A 20-Person Team Burned $2 Billion: AI Rewrites the Ceiling on Capital Efficiency
AI lets a 20-person team spend $2 billion efficiently, and venture capital shifts from zero-sum competition to strategic control points; the frontier labs' funding advantage will eventually fade, and the application layer captures more of the value.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Adding people never sped things up; adding money now does
Martin's example: a well-known multimodal model that everyone uses was built by a team of about 20 people and cost more than $2 billion. Ten years ago, $1 billion would only have bought you a huge headcount and a botched result; today the team knows how to convert capital directly into capability and growth. This is a level of capital efficiency without precedent in the history of engineering, and it redefines the upper bound on how much money venture capital can put to work. The old mythical man-month held that adding people cannot speed things up — but in the AI era capital buys compute and data directly, which is what lets a small team absorb an enormous sum.
— Martin CasadoEarly-stage investing reads position in the stack, not the financials
Early-stage investing cannot be done off the balance sheet, gross margin and churn alone. Martin points out that if you ignore a company's strategic position in the technology stack, you will fail to see the ones that command an enormous transfer of value even while losing money. A company may be a critical control point in the future stack; whether it monetizes itself or gets acquired, the return is enormous. This is the lens that makes the Cursor and OpenRouter deals legible, and it is the dividing line between early-stage investing and late-stage financial thinking.
— Martin CasadoThe case for and against the labs winning everything is equally strong
Martin separates out two paths: the labs win everything, or they don't. For winning it all: over the past three years they have held 95% of the market, they raise more than everything downstream combined, staying just slightly ahead is enough to keep pricing power, autocatalytic effects widen the gap, and they control supply. Against: applications keep expanding into territory that requires services and deep connection to the customer, the open-source model ecosystem is maturing, and cheap capital and GPU supply will eventually rationalize. He says plainly that there are strong arguments on both sides.
— Martin CasadoBy dollars the labs win; by tokens the long tail wins
Martin offers a specific guess: supply constraints ease around 2028; the big labs still take 80% of the market on a dollar-weighted basis, because that is how it has always gone with large incumbents in history; but on a token-weighted basis, 60% belongs to the long tail and open source. That implies the model layer fragments further than people expect, and application companies begin to eat into the labs' share of the profit. He cautions that this is not a settled call — it is, in his words, a complete guess.
— Martin CasadoOpenRouter's value is aggregating the long tail, not smart routing
On the surface OpenRouter is an API router; in practice it is a two-sided marketplace: developers get a single entry point, visibility and analytics, and model providers get distribution and demand. Its value lies in aggregating long-tail supply, not in intelligent routing. Martin says intelligent routing is an AI complete problem; the gains actually being realized in deployment today are cost-performance optimizations — holding quality constant while cutting token consumption.
— Martin CasadoA $200 subscription gets resold to someone else for $20
Martin reveals that highly sophisticated operators in China subscribe to a $200-a-month model plan, burn through the tokens within three days, cancel the subscription, collect a prorated refund for the remaining 27 days, and then resell the full service to other people for around $20. This kind of arbitrage shows subsidies on frontier models being captured by a grey market, and it shows that marketing spend is turning into a precisely calculable financial decision — the CMO's job becoming the CFO's job.
— Martin Casado$60 billion for Cursor does not actually count as a premium
SpaceX acquired Cursor for $60 billion, the largest standalone private acquisition in history. Martin thinks Cursor is worth that on a standalone basis too, because it could very easily raise money at that valuation. Cursor has the data, the product and the distribution; SpaceX has the compute and the resources; both sides believe code is the path to general intelligence, and combined they are worth more. He stresses that this is two companies added together, not any single resource.
— Martin CasadoNo single layer takes it all; every layer is growing value
Martin argues that right now no one layer is eating all the value — applications, models, inference and media are all creating it. This is the largest unlock of wealth he has seen in his career, far beyond the 1990s. His advice to founders is to stop thinking in zero-sum terms and to worry less about moats, and instead to think about what is strategic in this new world. That judgment is the direct explanation for why a16z is willing to pay a premium for strategically positioned companies that are losing money.
— Martin CasadoIn their own words · checked verbatim
what would have happened 10 years ago if I gave you a billion. What would you do, hired a ton of higher you would blow up the whole thing would be like a tunnel mess. right. And so now we actually know what to do with that money.
Martin Casado4:53
I think there's basically two paths that are meaningful to talk about. One of them is like the labs win everything. And then the other one is the labs don't win everything.
Martin Casado11:41
I think that you need a two sided marketplace. And in many ways, you can view one of the products of open router is the demand.
Martin Casado21:10
But what we've never been able to do in the history of this industry is put in $10 and get anything back. … But now it really is $10 in and then some amount out pretty directly.
Martin Casado26:11
one has the data, One has the compute, One has the distribution
Martin Casado28:11
Figures
| Team size behind the well-known multimodal model | About 20 people | 5:57 |
| Training cost of that model | More than $2 billion | 5:57 |
| Market share held by the labs | 95% | 11:41 |
| When supply constraints ease (prediction) | Around 2028 | 16:09 |
| Predicted dollar-weighted share for the big labs | 80% | 16:09 |
| Predicted token-weighted share for the long tail and open source | 60% | 16:09 |
| Cursor acquisition price | $60 billion | 28:11 |
| OpenAI monthly plan price | $200 | 25:00 |
| Share of a founder's time spent on recruiting | 30%-40% | 31:11 |
Glossary
- autocatalytic effects
- Using AI tools to accelerate AI's own R&D, forming a positive loop between capital and capability.
- AI complete
- A problem that can only be solved once you have general artificial intelligence — model routing, for example.
- token path
- The intermediate layers a token flows through during inference, such as an API gateway; a strategic choke point.
- two-sided marketplace
- A platform connecting the supply side of models with the demand side (developers and users).
- founder-market fit
- A founder's experience and endowments matching closely what the target market needs; central to an investment call.
How to listen
AI founders, early-stage investors, and corporate strategy and M&A teams — anyone trying to understand the competitive landscape at the model layer and the new limits of capital efficiency.
The opening discussion of academic credentials and the small talk are skippable; the substance starts at 4:53 with capital efficiency.