO @ Paraform

• Madison Morales - Talent @ Anthropic, prev. Talent @ General Catalyst

• Sebastian Cuadros - Head of Talent Engineering @ Fonzi AI

• Amit Matani - CEO @ Wellfound, founded AngelList Talent @ AngelList

• Johnny Sanchez - Head of Recruiting Operations & Technology @ Airbnb

• Danielle Poreh - Clay for Recruiting Lead (prev. Talent Scientist) @ Clay

• Scott Bianco - Head of Talent @ Hebbia

• Umesh Khanna - Talent Engineering Lead @ xAI

• Joshua Gill - Talent Engineering & Operations @ Luma AI

…and many more.

Apply Now

Why a Fellowship

Because the people doing this work are learning faster than any playbook can keep up.

Right now, one team is building sourcing agents. Another is mapping hidden talent pools from social data. Another has redesigned candidate evaluation around live AI-assisted work sessions instead of leetcode. Another turned referrals and backchannels into a real system instead of a pile of Slack messages. Another is building talent-brand surfaces that make exceptional people raise their hand before they’re looking.

Almost all of this knowledge is trapped inside individual teams.

We want to put these builders in the same room: trade notes, demo what they’re building, pressure-test the edge cases — and define what Talent Engineering becomes before the market settles into worse language.

And with all that’s possible to build, the best people in the world won’t engage with an automated screen, and tooling without judgment is just spam with better infrastructure. The talent engineers worth knowing understand that the system exists to earn the human conversation, not replace it. That sensibility is a requirement here, not a nice-to-have.

We’re Looking for Fellows Who:

• Have built or instrumented part of the recruiting workflow themselves — sourcing, screening, evaluation, or anything downstream in the talent stack

• Are engineers who run technical screens and think systematically about how to read candidate quality at scale

• Have real talent judgment: taste for non-obvious people, role fit, timing, and candidate experience

• Operate AI-first, with evidence they’re using modern systems in the work rather than just talking about them

• Can turn messy recruiting judgment into workflows, agents, dashboards, interview loops, enrichment systems, or relationship graphs

• Care enough about trust and candidate experience that better tooling makes the work better, not just louder

• Would make the strongest people in the room feel like they’d met a peer.

You may not call yourself a talent engineer. Maybe you’re an engineering leader, a sourcer, a recruiter, a people-analytics builder, a talent-infrastructure founder, or something that doesn’t have a title yet. Titles matter less than what you’ve built — the work speaks for itself.

If that’s you, apply here. We’re accepting applications on a rolling basis.

Questions? Email build@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

Nicolás Abad’s father was the kind of doctor patients never forget.

A physician in Colombia, he became known as “Telepatia” because of his uncanny ability to understand what was wrong before patients had even finished explaining their symptoms.

He died at 58 from a preventable drug interaction – just days before the release of ChatGPT, which helped usher in a new generation of AI capable of catching this type of risk.

Today, Nico is building the product that could have helped save his father’s life, and that his father would have loved as a doctor.

We’re excited to announce our lead investment in Telepatia’s $33 million Series A.

Latin America has 700 million people, but a severe shortage of clinicians. The region has 43% fewer physicians and 65% fewer nurses per capita than OECD countries on average. Where the United States has spent decades piling software on top of a fractured system and created a new kind of mess, Latin America has the advantage of starting from a clean slate, and the potential to leapfrog the clinician shortage entirely.

Telepatia is building the AI-native clinical platform for Latin America. Its platform combines AI documentation, clinical decision support, and a growing suite of AI healthcare employees—including AI doctors, nurses, and auditors—integrated across hospital systems and trained on clinical guidelines, peer-reviewed literature, and local institutional protocols.

Since launching in July 2025, the platform has deployed across 25+ hospital systems in Brazil, Colombia, and Mexico, reaching 14 million patients. Doctors use it all day. Protocol adherence has risen from 84% to 99% after deployment with physicians recovering 1.7 hours daily – time they now spend with patients. Telepatia has already prevented 60,000 medical errors in real time, allowing physicians to adjust course immediately.

Take it from the customers themselves. “The real test of any clinical tool is whether doctors actually use it,” says José Henrique Dias Salvador, CEO of Mater Dei in Brazil. “Our physicians use Telepatia eight hours a day. We’ve never seen adoption like that.”

The top 100 health institutions in Latin America employ 50% of physicians. This concentration means that a single platform can achieve outsized distribution once it breaks into the top systems. Win the right logos, and distribution compounds in ways that simply aren’t possible in more saturated markets. Telepatia is achieving this.

We are joined in this round by Palantir CTO Shyam Sankar, Rappi founder Simón Borrero, and Nubank founder David Vélez, three leaders who have built and scaled transformative technology in complex industries across Latin America and the world.

Nico’s father had a wall covered with portraits of Nobel laureates he admired – scientists like Allvar Gullstrand and Richard Feynman whose discoveries changed the course of medicine and science.

We asked Nico what winning a Nobel Prize would feel like. He said: earning a place on that wall.

Nico’s goal is to win the first Nobel Prize in Medicine awarded to a Latin American since 1984. As he puts it, “Recent Nobel Prizes in Physics and Chemistry have gone to breakthroughs powered by AI. We believe Medicine is next.”

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

Legend has it that in the early days of DoorDash, the whole delivery network ran on one guy named Steve. An order would come in, and Steve would find the closest driver and text them to go grab it. No algorithms, no dispatch system. Just Steve. He was great at it, until the company grew and there was no way to hire enough Steves to keep up. So Rohan Chopra, an early DoorDash employee, built the software that took the brokering off Steve’s plate. DoorDash didn’t replace Steve. They freed him up to solve other big problems to help the company grow.

What stuck with Rohan afterward was how often he saw the same setup everywhere else: a smart person in accounting or ops or account management holding a critical process together by hand because the tools never caught up. Your best people spend two or three hours a day clicking buttons, pulling reports, reconciling invoices, trafficking the same campaign for the hundredth time. It keeps the business running, but it’s rote, repetitive work and a poor use of their time.

AI was supposed to fix this. But most enterprise AI tools serve as contextless point solutions that companies have to adapt themselves to use, rather than the AI adapting itself to the company.

Steve got lucky. He had Rohan there to build him out of his manual matchmaking job. Most companies don’t have a Rohan sitting around to automate their Steves, and that’s exactly why Rohan started Convey. Convey lets non-technical teams build and run digital teammates: AI workers that take whole categories of operational work off your plate, with no engineering required. You onboard Convey the way you’d onboard a new hire. Share your screen, walk through a process, answer a few questions. After that Convey’s agents own the outcome, running inside your systems and asking the right person for help when they get stuck. This allows human employees to become 100x operators – someone who hands off the grunt work and spends their time on the things only they can do.

What makes Convey’s agents work in production is how they are provisioned and constructed. Most agents run a fresh prompt loop every time and wing it when something unexpected shows up. But you don’t run a business-critical process on a vibe. Convey compiles what it learns into real, versioned, testable programs, so it does the same thing reliably across hundreds of thousands of runs and flags a human when something is genuinely off. Each teammate has its own identity and permissions and lives inside your security setup instead of working around it.

The results they are generating are staggering. Convey teammates have already done over 1.1 million hours of real work, not demos or pilots, inside massive companies like NBCUniversal, TelevisaUnivision, Unity, Samsara, ChargePoint, and Faire. One large streaming service handed off reporting and ad ops workflows and got back 450+ hours a week. Savoya used Convey to lift EBITDA 40% year over year and is on track to save 10,000 hours this year.

Special products come from special groups of people – and the people behind Convey are as good as it gets. Rohan, Will, and Diego have been best friends since Stanford. Will and Diego sold their last company to p44 around the same time as Rohan was leaving DoorDash, and the timing finally aligned to start a company together. Anyone who has worked with one of Convey’s three founders will tell you to not miss the opportunity to join their team. That same energy is what made us so excited to invest.

We’re thrilled to be leading Convey’s $38M Series A. If your company is buried in repetitive work - reach out to the team at Convey. And if you want to help build the agentic employees reimagining how work is done at the enterprise, they’re hiring!

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’s post is by Ben Liu, co-founder and CEO of Formation Bio. It’s one of my favorite long reads in a while.

AI will transform pharma, but not evenly and not all at once. Some parts of drug discovery and development are already software-like: faster, cheaper, iterative, driven by model improvements. Others stay bound by the physical world, such as recruiting patients, dosing them, and waiting years for an endpoint to read out. The sequence in which these frontiers move determines where value concentrates and where it erodes.

Building an enduring company in the age of AI is hard precisely because advantages can dissipate rapidly as models improve. Each major release can redraw the line between what was previously a standalone product (or even company) and what a frontier lab now offers natively. Companies that looked durable in one generation get exposed in the next, not because they executed poorly, but because the capability layer moved forward and absorbed what used to be their edge. We have watched this play out in other categories and our industry is not immune.

The companies that win will predict the uneven frontiers correctly, absorbing each model improvement as a tailwind rather than a headwind. Our view: discovered drugs are becoming abundant while clinical development remains the binding constraint — so the durable position is built around that bottleneck, not in the path of advancing discovery models. Here is how we see it playing out, and why.

Discovered drug candidates become abundant

The first major shift is that drug discovery is becoming dramatically more efficient.

This trend began well before AI. Over the past decade, the industry has become much better at generating plausible therapeutic assets. The number of discovered drug candidates has roughly doubled, while the number of approved drugs has remained constant at around 50 per year.

Figure 1. The pipeline doubled. Approvals didn’t. FDA CDER approved 50 novel drugs in 2024, consistent with the general recent pattern of annual novel drug approvals clustering around several dozen per year rather than scaling with the total R&D pipeline. Sources: Citeline, Annual Pharma R&D Review (pipeline); FDA, Summary of Approvals (approvals).

That gap reveals a critical truth: discovery has not been the primary bottleneck — development and clinical trials have. In addition to unrealized commercial value, the cost of this gap is that treatments never reach the patients who need them. In an era of drug abundance, discovery will likely become less of a constraint than development. To fully harness the promise of AI for patients, we must solve the drug development bottleneck.

Markets already price this distinction. A discovered drug is worth relatively little before meaningful clinical validation, and dramatically more after convincing human proof of concept. Preclinical discovery deals often carry upfronts in the tens of millions per asset. By contrast, after strong Phase 2 data, pharma may pay hundreds of millions to billions of dollars for the same underlying asset class.

That spread is the market’s way of saying that molecule discovery is only one layer of value creation. The larger inflection comes when biology, patient selection, dosing, endpoints, and clinical strategy are validated by human data.

AI will accelerate this shift. Improvements in screening, structural biology, computational chemistry, biological datasets, platform biotech, and now AI are increasing the number of plausible therapeutic assets that can be generated. AI-enabled drug design is moving from a scarce technical frontier toward a broadly accessible platform capability. The rise of China as a drug discovery and development engine adds another accelerant.

This is a major scientific achievement. But it also changes the economics of the industry. If the number of discovered drugs increases substantially, then discovery itself becomes less scarce. There will be more drugs per target, more companies pursuing similar biology, more assets available for licensing, and more molecules that look plausible at the preclinical or early clinical stage.

The world will have far more discovered drugs than it can actually develop.

That likely means “we discovered a drug” becomes a weaker basis for durable company value. A discovered asset can still create value, but the market will ask harder questions: Is this the right drug? Is this the right target? Is this the right patient population? Is there a credible clinical path? Is the biology differentiated? Can this become a product?

In a world of drug discovery abundance, the scarce capability is no longer making drugs, but rather, knowing which drugs are worth developing.

More drugs per target likely means lower value per discovered asset

As discovery becomes more efficient, more companies will converge on the same attractive targets.

When a target is compelling, better tools will make it easier for the field to generate multiple molecules, modalities, and development hypotheses against it. The result will not be one perfect drug per target, it will be many plausible drugs per target.

This creates a paradox: AI may make it easier to create drugs, but that does not necessarily make each drug more valuable.

In fact, the opposite may happen. As the cost of generating a plausible asset falls, the supply of early assets will rise. More companies will be able to create something that looks credible enough to fund, partner, or advance. But abundance changes the economics. The marginal discovered drug becomes cheaper to produce, easier to replicate, and therefore less scarce.

We are already seeing this happen. Over the past two decades, the number of drugs being developed against each target has increased meaningfully — from roughly two to three drugs per target to seven or eight, and in some cases fifteen or more. If discovery tools continue to improve, it is reasonable to expect that number to rise further, potentially to twenty, thirty, or even fifty programs against the most attractive targets. For example, in 2025, there were more than 100 programs targeting PD-1 and GLP-1s each.

Figure 2. Drugs per attractive target keep climbing — and outliers are already in the hundreds (McKinsey Life Sciences, 2025). Range covers typical attractive targets. Mega-targets like PD-1/PD-L1 (200+ programs, DelveInsight 2025) and obesity/GLP-1 (193 assets, IQVIA Oct 2025) already far exceed these levels in 2025. The 2030 projection assumes ~2x current levels for typical targets, sourced from Quantum Black by McKinsey and others.

This has major implications for value creation.

Preclinical assets and discovery-stage companies will face increasing valuation pressure as the market recognizes that early assets are becoming more abundant. A molecule alone will no longer command the same premium if others can generate comparable molecules against the same target. Differentiation will shift toward what is harder and slower to prove: stronger translational evidence, prospective clinical data, the right endpoints, better patient selection, and a faster, more credible path to approval.

This does not mean discovery is unimportant. It means discovery alone is less likely to be the durable value inflection point. And we have already seen this empirically in the markets for the past 10 years — valuations for discovered drug candidates have decreased over time (pre-Phase 2 drugs).

Predictive models will improve unevenly

AI will improve prediction across pharma, but not all predictions are equally hard. The sequence of progress will be uneven, and that sequencing is important.

Prediction will improve first in the domains where data is abundant, variability is lower, benchmarks are clearer, and feedback loops are short. It will improve later in domains where biology is more heterogeneous, outcomes are harder to define, and validation requires long, expensive, prospective studies.

Put simply: drug discovery and design will improve first, toxicity prediction will improve next, and clinical efficacy will be last.

Figure 3. Three frontiers, three timelines. The diagonal traces the order in which AI prediction is likely to mature. Position reflects feedback loop length on the horizontal axis and benchmark clarity on the vertical axis.

Drug Design and Discovery — the most tractable frontier

We can define drug design as the process of identifying molecules that can meaningfully modulate a target of interest, whether by agonizing it, antagonizing it, inhibiting it, degrading it, or otherwise altering its function.

Among the major frontiers in drug development, discovery is likely the most tractable for AI because its feedback loops are comparatively short. Experiments can often be run in minutes, hours, or days. Benchmarks are very clear — we know what assays are needed to understand what binds to what, binding efficiency, potency, PK/PD properties. Assays are more scalable. The search space can be explored computationally. Models can generate hypotheses, test them against available data, and improve through repeated iteration.

This does not make discovery easy, but it makes it more software-like than other parts of drug development. Understanding whether a drug can be designed to bind, agonize, or antagonize a given target is becoming increasingly efficient. Billions of dollars are flowing into companies focused on solving this problem. But all this together means that drug discovery will likely be commoditized first, and value will not accrue at this part of the chain unless you translate these discoveries into post-clinical readout validation.

Toxicity — harder, but next

Toxicity prediction is harder than discovery, but it is likely the next frontier to improve meaningfully with AI. We can define toxicity prediction as the ability to anticipate whether a molecule will produce harmful effects in humans using evidence from in silico models, cellular systems, organoids, and lower-order animal models.

The biology is more complex, but the problem is still comparatively structured. We often know what a toxicity signal looks like. There are recurring failure modes, established assays, historical datasets, and staged experimental systems that progress from cells to organoids to genetically similar animal models and, eventually, to humans.

The feedback loops are longer than in discovery, but still much shorter than clinical efficacy. Experiments may take weeks rather than years, and many toxicity questions benefit from repeated patterns across molecules, tissues, species, doses, and mechanisms.

That makes toxicity a natural next domain for AI-driven improvement. It is not as software-like as discovery, but it has enough structure, repetition, and measurable signal for models to become increasingly useful over time.

Clinical efficacy — the hardest frontier

Let’s define clinical efficacy prediction as the ability to predict whether a candidate drug will work in patients without needing to run a trial. This is the holy grail, and we expect it will be the last frontier to be solved.

Clinical efficacy depends on the full complexity of human biology: disease heterogeneity, patient selection, dose, duration, endpoints, standard of care, adherence, trial design, site execution, and statistical power. A model may predict that a molecule engages a target. It may predict that the molecule is likely to be safe. But predicting whether that drug will produce a clinically meaningful benefit in the right patients, under real-world trial conditions, remains far more difficult.

One reason is that the datasets required to solve this problem largely do not exist yet. To predict efficacy reliably, we would likely need something like a UK Biobank on steroids: multimodal, longitudinal, deeply phenotyped, multi-omics datasets across millions of patients, linked to treatments, outcomes, disease progression, biomarkers, imaging, clinical notes, adherence, environment, and real-world follow-up. And we would need those datasets not just observationally, but across enough perturbations — enough drugs, doses, mechanisms, patient subtypes, and controlled interventions for models to learn which biological changes actually translate into clinical benefit.

Today, that data is sparse, fragmented, biased, and unevenly distributed across institutions, trials, EHRs, registries, biobanks, omics platforms, and pharma archives. Unlike internet-scale AI, where models can learn from enormous volumes of naturally occurring text, clinical efficacy prediction requires data from expensive, slow, ethically constrained human experiments. The most important labels — whether a drug works, in whom, at what dose, against which endpoint, and under which conditions — are generated through trials that take years.

Unlike toxicity, where many failure modes are observable, repeatable, and measurable in staged systems, efficacy often lacks clean benchmarks. We do not always know what “getting better” looks like in a way that is stable, quantifiable, and easy for models to learn, especially in complex diseases like Alzheimer’s, autoimmune disorders, metabolic disease, and many areas of neurology.

Even in diseases we understand better, efficacy remains difficult to predict because human response is heterogeneous. Patients vary in baseline biology, disease progression, comorbidities, prior treatment history, adherence, environment, and response to intervention. We are often still uncertain about natural variability, responder and non-responder populations, and how small-scale biological perturbations translate into meaningful outcomes at human scale.

The consequence of uneven frontiers and rates of change

This creates an important sequencing implication:

1. First, the industry will have more promising discovered molecules than it can develop.

2. Then, it will have more molecules that appear likely to be safe.

3. The last major unknown will be whether those molecules are actually efficacious in patients — and that uncertainty will remain hard to resolve until the industry has much richer human datasets and many more prospective clinical feedback loops.

In other words, clinical efficacy prediction is not just a harder modeling problem. It is a missing-data problem, a human-biology problem, and a feedback-loop problem. That is why clinical development remains the central bottleneck.

What if we are wrong, and clinical efficacy prediction happens far faster than expected?

Even if we eventually have a black-box model that could predict clinical efficacy with high accuracy, it is unlikely that regulators would immediately allow drugs to be approved without prospective clinical trials. They would need to see repeated evidence that the model works in practice. How many prospective trials would need to read out correctly before regulators became comfortable relying on such predictions? Five? Ten? Twenty? And even then, validation in one disease area would not automatically transfer to another. A model that performs well in oncology would still need to prove itself in neurology, immunology, cardiometabolic disease, and other therapeutic areas.

Because each prospective validation cycle can take years, regulatory adoption will lag model capability. AI may improve our confidence, sharpen our development decisions, and increase the probability of success, but it will not eliminate the need for clinical proof anytime soon.

As long as regulators require prospective human clinical trials, clinical development will remain the bottleneck. Unlike discovery or toxicity prediction, it is constrained by the physical realities of drug development: manufacturing clinical supply, completing CMC work, recruiting patients, treating them for the required duration, and waiting for endpoints to be met. These steps can be made faster and more efficient, but they cannot be compressed from years into days or weeks. The limiting factor is not just computation; it is the irreducible time required to test medicines in humans. We suspect that one day, trials can be done in silico, but the data required to train such predictive AI models don’t yet exist (e.g. UK Biobank on steroids: millions of patients, genomic, proteomic, transcriptomic, imaging, longitudinal study, multiple perturbations).

This is why clinical development remains the central bottleneck in pharma. Markets already reflect this reality. Assets are worth far more after Phase 2 readouts because that is where the most important uncertainty begins to collapse: not whether a drug can be discovered, and not only whether it can be safe, but whether it can actually help patients and modulate a clinical endpoint that matters.

Figure 4. A hypothetical $1B peak-sales drug — where value gets created. Risk-adjusted asset value at each development stage. Phase 2 readout anchored at $1B; values use industry-standard probability of success and a 13% discount rate (biotech-typical). The single steepest inflection — a 4x step — occurs from Phase 1 readout to Phase 2 readout.

That final unknown of clinical efficacy is where the most durable value will continue to accrue.

The Key Epistemological Question: How and when do you know you are better than industry averages?

There is also a fundamental epistemological question that almost every biotech company must grapple with. Even if you have a model or platform that you believe predicts clinical stage efficacy at a higher rate, when and how do you know you are better than the industry average?

Clinical proof takes time, and small numbers are hard to interpret. If the baseline Phase 2 probability of success is roughly 30%, and you run five Phase 2 trials and succeed in three (i.e., 60%), are you meaningfully better, or just lucky? If you run ten Phase 2 trials and succeed in five, that is encouraging, but still not statistically significantly better performance. Because Phase 2 trials can take two to four years to read out, it may require a decade or more, and many prospective readouts across therapeutic areas, before the market, regulators, or even the company itself can confidently distinguish true model-driven advantage from variance.

Even if models perform well on historical data, markets are unlikely to give full credit before prospective clinical proof. This is especially true for first-in-class drugs, where there is no clean precedent and where the core question is not merely whether a model can interpolate from the past, but whether it can predict a new biological and clinical reality. For best-in-class drugs (i.e., optimized molecule on known biology), investors may discount positive readouts as efficacy is generally already known.

That could create a type of valuation overhang. Investors may believe AI can improve probability of success, but they are unlikely to underwrite that improvement fully until they see repeated prospective evidence. This is especially true because the first generation of AI biotech companies has not yet clearly demonstrated a step-change improvement in clinical probability of success.

Put simply, the advantage still has to be proven in the clinic. Credibly claiming a statistically significant edge likely requires 10–20 prospective clinical shots that materially outperform industry benchmarks (need to hit at least 6 out of 10 or 10 out of 20 to hit statistical significance at p = 0.048). But if a company starts in discovery, it may run out of time or capital before it can generate that many Phase 2 readouts.

Proof, for now, still comes from the clinic.

Two ways to win, only one of which is more measurable now

But there is one advantage we can know much earlier: whether you can do drug development and clinical trials faster and more efficiently.

That is why the winning strategy is not to rely solely on better success prediction. It is to combine better prediction with a structurally superior drug development engine. If our models help us choose better assets, that is powerful — but you won’t know if you are actually better at prediction until after the drugs are read out clinically. If we can also develop those assets faster, cheaper, and with greater operational consistency, then we create an advantage that is measurable. Functionally, being able to take more shots for the same cost and time as a traditional company meaningfully increases your probability of success of hitting a winner.

This is the strategic wedge.

The company that wins will be the one that can both make better bets and resolve those bets faster and more efficiently.

Inverse Pascal’s Wager: Underwrite to industry averages and ensure you can take enough shots

Pascal’s Wager is an argument about decision-making under uncertainty. Pascal argued that, even if we cannot know whether God exists, belief is rational because the payoff structure is asymmetric: if you believe and God exists, the upside is infinite; if you believe and God does not exist, the downside is limited. The expected value is dominated by the possibility of infinite upside. And, of course, if you do not believe and an angry God does exist, the downside is infinitely bad.

For prediction of clinical probability of success, we subscribe to an inverse Pascal’s Wager.

Figure 5. A depiction of an inverse Pascal’s Wager.

It is tempting to believe that better models should translate into materially higher clinical probabilities of success. And over time, they may. But if we underwrite too aggressively to that belief and we are wrong, the downside is existential. We overpay for assets, take too few shots, overbuild around false confidence, misallocate capital, and risk killing the company.

The safer strategic posture is to assume industry-average probabilities of success until proven otherwise, and build a business model that works under that assumption.

Build a company that can take the appropriate number of well-underwritten shots even at industry-average odds.

The goal is to build a company that can take enough well-underwritten shots that it does not depend on needing to have a higher PoS (but of course do everything you can to achieve a higher PoS).

If the baseline Phase 2 probability of success is roughly 30%, then any single program is more likely to fail than succeed. But a portfolio changes the risk profile. A portfolio of 10 independent Phase 2 programs has only about a 2.8% chance of producing zero winners. Said differently, even at industry-average odds, taking enough disciplined shots dramatically reduces the risk of complete failure.

Figure 6. A portfolio of 10 has only a 2.8% chance of zero winners. Binomial distribution with n = 10 trials and p = 0.30 (industry average Phase 2 probability of success). Sources: BIO and Wong, Siah & Lo. Even at average odds, a well-constructed portfolio of 10 produces a positive expected outcome with extremely low probability of total failure.

That is the asymmetry we want.

If we underwrite to industry averages and our models are actually better, we outperform. If we underwrite as though our models are better and they are actually average, we can misprice risk, overpay for assets, take too few shots, and destroy the company.

So the principle is simple: underwrite the company to industry-average probabilities of success, and let any predictive advantage show up as realized upside.

There is an important nuance. We should absolutely use predictive models to do everything you can to maximize PoS and rank assets, prioritize programs, estimate rNPV, and decide which shots are worth taking. The AI models should shape portfolio selection. But the company strategy should not depend on assuming that our predicted probability of success is already superior to the industry’s as that exposes you to existential risk.

If we have better predictive models, can run trials cheaper and faster, and can commercialize more efficiently, then we win twice: first through disciplined portfolio construction, and second through realized performance that exceeds the baseline.

Until those questions are answered, the right strategy is humility at the asset level and discipline at the portfolio level: assume average odds, take enough well-priced shots, and let better prediction create upside rather than fragility.

Our strategy and capabilities are built around the bottlenecks that emerge as AI creates uneven frontiers

For Formation Bio, the opportunity is to build around the bottleneck we believe will persist as AI reshapes pharma: clinical development. We license the most promising drugs and move them through clinical trials faster, cheaper, and more efficiently. In the near term, we sell or partner validated assets back to pharma after major value inflection, such as a Phase 2 readout. Over time, we will build toward commercializing our own medicines.

If we cannot find the assets we want, we may eventually pursue drug discovery ourselves, especially as discovery timelines collapse further. But in a world with more good drugs than available development capacity, even spending one to two years generating another candidate carries real opportunity cost. That time is better spent on the scarce capability and one most proximal to creating durable market value: identifying the best assets already available, matching them to the right indications, and moving them through clinical proof with greater speed and discipline. That is the strategic premise behind Formation.

Our platform is not a generic AI strategy. It is built around our specific view of the uneven frontier: discovery gets easier first, toxicity prediction improves next, and clinical efficacy and development remain the hardest and most valuable bottlenecks. We are building systems around the parts of pharma where scarcity and value will concentrate as AI changes the industry and the sequencing of what we work on is important.

If the uneven frontier plays out, AI will first enable the discovery of far more drug candidates, then more candidates with lower predicted toxicity risk, and eventually a growing abundance of plausible drugs whose biggest unresolved question is whether they work in humans.

That creates a very different world: not one where the scarce resource is a molecule, but one where the scarce resource is knowing which molecule, in which indication, for which patient population, with which trial design, is actually worth advancing. Many assets may look promising on paper, bind the intended target, and appear developable preclinically, yet still have large unanswered questions about clinical efficacy. At the same time, many target-indication pairings will remain unexplored, under-tested, or poorly understood.

If that is the case, three capabilities become most important.

First, we need to be the best at picking drugs

In a world with more assets than the industry can develop, asset selection becomes one of the most important capabilities in pharma. The company that wins won’t be the one that sees the most drugs, it will be the one to identify, earlier and more accurately than others, which assets are most likely to become valuable medicines.

This has always been the case, but now the universe of discovered candidates will dramatically increase as AI and China transform our industry.

Our platform comprises several key capabilities across the drug development lifecycle. Our AI system Atlas allows us to map and screen the universe of drug assets, competitive landscapes, and target-indication opportunities more efficiently and with greater scale than manual teams and processes. But seeing more opportunities is only useful if we can judge them better. The real advantage comes from combining AI-scale search with exceptional human judgment.

The best teams will still be differentiated by taste, judgment, and pattern recognition: the ability to know which risks are fatal, which are manageable, and which signals actually matter. The goal is not to make isolated one-off decisions, but to embed expert judgment (e.g. former Heads of R&D, biotech investors, PhDs) into the AI systems that evaluate assets, so Formation’s asset-selection capability compounds over time.

Our Delphi system closes that loop. It tracks clinical readouts across the industry, allowing us to learn not only from the drugs we own, but from every relevant trial the market runs. By comparing outcomes against our prior predictions, Delphi continuously sharpens our understanding of probability of success. Every readout becomes a learning event: which risks mattered, which signals translated, which assumptions were wrong, and which patterns should inform the next asset we evaluate.

Together, Atlas and Delphi form the foundation of a compounding drug-picking engine. Atlas expands the universe of assets we can see. Expert judgment teaches the system what to value. Delphi turns clinical outcomes into feedback that improves future decisions. Over time, asset selection becomes an institutional capability that strengthens with every asset we evaluate and every trial the industry reads out.

Second, we need to be the best at precision drug development and matching the right drug to the right patients

As more drugs emerge for each target, value will increasingly depend on matching each asset to the right indication, patient population, endpoint, and development path. The same molecule can be mediocre in one setting and highly valuable in another. The difference is often not the molecule alone, but the development strategy around it.

This is too large a search problem for a purely human process to do comprehensively and exhaustively. For every asset, there are many possible indications, patient segments, biomarkers, endpoints, comparators, trial designs, and commercial paths. Evaluating those options requires integrating genetics, real-world evidence, literature, omics, competitive intelligence, regulatory precedent, and market structure.

AI can compress that process. It can help us evaluate far more drug-indication pairs, compare more development strategies, and identify the settings where an asset has the strongest chance of becoming a valuable medicine.

That is the role of our Forge and Cassini systems. Forge helps evaluate development paths, simulate program options, and benchmark against prior studies. Cassini helps determine whether a program can meet a target product profile that creates meaningful market value. Together, they allow us to move from “Is this a good drug?” to the more important question: “Where, how, and for whom can this drug become a valuable medicine?”

Third, we need a development and capital advantage (more shots on goal)

If clinical trials remain the bottleneck in drug development, then the company that can run more high-quality trials, at lower cost and greater speed, will have a meaningful advantage. This is where Formation has a structurally different model. If we can take more shots per dollar, and finance those shots at a lower effective cost of capital, we can increase the probability of producing winners without proportionally increasing risk to the parent company.

Drug development is a probabilistic business. Even well-selected assets fail. Advantage therefore comes not only from picking better programs, but from being able to fund and run more well-underwritten programs efficiently. A company that can develop assets faster, cheaper, and with less reliance on corporate equity can pursue a broader portfolio for the same balance-sheet burden.

That changes the math. More shots per dollar increases portfolio breadth. Lower cost of capital preserves upside and reduces dilution. Faster execution resolves uncertainty sooner. Together, these advantages increase the likelihood that Formation can generate multiple successful readouts while avoiding one of the most common failure modes in biotech: running out of capital before reaching meaningful clinical value inflection.

In a probabilistic business, the ability to take more intelligent shots is enormously valuable. One meaningful winner in our industry can pay for many losses, but only if the company can take enough well-underwritten shots without compromising capital discipline. More shots, selected intelligently, developed efficiently, and financed creatively, can compound into a durable advantage.

Figure 7. Probability that a portfolio of n independent Phase 2 programs produces zero winners, at industry baseline Phase 2 PoS = 30% (See sources in Figure 6.)

Apollo (our AI-enabled CRO), built over years from our TrialSpark foundation, is designed to give us an execution advantage: the ability to run trials faster, cheaper, and with more operational control. Faster startup, lower trial execution costs, better patient recruitment, stronger protocol design, automated workflows, and tighter operational feedback loops all increase both the number and quality of development bets we can make. If clinical proof is the industry bottleneck, then reducing the time and cost required to generate that proof is one of the most important sources of leverage.

Alternative capital vehicles give us a capital advantage. Clinical-stage assets may require tens of millions of dollars before a meaningful readout. If every program must be funded directly from the company’s balance sheet, the cost of capital compounds quickly, and so does dilution. The company must either raise more equity, accept dilution, slow down, or take fewer shots.

Alternative capital structures can change that equation (more to come on this shortly!). If a company can find a way to finance more programs through structures designed specifically for asset development, rather than needing to solely finance each drug on our balance sheet, this can reduce dilution, improve capital efficiency, and allow us to pursue more opportunities for the same corporate balance-sheet burden.

This creates a structural advantage. Better prediction helps us choose better assets. A faster and cheaper development engine lets us test those choices more efficiently. A dedicated capital vehicle lets us scale the number of shots we can take without overburdening the company. And every trial creates institutional learning: each protocol, enrollment curve, regulatory interaction, endpoint decision, and readout should make the next trial faster, cheaper, and better informed.

The compounding advantage

At Formation, we won’t be able to harvest the fruits of AI, and ultimately deliver more treatments to patients, unless we solve the core bottlenecks that emerge as AI transforms drug discovery.

As the uneven frontier takes shape, the systems we are building are designed to address each of these bottlenecks in turn. Atlas helps us win in a world of drug abundance by mapping the universe of candidates and surfacing the assets most worth developing. Delphi helps us estimate probability of success, model value, and stress-test development paths as predictive models improve unevenly. Forge helps us design and simulate clinical programs in a world where drug-indication pairing and study design determine value. Apollo helps us compress the time and cost of trial operations, where clinical execution remains the binding constraint. ARK is the data foundation underneath all of it, integrating clinical, genomic, real-world evidence, and competitive intelligence so the entire system compounds.

Over time, as the cost and time required for discovery collapse further, we may enter discovery ourselves and close the full loop. But the right place to start is where value is most concentrated today: closest to clinical proof and the major value inflection of a Phase 2 readout.

The most important advantage is that this system compounds.

Traditional pharma often treats each deal, asset, and trial as a discrete event. Knowledge is fragmented across teams of thousands, documents, consultants, vendors, and individual memories. Lessons are learned, but they are not always systematically captured. Mistakes repeat. Insights decay. Institutional knowledge walks out the door.

Formation is designed to be different.

Our platform creates a flywheel where every asset evaluated and every trial executed should make the entire system smarter. Every asset screened improves our landscape models. Every evaluation sharpens our prediction systems. Every trial refines our benchmarks for endpoints, recruitment, site performance, cost, timelines, and execution risk. Every regulatory interaction improves our understanding of viable development pathways. Every commercial assessment improves our understanding of what readouts create value.

Over time, every program should add to Formation’s exocortex: an enduring system of record for how to develop drugs better.

In an AI-enabled pharma world, the winning company will not simply be the one with access to the best public models. Those models will become broadly available and their value ubiquitous. The more durable advantage will come from proprietary workflows, proprietary data, repeated decisions, and organizational learning loops that competitors cannot easily copy.

Formation’s goal should be to make every deal and every program improve the next one. Each asset we evaluate should make us better at asset selection. Each trial we run should make us better at execution. Each readout should make us better at prediction. Each financing should make us better at structuring risk.

The frontier is uneven, and that is the opportunity

The next decade of pharma will not be defined by a single AI breakthrough. It will be defined by a sequence of capability shifts.

Drug discovery will become faster and cheaper. Toxicity prediction will improve. Clinical efficacy prediction will lag. Development will remain the bottleneck. The world will have more drugs than it can develop, more hypotheses than it can test, and more assets than capital can efficiently support.

That unevenness is the opportunity.

It also raises the question that sits underneath every AI strategy. Defensibility is not just about what works today. It is about whether your position becomes more valuable or less necessary as the models improve. That is why we do not try to predict the future perfectly. In a world of rapid change, strategy is not built on certainty; it is built on informed judgment about timing.

The answer is to build around the bottleneck. If we build the best system for picking drugs, pairing them with the right indications, designing better trials, financing development efficiently, and executing faster, then each wave of AI progress strengthens our model rather than threatening it.

The uneven frontier rewards companies that know where on the continuum to stand. Formation stands at the point where drug abundance meets development scarcity. That is where value will be created.

And the stakes are enormous. In a post-AI world, one of humanity’s defining ambitions will be to live longer, healthier lives. One of the most important companies of the next era will be the company that helps make that possible by turning the right molecules into medicines with greater speed, precision, and scale.

Subscribe now

This essay reflects the personal views and opinions of the author and is provided for informational purposes only. It contains forward-looking statements regarding artificial intelligence, the evolution of biopharmaceutical research and development, and Formation Bio’s strategy, platforms, and positioning. These statements are based on current expectations, assumptions, and information available as of the date of publication, and are subject to significant risks and uncertainties — including those relating to clinical development, regulatory review, competition, and broader market conditions. Any hypothetical asset valuations, portfolio probabilities, or projections shown are illustrative only and do not represent any Formation Bio asset, program, financial result, or forecast. Nothing in this essay constitutes an offer to sell, or a solicitation of an offer to buy, any securities. References to specific companies, products, or third-party data are illustrative and do not constitute endorsements. Third-party data has not been independently verified.

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

News as of this morning: SpaceX is acquiring Cursor. Here’s founder Michael Truell and a16z’s Martin Casado at last year’s Runtime summit, talking about many of the themes that matter in the coming AI decade:

Of all the seats in a company, the most senior finance person is one of the loneliest. You sit amongst the board and the CEO, you see every number before anyone else, and yet you rarely have a true peer inside your own walls. When something hard lands: a challenging relationship dynamic, a complex financing decision, the business taking an unpredictable turn, a hire you’re unsure about, there often isn’t anyone you can talk to.

We want to do something to fix this.

It wasn’t even our idea, at first. We hosted a number of brainstorming sessions and dinners and the ask came back loud and clear: our finance leaders wanted a real community, one built for both tactical learning and genuine relationships, something closer to YPO (long established network for CEOs — less networking dinner, more trusted inner circle: the same small group of peers over time, who know your real situation and become the people you call first). We listened, and we’ve spent the months since designing the a16z CFO Community very deliberately.

Here’s what it is. Stage-matched cohorts of roughly eight finance leaders, meeting in person every other month for a year — six sessions in all. Each session pairs a candid, moderated small-group conversation with a tactical block on the topics CFOs actually wrestle with: annual planning, forecasting and benchmarks, building finance teams for scale, AI tooling and efficiency, and GTM and engineering productivity. There’s also a lot of room and space to get to know each other on a more personal and candid basis, and build relationships that will last years, where you can share trusting feedback. The first cohort kicks off this month.

Apply

A few principles shaped every decision:

Curation is everything. The group only works if you’re in the room with people at your stage facing similar problems, so in our estimate, roughly 80% are true peers, with a few slightly ahead so you can see where you’re headed. Small and selective beats big and broad every time.

Trust over scale. Communities die when they become impersonal and transactional. We’re keeping cohorts small enough that your absence is noticed, and we’re protecting the room — no investors, no observers. Just operators talking straight.

Accountability is a feature, not a bug. You opted in, we selected you, and we ask you to show up — one miss across the six sessions is ok. That friction is the point. Commitment is what turns a calendar invite into a relationship you lean on for years.

The content is the carrot; the relationship is the outcome. The moderators and their experience get you in the door. What you walk away with is a short list of people you can actually call upon whenever you need, and they can call you.

Why join? Because the job doesn’t get less lonely as you scale — it gets more so. This is a place to pressure-test the decisions you can’t put in a board deck, to learn from people a step ahead, and to give back to those a step behind. Over time, it becomes a brain trust: a group of finance leaders who know your context and pick up your message.

We’re starting small and intentional. If you’ve ever wished for a room where you could ask the real question without worrying how it looks — we are building this for you.

Cheers and hope you join us!

Learn More

Subscribe now

This material is solely for educational purposes and is not investment advice or an offer of investment advisory services. This material should not be used as the basis for an investment decision. a16z is an investor in SpaceX and Cursor through its managed funds, and thus has a financial interest in the company’s performance and future prospects. In particular, a16z benefits if the company grows in value; and a16z funds will receive any customary dividend payments in connection with their status as a shareholder of the company. However, a16z is not being compensated by SpaceX or Cursor for this material.

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 is the inaugural issue of Profiles of the Future, a new monthly series from a16z that profiles the future of an industry through the lens of the company and founder defining it. - Alex Danco

Elon Musk’s compensation package at SpaceX is structured around two targets. The first award vests if the company reaches a valuation of $7.5 trillion and establishes a permanent human colony on Mars of at least one million people. The second vests if SpaceX operates data centers in space that draw at least 100 terawatts of power, more than 1,000x the consumption of every data center on Earth combined. Miss both, and Musk earns nothing but the $54,080 salary he has been paid since 2019.

The board members who signed this package spent two decades watching Musk make predictions about SpaceX that sounded impossible before they came true. He said SpaceX would put humans in orbit when no private company ever had; it now flies NASA’s astronauts routinely. He said it would land and reuse an orbital rocket when the entire industry treated boosters as disposable; SpaceX has since done it hundreds of times. He said a satellite internet business could be worth tens of billions when satellite internet was a graveyard of bankruptcies; Starlink’s revenue has climbed from zero to $11.4 billion in a few years. The predictions were often aggressive on timing but almost never wrong on direction. And the original direction, written down in 2002 as the company’s mission, was to make humanity multiplanetary. So the board tied his pay to the mission itself.

If that mission sounds like something from a science fiction novel, that might be because it is.

Iain M. Banks spent twenty-five years writing about a civilization called the Culture. It is, by most reasonable measures, the best utopian society ever imagined. Humans live alongside Minds, the superintelligent AIs that run orbital habitats the size of small worlds, in a relationship that is neither servitude nor rivalry but partnership. Nobody works who doesn’t want to. Nobody starves. The Minds handle the staggering computational load of running cities in space. Humans handle being human, which turns out to be a full-time job.

SpaceX’s three autonomous drone ships, the floating platforms in the oceans where Falcon 9 boosters land, are named after sentient starships in Banks’s novels: Of Course I Still Love You, Just Read the Instructions, and A Shortfall of Gravitas. In a 2023 interview at the UK AI Safety Summit, Musk was asked what a good AI future looks like. “The Banks Culture books are by far the best envisioning of an AI future,” he answered. “There’s nothing even close that’ll give you a sense of what is a fairly utopian or protopian future with AI.” He has been telling us, on the sides of his landing pads, exactly what he is trying to build.

The Culture is not a frictionless paradise. Banks’s novels are full of war, intrigue, and moral complexity. It is utopian because civilization has solved the prerequisites of survival well enough that trillions of humans are free to take care of, in the words of Banks, “the things that really mattered in life, such as sport, games, romance, studying dead languages, barbarian societies and impossible problems, and climbing high mountains without the aid of a safety net.”

A future like this has four prerequisites. First is access to a meaningful fraction of a star’s energy output (orders of magnitude beyond what human civilization produces today). Second is physical intelligence at scale: machines that can build, mine, refine, and repair anything, anywhere, without a human in the loop. Third is cheap digital intelligence that exceeds biological intelligence. And fourth is a way to move mass off the planet cheaply, frequently, and reliably, because none of the above scales on Earth alone.

Working Back from the Future

Most analyses of SpaceX work forward from the present: rockets, satellites, contracts, revenue. But to see what’s actually happening, it’s more useful to start at the destination and work backward.

The Mars city. The operational target is a self-sustaining city of a million people on Mars within the lifetime of people alive today. Self-sustaining is the hard part. It means the city has to survive if Earth stops sending ships, which requires manufacturing its own everything: food, water, air, energy, medicine, machinery, and eventually more humans. Getting a million people and millions of tons of cargo there over a few decades will take, by SpaceX’s own math, several thousand Starship flights at more than ten launches per day during each transfer window. Those windows, set by Earth-Mars orbital mechanics, are only a few weeks wide and open just once every 26 months.

The Moon city. This is the closer, easier dress rehearsal. The lunar south pole has ice in permanently shadowed craters and continuous solar exposure on certain ridges, which makes it the natural site for a base. But Musk has talked about something more ambitious than a research outpost. He envisions factories on the Moon building AI satellites with a mass driver shooting them into space one after another. Another idea Musk has borrowed from science fiction, a mass driver is an electromagnetic launch system that exploits the Moon’s one-sixth gravity and absent atmosphere to fling solar-powered satellites into deep space at industrial scale. The satellites could be built on the Moon itself given that lunar regolith is roughly 20% silicon and 10% aluminum by weight, the two main inputs for solar cells and satellite structure. “If you want to go beyond a mere terawatt per year,” Musk explains, “you have to go to the moon.”

The orbital data centers. Musk is betting that in a few years the most economically compelling place to put AI data centers will be space. The bottleneck on AI is energy, which is barely growing outside of China while demand for AI compute grows exponentially. Solar panels in orbit deliver four to ten times more power than the same panels on Earth (depending on how sunny the ground location is) because there’s no atmosphere, no day-night cycle, no clouds, and no seasons. NASA worked this out decades ago, and rockets are finally cheap enough to make it real. In five years, Musk projects, SpaceX will be launching more AI compute to orbit per year than the cumulative installed base on Earth. This is why SpaceX merged with xAI in February. Rockets and intelligence are becoming the same problem.

Starship is the vehicle that makes everything upstream possible. Starship V3, which made its debut flight this year, is the largest and most powerful rocket ever built – taller than a 40-story building and more than twice as powerful as the Saturn V that carried astronauts to the moon. By NASA’s accounting, reaching orbit historically cost around $18,500 per kilogram. In 2010, the first Falcon 9 brought that down by about 85% to roughly $2,700. In 2018, the Falcon Heavy cut it further to about $1,400. Starship, designed to be the world’s first fully and rapidly reusable spacecraft, aims to further reduce it to $100-500 per kilogram. Spaceflight that once cost billions per launch now costs in the tens of millions.

Starlink is the cash flywheel that helps pay for everything else. According to SpaceX’s IPO filing, the Connectivity segment (almost all of it Starlink) brought in $11.4 billion of revenue in 2025, up roughly 50% year over year, at an adjusted EBITDA margin north of 60%. As of March 2026, the service had 10.3 million subscribers in 164 countries running on more than 9,600 satellites. Starlink started as a side project to fill the company’s own launches, and it’s becoming one of the great consumer businesses in history. When a16z was doing diligence on SpaceX in 2019, several people told us the economics would never work. The dish required antenna technology previously reserved for F-22 fighter jets and Navy destroyers that were never mass-produced for consumers. SpaceX’s first units cost about $3,000 to build and sold for $499. But they figured out how to drive the manufacturing costs down and prove the skeptics wrong.

Falcon 9 is the workhorse that buys time for everything else. It is the only orbital-class booster on Earth reused at scale, with individual boosters routinely flying more than twenty missions each before retirement. In 2025, SpaceX launched 83% of the mass sent to orbit from Earth. The company has now launched more payloads to orbit than the rest of the world combined, despite giving everyone else a half-century head start.

That’s the stack, top to bottom. The Culture lives on top, generations from now. Falcon 9 and Starlink sit at the bottom, paying the bills today. Each layer makes the next one possible.

SpaceX CFO Bret Johnsen describes what it looks like from inside the company:

“[Musk] creates a culture where you set out what initially look like audacious goals, and then step-by-step, you realize that you’re marching toward something that is absolutely achievable… If I think about going to Mars, for example. When I first got here in 2011, people would be rolling their eyes when they talk about Mars and being a multiplanetary species. Nowadays when we say that, the response is literally, ‘What year?’... And I think what Elon has done a masterful job on is setting out these targets and creating a fantastic business model around each piece of IP that you need for that end goal.”

The Idiot Index and the Algorithm

Musk didn’t originally set out to build a rocket company. In 2001, a thirty-year-old Musk was figuring out what he wanted to do after PayPal. He had always been interested in space, and when he went looking for NASA’s plan to put humans on Mars, he was surprised to learn there wasn’t one. So he devised a plan to send a small greenhouse to Mars and broadcast the picture home. The idea was that a green sprout on a dead red planet might reignite public interest in space and the political will to fund a real Mars program. He just needed a rocket to get it there.

Later that year he traveled to Moscow to buy a refurbished intercontinental ballistic missile, the first of two trips. The meetings were reportedly fueled by vodka and a lot of posturing. “We’d all go in this little room and every single person had his own bottle in front of him,” Adeo Ressi, Musk’s best friend from Penn who came along on the trip, told Esquire in 2012. The Russians didn’t take Musk seriously, and on one occasion, a chief designer spat on him and his team in a display of contempt. On the second trip, in February, Musk asked how much a missile would cost. $8 million each, they said. When Musk countered with $8 million for two, Musk’s aerospace advisor Jim Cantrell remembers them saying something like “Young boy. No,” and implying he didn’t have the money. Musk decided they weren’t serious and walked out.

Cantrell figured the trip was over. He and Mike Griffin, who would later run NASA and had come along as an advisor on the second trip, ordered drinks on the flight home and clinked their glasses, glad to be out of Moscow. Musk sat in the row ahead of them, hunched over his laptop. Then he turned around in his seat. “Hey, guys,” he said, “I think we can build this rocket ourselves.” He showed them a spreadsheet listing the raw materials in a rocket – aluminum, titanium, copper, carbon fiber – and what each one cost. The materials cost only two percent of the quoted price, and as Musk would later put it, “clearly you just need to think of clever ways to take those materials and combine them into the shape of a rocket.”

Within months, Musk decided he could risk $100 million on a rocket company (more than half of the roughly $180 million he would receive from the sale of PayPal), and founded SpaceX in a warehouse in El Segundo, California. He offered five people spots on the founding team. Three said no, including Cantrell and Griffin. The two who said yes were Tom Mueller, who became VP of Propulsion and employee number one, and Chris Thompson, employee number two, who took operations and production.

Years later, Musk would dub the principle underlying his spreadsheet diagnostic tool the “idiot index.” If the ratio of a part’s cost to its raw materials is high, you are either an idiot or you are working with idiots. This sounds like a joke, but it’s the foundation of SpaceX’s strategy.

Every part SpaceX bought was accompanied by an idiot index calculation. One of the legendary stories from the early days of the company involves Steve Davis, who joined SpaceX straight out of Stanford as the 14th employee and was tasked with sourcing an actuator to steer the Falcon 1 rocket’s upper stage. When he reported that a traditional aerospace supplier wanted $120,000 for the part, Musk laughed, telling him the component was no more complex than a garage door opener. Musk gave Davis a budget of $5,000 to build it from scratch. As biographer Ashlee Vance recounts, Davis spent nine months toiling over the design, ultimately producing a functional actuator for just $3,900. When Davis sent a technical breakdown of the triumph, Musk responded with a characteristically brief, two-letter email: “Ok.”

To drive the idiot index toward its theoretical lower bound, you must vertically integrate and control the process end-to-end. But vertical integration creates fixed costs that only pay off at high volume, and high volume in the rocket business required breaking with how the industry had always operated.

Traditional launch providers like ULA and Arianespace treated each mission as a custom job. The customer specified the orbit, the payload, and the integration requirements, while the launch provider designed a custom mission around the satellite. That model assumed a few launches per year at very high per-mission cost, and it made manufacturing at scale impossible.

SpaceX inverted this. They published a Falcon User’s Guide that defined the rocket’s exact specifications and told customers to design their satellites to fit. At the time, this was considered radical, and it cost SpaceX some early business. But it unlocked a manufacturing flywheel.

Standardization and reusability fed each other. Because every Falcon 9 was the same, a recovered booster could become a finished, qualified product ready to fly again. The first Falcon 9 booster to fly twice did so in 2017. By 2020, individual boosters were flying five times. By 2021, ten times. Today, the record-holding booster has flown 35 missions. This reusability has changed the economics of spaceflight, and it’s difficult to see how the competition catches up. In 2021, Musk estimated the Falcon 9’s best-case marginal cost of launch (excluding overhead allocation) for 15 tons to orbit was about $15 million, which he said was “about half to one third of the cost of alternatives.” Today, SpaceX launches a rocket every two to three days on reused boosters while competitors launch only a handful of custom rockets a year.

But SpaceX’s advantage is not just economies of scale, vertical integration, and a better strategy. It’s also speed and culture.

Traditional aerospace companies eliminate uncertainty through analysis. Boeing’s commercial crew program, in NASA’s polite phrasing, “utilizes a well-established systems engineering methodology targeted at an initial investment in engineering studies and analysis to mature the system design prior to building and testing.” Measure twice, cut once. SpaceX inverted this. The company builds many cheap prototypes, pushes them to failure, learns from the failure, and iterates. The Starship test campaign has produced more spectacular explosions than any rocket program in history, but each failure is a data point about where reality diverged from the model.

The contrast was visible to anyone who worked in both worlds. Garrett Reisman was a NASA astronaut who flew two Space Shuttle missions, then left NASA in 2011 to join SpaceX as a senior engineer. He has described the prevailing NASA view of SpaceX in those years: “They’re cowboys; they’re dangerous; they’re going to kill somebody.” What changed his mind was watching SpaceX work. “They were making things in a month that would have taken NASA like a year. We were just amazed.”

The clearest example is the Falcon 1 program. Between 2006 and 2008, SpaceX launched four Falcon 1 rockets from a tiny atoll in the Pacific called Kwajalein. The first three failed, but each failure was different and instructive. A fuel leak on flight one. A propellant sloshing anomaly on flight two. A separation collision caused by residual engine thrust on flight three. By September 2008, the company had funds for exactly one more launch. And it wasn’t Musk’s only company on the edge. Tesla, the electric car company Musk was building in parallel with SpaceX, was also weeks from bankruptcy, and he had to decide if he would consolidate his remaining PayPal cash behind one company or split it between both.

“That was a really tough decision. I decided in the end to split what I had to try to keep both companies alive, but that could’ve been a terrible decision that resulted in both companies dying,” Musk recalled. “I never thought I’d have a nervous breakdown, but I came pretty darn close.” He couldn’t choose because, in his worldview, both missions were essential – Tesla to accelerate the world’s shift to sustainable energy and SpaceX to make humans multiplanetary. “All available resources had to be plowed into the companies,” Talulah Riley, Musk’s then-fiancée, said in the BBC documentary series The Elon Musk Show. “He gave me an out. He said, ‘This is going to be the hard part, you don’t have to stay for it.’”

Flight four worked. That December, weeks before SpaceX would have run out of money, NASA awarded it a $1.6 billion cargo contract. When they called Musk to let him know, he was so overwhelmed with emotional relief that he blurted out, “I love you guys.”

The pattern that emerged from this experience of failing fast and correcting mistakes quickly became the culture of every program that followed. It is the same pattern that now lets SpaceX iterate on Starship between flights, while a traditional aerospace program takes years to go from a flight anomaly to a redesigned vehicle.

The reason this works better than the alternative is because you cannot think your way to perfect solutions for problems you do not fully understand. Reality is the only adequate validator, and the trick is making it cheap enough to consult often.

This is SpaceX’s iteration loop told through stories, but there’s also a written-down version. Musk has, over the past two decades, codified the SpaceX approach into a five-step operational process that the company calls the “Algorithm.” Tim Berry, who spent ten years at SpaceX leading the upper-stage production team for Falcon 9 and Falcon Heavy, has described it as “drilled into our minds.” Walter Isaacson published the canonical version of it in his Musk biography:

1. Question every requirement. Each requirement should come with the name of the person who made it. You should never accept that a requirement came from a department, such as the legal department or the safety department. You need to know the name of the real person who made that requirement, and you should question it no matter how smart that person is. Requirements from smart people are the most dangerous because people are less likely to question them. Then make the requirements less dumb.

2. Delete any part or process you can. You may have to add them back later. In fact, if you do not end up adding back at least 10% of what you deleted, you did not delete enough.

3. Simplify and optimize. This one should come after step two. A common mistake is to simplify and optimize a part or a process that should not exist.

4. Accelerate cycle time. Every process can be sped up. But only do this after the first three steps. In the Tesla factory, Musk has said, he mistakenly spent a lot of time accelerating processes that he later realized should have been deleted.

5. Automate. That comes last. The mistake at Tesla’s factories in Nevada and Fremont was that automation was attempted first, before the requirements had been questioned, the parts and processes deleted, and the bugs shaken out.

Most engineering organizations skip directly to step five. They take a process that should not exist and then automate it. SpaceX runs the steps in order, every time, on every part of the company. When the Algorithm has been run enough times on a piece of hardware, it starts to look like nothing else in the industry.

The Raptor 3 is what it produces when a team iterates on the same engine for ten years. It generates 22% more thrust than the Raptor 2, weighs 40% less, and requires no heat shield because the plumbing and wiring previously hanging on the outside have been fused into the engine’s metal structure through 3D printing. “The amount of work required to simplify the Raptor engine, internalize secondary flow paths, and add regenerative cooling for exposed components was staggering,” Musk has said. “Getting close to the limit of known physics.”

No known engine program in the history of aerospace has iterated this fast. The Space Shuttle Main Engine flew essentially the same design for the last thirty years. The RD-180 that powered the Atlas V is a derivative of an engine designed in the 1970s. SpaceX is on its third clean-sheet redesign of the Raptor in less than a decade, and each version is dramatically better than the last.

The same philosophy applies to people. By mid-2018, Falcon 9 reusability was on a reliable cadence, and Musk turned his attention to the satellite internet constellation that would eventually fund everything upstream. The Starlink team was based in Redmond, Washington, and many of its senior engineers had come from Microsoft, where development moved slower than Musk wanted. In June he flew to Redmond and fired the senior leadership team. He then transplanted young star engineers from the rocket side and gave them a year to launch the first operational batch. It was a cutthroat way to run the company, and from the press coverage of the firings, it looked like the division was imploding. But eleven months later, in May 2019, the first batch went up. Musk had cleared the bottleneck and could move on to the next problem.

This is how he runs everything. In 2018, when Tesla was in the middle of “production hell” trying to scale Model 3 manufacturing and burning cash at a rate that was existential, Musk literally moved into the factory. “I was living in the factory in Fremont and the one in Nevada for three years straight,” he recalled in an interview years later. “I slept on the floor under my desk so that during shift change, the entire team could see me. This is important because if the team thinks their leader is off somewhere having a good time, drinking Mai Tais on a tropical island, it’s demoralizing. Since the team could see me sleeping on the floor during shift change, they knew I was there. That made a huge difference, and they gave it their all.” He later made it a company-wide rule: the more senior you are, the more visible your presence must be.

To find a comparison for how Musk operates as a CEO, you have to go back in history to the industrialists of the late 1800s and early 1900s: Henry Ford, Andrew Carnegie, Thomas Watson, Andrew Mellon, Cornelius Vanderbilt. What makes Musk’s operating style distinctive is his relationship to the work. He reportedly shows up every week at each of his companies, identifies the biggest problem, and fixes it. He does that every week for 52 weeks in a row and then each of his companies have presumably solved the 52 biggest problems that year.

One engineer who joined SpaceX from another aerospace company described the experience as “being dropped into a shocking zone of competence. Everybody around me is so absolutely competent.”

The Constellation

SpaceX looks like a company, but it’s more useful to think of it as the central node of a constellation of companies, all run by the same person, building toward the same long-term mission, and almost impossible to disentangle from one another. Musk has spent over two decades assembling a set of companies that each addresses one constraint that would otherwise bottleneck the others. And they are beginning to compound.

The merger with xAI in February is the epitome of what SpaceX is becoming. If compute ends up in orbit, which is Musk’s bet, SpaceX has the most credible path to deploying it at the scale AI will need. Moving mass to orbit and producing intelligence at scale could be the two defining capabilities of the next few decades, and they now reinforce each other under one roof.

xAI brings Grok, a frontier model that is uniquely positioned on real-time information through its access to X’s data firehose. It also brings the engineers who built the Colossus 1 and Colossus 2 supercomputers faster than many in the industry thought possible.

The Colossus buildout is worth pausing on. xAI took over an old factory in Memphis and had 100,000 GPUs training in 122 days. Once the racks started arriving, it took just 19 days to get the cluster up and running. “From the moment of concept to building a massive factory, liquid-cooled, energized, permitted, in the sort of time that was done, that is superhuman,” Nvidia CEO Jensen Huang said of Musk. “And as far as I know, there’s only one person in the world who could do that. What they achieved is singular. It’s never been done before. 100,000 GPUs is easily the fastest supercomputer on the planet as one cluster [in 2024]. It’s a supercomputer that would normally take three years to plan, then they deliver the equipment, then it takes one year to get it all working.”

A project that would have taken the rest of the industry at least four years took Musk and the xAI team four months.

In May of this year, Anthropic agreed to pay SpaceX $1.25 billion per month for all of the compute from Colossus 1. Weeks later, in an amendment to its IPO filing, SpaceX disclosed that Google will pay $920 million per month for access to 110,000 GPUs, about half the compute Anthropic is getting. Together, the two deals represent roughly $26 billion a year in revenue, from just two customers, for a business SpaceX didn’t have until it absorbed xAI early this year. Chips, power, and land are all scarce, and SpaceX is emerging as one of the few companies with enough AI infrastructure to both lease compute capacity to others and pursue its own ambition of building a leading frontier model.

What xAI gets from SpaceX is a more durable solution to the power constraint that Musk believes will bottleneck AI in the coming years. Producing enough electricity to meet the demand he anticipates for intelligence would require grid buildout, new power plants, and years of permitting the industry doesn’t have. Solar energy in orbit, in his view, is the way out because it’s effectively unlimited. And SpaceX is the only company with a vehicle that can put compute there at scale. Whether he’s right is one of the most important open questions in technology, but SpaceX’s IPO filing shows how seriously the company is taking this bet: it projects AI to be its largest future market by far. The space business that built the company looks almost like a rounding error next to these ambitions.

Tesla is the other major piece of the constellation, and the integration there runs deep in a different way. Tesla and SpaceX share a founder, a talent pool, an operating culture, and an increasingly overlapping set of technology roadmaps.

Tesla provides three things to the SpaceX-xAI side of the constellation. First, chips: the AI5, AI6, and Dojo3, designed in-house at Tesla. Musk has been explicit that these are not just for cars but building blocks of the broader constellation compute stack. The AI5 handles self-driving inference, the AI6 is built for Optimus and AI data centers, and the Dojo3 (paired with a planned AI7) is engineered for orbital compute. Second, robots. Tesla’s bet is that Optimus becomes the physical AI layer for factories, warehouses, homes that want to operate without human labor, and eventually for the Moon and Mars cities Musk envisions. Third is solar. Musk has said that Tesla and SpaceX are separately building toward 100 gigawatts per year of solar-cell production to power the AI buildout on Earth and in orbit.

Then there is the TeraFab. In April, Tesla disclosed that it had begun ordering equipment for a research semiconductor fab on the company’s Giga Texas campus. “This is something we expect to be probably a $3 billion-ish initiative, and capable of maybe a few thousand wafers per month,” Musk told investors on Tesla’s Q1 2026 earnings call. SpaceX, separately, is funding the initial buildout of a much larger facility designed to produce on the order of a million wafers per month at maturity because no existing fab can scale fast enough for what Musk has in mind. And what he has in mind is measured in gigawatts. “This is not a promise of what we’ll do,” Musk said last week. “This is what we’ll try to do and think we probably can do, which is to get to roughly an annualized rate of a gigawatt per year by the end of next year in terms of space AI compute. Then, aspirationally, scale that by an order of magnitude per year. So in two and a half years, hitting an annualized rate of 10 gigawatts a year in space. In three and a half years, maybe 100 gigawatts. Then, depending on progress in chipmaking in the rest of the world and with the TeraFab, going beyond that to scale to a terawatt per year, which is 1,000 gigawatts. That’s twice the electricity consumption of the United States.”

The Gilded Age comparison gets at something real, but it also points to what’s different. Carnegie built steel; Vanderbilt built railroads. Each dominated one sector of the era’s industrial base. Musk is attempting several at once – space, energy, artificial intelligence, robotics, tunneling, brain-computer interfaces, autonomous cars – and bending all of them toward a single goal most people consider fanciful. Whether it all works is genuinely unknown; plenty of it may not. But the attempt itself has no historical precedent and may be the staging ground for a different kind of century.

The World SpaceX Enables

A kilogram of cargo to orbit cost roughly $54,500 on the Space Shuttle before it was retired in 2011. On Starship at maturity, Musk projects $100 per kilogram. When the cost of getting to space falls by a factor of more than five hundred, every industry that could theoretically exist in space starts becoming economically possible. There are many.

The closest historical parallel might be the Transcontinental Railroad. Before 1869, traveling from New York to San Francisco took six months by wagon, cost roughly a year’s wages, and came with a material chance of death. After 1869, the trip took a week. The railroad itself was a remarkable engineering feat, but the real story was everything it enabled: Sears Roebuck, meatpacking giants like Swift and Armour, Standard Oil, and eventually U.S. Steel, which consolidated the industrial empires born during the rail boom.

If the Falcon 9 is the Transcontinental Railroad equivalent for space, Starship could be an upgrade on par with the airplane. The railroad opened a continent. The jet age opened the planet. Starship will open the solar system.

The Industrial Moon

The Moon has been scientifically interesting for as long as humans have looked up at it. It is now becoming economically interesting because it is an entire world made of the raw materials of industry.

Start with how you get things off it. As described earlier, the Moon’s one-sixth gravity and absent atmosphere make a mass driver, not a rocket, the natural way to move cargo off the surface. That changes the economics of shipping completely. Once the track is built, the marginal cost of delivering manufactured goods is dominated by electricity, not fuel, and electricity on the Moon is just sunlight. A package gets flung off the surface, reenters Earth’s atmosphere behind a heat shield, pops a parachute, and lands at a recovery site. At sufficient throughput, the marginal cost starts to look less like spaceflight and more like freight.

Then there’s what you make there. The same lunar regolith that yields silicon and aluminum for solar cells and satellites is the feedstock for an entire industrial base. The space revolution of the 2030s and 2040s could feature autonomous mining vehicles working the regolith around the clock, refineries turning out aluminum and silicon, and factories assembling satellites, solar panels, and the chips to run them. Most industries on Earth have a lunar version waiting to be built, and SpaceX can’t build all of it alone. The people who build the lunar Alcoa, the lunar Caterpillar, and the lunar Union Pacific will be among the titans of the twenty-first century.

Compute in the Sky

The bottleneck on artificial intelligence in 2030 likely won’t be chips but electricity. The obvious response is to build more solar in Texas or Nevada, but this hits a wall faster than people realize. One terawatt of continuous solar power requires roughly 1% of the land area of the United States, and permits for new utility interconnects take a year or more. The xAI Colossus buildout in Memphis required deploying a fleet of temporary gas turbines, fighting state permit battles, and standing up a separate power hub across the state line in Mississippi to bring a single gigawatt online. Scaling that to the hundreds of gigawatts the AI buildout will need is a non-starter. Even the vanes and blades in the gas turbines that would back up the solar are backlogged through 2030.

The solution is to move compute to where the sunlight already is. Once Starship is flying daily and orbital deployment becomes routine, that becomes easier. And the economics improve with the cost curves of rocket launches, solar panels, and chips. “We’re ramping up factories and benefiting from silicon cost reductions, so our costs are going to go down over the next few years,” SpaceX CFO Bret Johnsen explains. “If you look at the terrestrial solutions, the curve is going the other direction. Everything is getting more expensive: the way you’re doing the cooling, power bills are not going down, and land/regulatory is getting more challenging.”

One common objection comes from people who hear “data centers in space” and imagine launching a building the size of Colossus into orbit, but that’s not what it is. “It’s probably about the size of a Blackwell rack, and it has these solar wings that are probably 500 feet long on each side. You keep it in a sun-synchronous orbit so the solar panels are always in the sun,” says early SpaceX investor Gavin Baker. “I’ve spent a lot of time at Starbase over the years and I’ve talked to a lot of SpaceX engineers. I do think it is the most talented group of engineers on planet Earth, and they’re very confident they have solved this.”

In fact, Musk believes that the AI Sat Mini will be easier to build than a Starlink satellite. “You still have some laser links but you don’t need all of the super complex antennas on a Starlink satellite,” Musk explains. “Given the two, the easier one to design for is the AI satellite… There’s not some magic that’s necessary for the AI satellites. A lot of this is technology we’ve already made for the Starlink V3 satellites. We don’t think this is a super hard problem compared to things we already do.”

He projects that within five years, SpaceX will be launching more AI compute to orbit each year than the cumulative installed base on Earth. The math is roughly 10,000 Starship launches per year, or more than one launch every hour, around the clock. By the late 2030s, with the lunar mass driver online, the petawatt threshold comes into view: a thousand times the compute deployed in 2030, launched into deep space at a cadence of one satellite every few minutes.

Mars

The Mars trajectory was supposed to start this year. Musk announced in September 2024 that SpaceX would launch five uncrewed Starships to Mars in the November 2026 transfer window, carrying Optimus robots that would test landing systems, scout for ice, and begin setting up the infrastructure for a future human mission. He said in May 2025 there was a 50-50 chance of hitting it, but that changed earlier this year.

In an X post on February 8th, Musk announced that SpaceX was deferring its Mars timeline and shifting its near-term focus to a self-sustaining city on the Moon. The reasoning was that the Mars launch windows open every 26 months and require six-month transit times, whereas the Moon is reachable every ten days with a two-day transit. “This means we can iterate much faster to complete a Moon city than a Mars city,” he wrote. “That said, SpaceX will also strive to build a Mars city and begin doing so in about five to seven years, but the overriding priority is securing the future of civilization, and the Moon is faster.”

On the surface this looks like a pivot, but it is actually the moment a path to a million-person city on Mars became clear.

The orbital data center thesis, which sharpened over the course of late 2025 and early 2026, gave the Moon a new role. Reaching petawatt-class orbital compute requires lunar mining, lunar refining, and lunar manufacturing of solar panels, radiators, and satellite structures, launched into orbit by a mass driver powered from the lunar surface. An industrial base of that scale requires a permanent population, which requires a city. And that city can be funded entirely by the orbital compute industry while serving as the dress rehearsal for Mars. Every problem SpaceX has to solve to build a self-sustaining city on Mars – radiation shielding, life support, in-situ resource utilization, governance of a permanent off-world population, supply chains across a gravity well – is a problem they have to solve first to build the Moon city. Building the Moon city teaches SpaceX how to build the Mars city with a much faster iteration loop.

The first uncrewed lunar landing demonstration is targeted as soon as 2027, and the Moon city will follow in less than a decade, on Musk’s stated timeline. The mass driver, the lunar industry buildout, and the lunar manufacturing of orbital compute infrastructure all spin up in parallel. Then Mars.

But the hardest part won’t be transporting people. It will be building Mars-side infrastructure that can absorb them. The Moon dress rehearsal will help. So will Optimus. The plan, repeated by Musk in his May 2025 Mars presentation at Starbase, is for the early uncrewed Starships to carry Optimus robots that scout for resources and begin setting up the infrastructure for human arrivals. The company is building a one-million-unit-per-year line at Fremont and a ten-million-unit-per-year line at Giga Texas. The robots are still in early production and have not yet done meaningful useful work in Tesla’s factories, but the production capacity coming online over the next two to three years will be essential for bootstrapping the initial Mars base.

The Sentient Sun

The mission statement SpaceX adopted when it absorbed xAI in February reads: scaling to make a sentient sun to understand the Universe and extend the light of consciousness to the stars.

This is, depending on how you read it, either the most ridiculous thing a serious company has ever put on its mission page or the most honest. We think it’s the latter.

If you squint at the org chart, SpaceX is a launch provider with an internet subsidiary and a recently-acquired AI lab. If you squint at the technology roadmap, it’s the only company on Earth assembling the full prerequisite stack for the post-scarcity transition. If you squint at the mission statement, it’s a serious attempt by one of the most operationally capable founders of our time to push humanity through the bottleneck that ends with us either as an interplanetary species sharing the cosmos with intelligent machines we built, or as a footnote on one rocky planet that didn’t make the leap.

By the time the first child born on Mars asks her parents why their family is there, Starship will have been flying daily for thirty years. The factory down the street will be staffed by Optimus robots running a descendant of Grok that has been self-improving for two decades. The compute that keeps her city alive will come from data centers in space, manufactured from lunar regolith by other robots, launched by a mass driver that has been flinging satellites into deep space at one every few minutes for the better part of a generation. Her parents will have come to Mars on a vehicle named after a starship in an Iain M. Banks novel, because somewhere in the early twenty-first century a man who read those books as a teenager decided to spend his life making them real.

Banks understood something about the people who will choose to go to Mars. The Culture is paradise, but his most interesting characters are the ones who leave it. The civilization solves scarcity, and what’s left over is the human appetite for hard journeys. The frontier is where meaning lives, even when paradise is available next door.

The pitch to early Mars colonists, Musk has said, will be the Shackleton pitch, after the famous recruitment ad placed for the 1914 Trans-Antarctic Expedition: “Men Wanted for Hazardous Journey. Small wages, bitter cold, long months of complete darkness, constant danger, safe return doubtful. Honour and recognition in case of success.” The ad is almost certainly apocryphal, but the story has been retold for a hundred years because it captures something true about those who choose to go.

Why does anyone find this appealing?

“Life cannot just be about solving one miserable problem after another,” Musk says. “There need to be things that inspire you, that make you glad to wake up in the morning and be part of humanity. Earth is the cradle of humanity, and you cannot stay in the cradle forever. It is time to go forth and become a star-faring civilization, to be out there among the stars, to expand the scope and scale of human consciousness. I find that incredibly exciting. That makes me glad to be alive. I hope you feel the same way.”

Subscribe now

This material is solely for educational purposes and is not investment advice or an offer of investment advisory services. This material should not be used as the basis for an investment decision. a16z is an investor in SpaceX through its managed funds, and thus has a financial interest in the company’s performance and future prospects. In particular, a16z benefits if the company grows in value; and a16z funds will receive any customary dividend payments in connection with their status as a shareholder of the company. However, a16z is not being compensated by SpaceX for this material.

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.