May 31, 2026
Future of $900B US AI Labs
tl;dr: with companies being this big, what will the world look like?
A few days ago, Anthropic announced Series H with a market cap of $900B. For reference, Uber is $145B, and Google is $4.6T. OpenAI will probably raise and be similar. And whey they IPO, their market cap will most likely jump.
But what will the world look like after this? After 1-2 years? I thought about this from first principles.
One. The AI labs will need to increase their revenue/profit (whether they IPO or not) to justify their valuation. With the types of datacenter deals signed and current business model on their compute, if they stay the course, these labs will be out of business (tldr: they are burning a lot of cash). Anthropic did say they finally achieved net profitability, but I would like to see them doing at least 3 consecutive net-profit quarters before buying this. Also, because the GPU they are contracted to buy (don’t have them yet) become obsolete in 5-6 years, with the current slow progress of datacenters being built, their GPU they promised to buy may only be available online as they become obsolete.
Two. The promise of AI. I read Simon Willison’s article that the AI labs may finally have found product market fit with the coding models. (tldr: coding agents for enterprise b/c if they can master coding, they can control computers, which is most of enterprise job functions). And enterprises have spent TON of money on enterprise ClaudeCode/Codex. However, what did that result in? Have the companies launched more projects or launched them faster due to the engineers using claude code? I haven’t yet seen concrete examples of companies and/or products that have seen this benefit due to agentic coding. (it’s good don’t get me wrong). So I wonder at what point the enterprise customers will say: is the $500MM ClaudeCost budget worth it? And I agree with his article: “consumer AI” aka chatbot didn’t seem to be the moneymaker or has a good product market fit given the cost. I think the future of AI lies in B2B.
Three. Chinese AI. Chinese OSS models are about 8 months behind Anthropic/OpenAI, but by this statement, in 8 months, the Chinese OSS models will be as good as the current Anthropic/OpenAI models. And I would posit that the gap will only get smaller (2nd mover advantage). Then the enterprises may say: ClaudeCode/Codex is too expensive and I’m happy with sloppy third or fourth. They may say: let’s just build our own AI data center and put Chinese OSS models; we also get the benefit of our proprietary data never touching the AI labs. Then the AI labs may find themselves footing the bill for compute while their revenue declines.
This implies their models have to be so good that the enterprise consumers will not move off. But I think LLMs have reached diminishing returns (e.g. most of the pre-training knowledge cuts off at Jan 2025). And what they promise to the enterprise customers have to be so good to justify the high cost of their services.
Four. A good strategy for AI labs then should be: copy/kill other AI unicorns. Naively speaking, copy decacorn/unicorns and “add” their valuation onto top of theirs. I haven’t seen them acquiring unicorns yet. Distribution is a thin moat, and they can just build their products easily using claudeCode/Codex. We’ve seen this with legal (Harvey/Legora), search (Perplexity), website design/builder (Lovable).