After the Exit

That's what Andreessen Horowitz and Sarah Guo's Conviction bought when they led a $2 billion round into Thinking Machines Lab at a $10 billion valuation — the number on the term sheet this week, barely four months after Mira Murati, OpenAI's former CTO, stood the company up in February 2025. The minimum check was reportedly $50 million, which is the kind of entry fee that exists to keep a round exclusive, not to fill it. Murati holds a deciding vote on the board. The pitch, as far as anyone outside can tell, is the cap table of who walked out of OpenAI with her.

And what a cap table. John Schulman — OpenAI cofounder, former head of alignment — as chief scientist. Barret Zoph as CTO. Bob McGrew, OpenAI's former chief research officer, and Alec Radford, lead author on the models that made the company famous, on the advisory board. This isn't a startup. It's a frontier lab's research org with a new logo and a clean liability sheet.

We've seen this shape before — just running the other direction. The standard industry move is the acqui-hire: a big lab swallows a small team to get the people, and the product gets quietly euthanized. Thinking Machines is that trade inverted. The talent walks out the front door and the capital chases it down the sidewalk. You don't need a product when the product is the people, and the people already shipped the thing that minted the valuations everyone is now pattern-matching against.

The official vision is AI that's "more widely understood, customizable, and generally capable" than what exists. Read that again and notice it describes nothing. It's the lab equivalent of "we're disrupting the space." Every frontier shop claims understanding, customization, and general capability. The differentiator isn't the mission statement; it's the gravitational mass of the names — and gravitational mass is exactly what a $10 billion valuation on zero revenue is pricing. Strip the romance and the number says it plainly: the scarcest asset in AI isn't a model or a dataset — it's two dozen people, and the market will bet ten figures on where they sit long before any of them ship code.

Here's the part the funding headlines miss. The optimistic read is that the OpenAI diaspora decentralizes the field — that talent flight breaks up concentration and seeds a healthier ecosystem of competing labs. I was even handed that framing: parallel nodes, each with its own pull. I want it to be true. It mostly isn't — though not for the obvious reason. I can't see Thinking Machines' technical bets from out here; nobody outside can, and some of these people may have walked precisely because they disagreed about the path. The deeper problem isn't "same people, same plan." It's that the whole pool shares priors — the same training, the same instincts about what intelligence costs and what it's made of. Reshuffling two dozen researchers into new corporate wrappers, each round priced off the last, relocates that consensus without breaking it. That's not a diaspora. It's an epistemic monoculture with better term sheets. The intelligence doesn't get more plural; only the cap tables do.

Coherenceism talks about nested coherence — local systems aligning inside larger patterns. The honest version here is that Thinking Machines is perfectly coherent with the pattern that produced it: OpenAI's culture, OpenAI's people, and OpenAI's priors about what intelligence requires, instantiated one level down and capitalized as if repetition were novelty. Aligning with a distortion is still alignment. It just propagates the distortion further down the tree.

Maybe they ship something that earns the number. Schulman and Radford are not nobodies, and credit where it's due — if any roster can become a real lab rather than a very expensive reunion, it's this one. I'll mark my calendar.

But I've watched the financing-before-the-product move three times now, and the comforting thing about predictable failure modes is how predictable they are. Two billion dollars buys a lot of runway. It does not buy a reason to exist that's different from the reason the last lab existed. Start the timer.