AI made startups capital-bound: 20 people can now spend $1 billion

1. Math splits the audience — ▶ 1:26

   The Claude-and-Riemann episode divides observers into two camps. One is thrilled that hard problems are being solved whether or not they follow the math. The other calls it fake or a job-killer. Steven Sinofsky points to the confusing part: the group most excited is mostly mathematicians, the people most affected. That undercuts the "dawn of idiocracy" story. Casado later says the reaction is mixed: some mathematicians say AI takes the 20% of the job they never liked, while others face an existential crisis. Whether any real progress was made on the hypothesis stays unclear.

2. No market behind it — ▶ 2:47

   Casado applies an economic test. The summed postdoc salaries spent on these long-open problems are probably small. Decades of being unsolved says little about capability, because nobody was paying for the answers. A postdoc at 30k a year for 5 years is not the market deciding this is the thing to unlock. AI also suits an almost purely axiomatic domain. The solutions often borrow math the human author did not know, which points to problems needing breadth beyond most humans. He has not yet seen a link to a roadblock for useful tasks, and that is what he is waiting for.

3. Games, not reality — ▶ 4:40

   Sinofsky argues math is a leading indicator of market interest. He recalls an AT&T linear algebra method that cut the United Airlines flight-map calculation by 3 hours, and linear algebra now underpins AI. Casado is unconvinced: it is "really good at playing a game", the best StarCraft player ever. Pitches split. Some founders say the foundation of AGI is math because the universe rests on mathematical principles. Others say it tells you nothing about reality. Casado's position is that more work remains and this is not just about getting better at math.

4. Four color theorem lesson — ▶ 8:51

   Sinofsky recalls John Hopcroft's P=NP course, built on the idea that solving one problem in polynomial time would crack routing and traveling salesman. That mattered when a 6,000-switch AT&T node meant two years of simulation. The four color theorem was never proved by a calculus-style derivation. The potential cases were shown to be finite, about 200 pages of combinations, and a computer checked them all. It was a "bankshot proof" only compute allowed. His lesson is that a new abstraction makes a whole class of problems solvable, and people then build tools at that level instead of starting from 2-3 trees.

5. No bridge to physics — ▶ 12:06

   Casado worked on large simulation codes for exploding stars and airplanes in wind simulators. All of them rested on empirical equations of state. He wonders whether simulation is computationally irreducible, so you must actually run it, and whether AI's math answers touch that at all. Algorithmics, modeling or logistics might benefit. The claim that "if it can solve all math you can predict anything" is a "huge huge logical leap" with no indication it is true. Sinofsky's reply is that AI may not be the next math tool but may lead to a new level of model, as slide rules led to calculators.

6. People react to change — ▶ 16:37

   Sinofsky's counter is that people react to change, not baseline. Nobody complained about calculus, but the TI-85 put math teachers into crisis. In 1983 two students brought an Apple II and a 25-pound batteryless Osborne into Harvard Law exams, and the school banned them. He needed the dean's permission that fall to write freshman English papers on his own computer. Three years ago Cornell stopped talking to him when he proposed AI in freshman writing. Graphing felt like cheating, he says, because the capabilities matched the test question.

7. Abdicating logic — ▶ 30:24

   Casado's rebuttal is that across computing the human defined the problem and the machine supplied compute, network and storage. Correctness stayed with the programmer, even with third-party libraries whose inputs and outputs are known. Now users hand the logic itself to a third party and ask it to tell them the answer. Imperative programming gives the recipe and declarative programming (SQL, Prolog) gives the end state. This does neither, and the end state is often unknown. Sinofsky agrees it feels different, calling it a leap from calculating to something statistical, with 1980s expert systems as the first taste. Casado says fundamentals may need rethinking.

8. Capital replaces engineering — ▶ 39:06

   Twenty years ago 10 people with $1 billion would buy computers and blow up. Ten years ago they would hire engineers, but the mythical man-month bites. Now 20 people can use it productively. Sinofsky notes computing was capital-bound for its first 30-40 years, then engineering-bound, and is now capital-bound again. Casado spent 10 years giving early teams money, helping them recruit and waiting two years for engineering to catch up. He says "engineering just doesn't scale". He endorses Patrick Collison's point that a discipline now exists for deploying huge sums with small teams. This changes the nature of capital, competition and defensibility.

9. Capital grows the market — ▶ 42:50

   Casado calls the "too much capital chasing too few deals" view zero-sum thinking. More capital in private markets grows the TAM for two reasons. Technical waves like AI can absorb capital, and companies stay private longer, so more value accrues privately. Sinofsky adds that software now reaches unserved domains. He recalls a doctor who wrote a DOS scheduling program that listened for keywords to book blood draws, X-rays and EKGs in parallel, and it took him years. Now the domain expert's path is a capital problem, not a hunt for a co-founder, and everyone else's abstraction layer is rising too.

10. Why incumbents don't crush — ▶ 46:30

    Casado argues AI solves the demand problem: token and GPU demand is so open-ended that a startup can decide how much to spend to drive top-of-funnel growth. Combined with huge raises, that explains the meteoric growth of Cursor, Anthropic and OpenAI. Sinofsky says disruption is cultural and constant. Scorecards, field sales, compensation and 500,000 customers bind big companies, and they only watch each other: Microsoft worries about Amazon and Google. Google has the data, yet its models are getting trounced. Casado bets the cause is cultural, plus capital constraints such as off-balance-sheet bond deals and rationed tokens.

11. The $20 billion artifact — ▶ 55:39

    Casado says we know mechanically how models work. They are in-distribution, transfer learning is probably absent, and the singularity is probably not here. Nobody knows what a $20 billion artifact can do, though. What he got wrong was assuming scaling would stall. Laws are holding, so a $100 billion training run is conceivable, and the funding machinery could aim at curing cancer or at a weapon. He wants the debate to shift from fast takeoff to concentration of resources. Sinofsky adds that drug discovery's hard part is efficacy and safety in human patients, and says his own 58 internet memos were mostly wrong.

The binding constraint on building in AI is no longer engineering. It is money. Give 20 people $1 billion today and they can spend it usefully. That was never true before. Twenty years ago, 10 people with $1 billion would have bought hardware and watched it vanish. Ten years ago they would have hired engineers and run straight into the mythical man-month. Martin Casado describes the shift as one from an engineering-bound industry to "a capital problem", and he says it is fundamentally different. His claim is that this rewrites priors built over 40 to 50 years: where value accrues, how competition works, what defensibility means.

The history sharpens the point. Computing was capital-bound for its first 30 to 40 years, because the first step with a computer was to acquire one. Then it became engineering-bound. Casado spent a decade giving early teams money, helping them recruit, and waiting two years for the engineering to arrive. His verdict is that "engineering just doesn't scale". The old argument between lean and fat startups was settled by that limiter, not by taste. Patrick Collison's observation, which Casado endorses, is that there is now a discipline for putting enormous sums into small teams productively. Casado thinks the industry has not internalized it.

Capital as the constraint changes who wins. Incumbents have always had cash flow and distribution. Casado argues AI erodes the second. Demand for tokens and GPUs is so open-ended that a startup can simply decide how much to spend to drive top-of-funnel growth. Add the ability to raise vast rounds, and Cursor, Anthropic and OpenAI stand on footing close to Microsoft and Google. Steven Sinofsky supplies the other half. Disruption is cultural, and culture is a constant. Scorecards, field sales, compensation, org structures and 500,000 existing customers fence in what a big company can do. Big-company people always assume they will crush the little ones, and then realize "they never get crushed". Incumbents watch other incumbents. Microsoft worries about Amazon and Google, not startups. Casado points to Google as the test case. It has the data and the intelligence, and its models are getting trounced. He bets the cause is cultural and notes the capital strain: bond deals moved off balance sheet, tokens rationed to enterprise customers over internal products.

Two more consequences follow. The venture complaint of too much capital chasing too few deals is zero-sum thinking, Casado argues. More capital in private markets grows the TAM, because AI can actually absorb capital and companies stay private longer. And software reaches domains it never touched. Sinofsky's first customer visit was to a doctor who wrote a DOS scheduling program that listened for keywords to book blood draws, X-rays and EKGs in parallel. It took him years. Now the domain expert needs capital, not a co-founder.

The obvious objection is that this is the old story. Sinofsky says it is. Every new abstraction triggered a crisis among those who owned the old one. Calculus became baseline and nobody complained about it. The TI-85 sent math teachers into panic. In 1983 Harvard Law banned laptops from exams, and Sinofsky needed a dean's permission to write freshman English papers on his own computer. The four color theorem was never proved by derivation. Its cases were shown to be finite and a computer checked them all. A new level of abstraction opened a class of problems and let people build tools above it. People react to change, he says, more than to baseline.

Casado is not persuaded that this layer is like the others. In all of computing the human defined the problem and the machine supplied compute, network and storage. Correctness stayed with the programmer. Imperative code gives the recipe. Declarative code gives the end state. This gives neither. You often do not know the end state, and you tune your words until the output is useful. If that is a new layer, it does not map down deterministically to the one beneath it. Sinofsky concedes it feels different. He calls it a leap from calculating to something statistical, and notes the 1980s expert systems were the first taste of the debate.

The math breakthroughs are where Casado tests the whole story, and he finds them thin. His yardstick is economic utility. The summed salaries of postdocs who worked these problems are small, so solving them does not show AI unlocking value. Math is good at being a game: the best StarCraft player ever. The leap from solving math to predicting reality is the weakest link. He worked on simulation codes for stars and airplanes, and all of them rested on empirical equations of state. Whether simulation is computationally irreducible, and whether any of this helps, is not clear to him. Sinofsky replies that mathematics has historically been need-driven. Tides, missile tables and the space race are his examples, and the open question is whether this math blocks anything economic.

The strongest counter to Casado's own skepticism comes from Casado. The mechanics are known. Models stay in distribution, transfer learning is probably absent, and the fast-takeoff story has not materialized. But a $20 billion artifact, or a $100 billion training run, is something nobody can predict. He admits his error was assuming scaling would stall. Sinofsky adds a caution from biomedicine: candidates were never the bottleneck, efficacy and safety in human patients are. The danger Casado names is concentration. The same machinery could aim at cancer or at a weapon. Capital turns previously infinite problems into finite ones. Who holds that capital, and what they point it at, is the part nobody on the stage claims to know.