At around 5 a.m. on Thursday, still awake after leaving his job the day before, Andrew Ho typed out a warning for the people he’d just walked away from: sell while you can. The former OpenAI researcher, eight months into the company and now out the door, told colleagues on X that the frontier AI labs are overvalued. “I would strongly recommend taking liquidity if you’re eligible for tender offers,” he wrote. “It seems somewhat implausible that the valuation is going to, like, 2x after the IPO, but it does certainly seem plausible that it could go down by 50%.”
By the time he woke up, hundreds of thousands of anxious eyes had found the posts. They landed in the middle of a tech selloff that had already dragged the Nasdaq 100 into correction the previous day. “It was kind of remarkable,” Ho told Fortune on Thursday.
Here’s the twist that makes his advice sting. Ho is holding roughly $700,000 in OpenAI shares he legally can’t touch until after an IPO and the lockup that follows it. “So I’m just stuck,” he said. He was, in effect, talking against his own book.
The treadmill that never slows down
Ho isn’t a pessimist about AI demand. He expects inference to climb sharply, a compute crunch to arrive soon, and the data center buildout to be genuinely necessary. What keeps him “paranoid” is something narrower: cheap models are forcing frontier labs to spend faster and faster, and revenue may never quite catch up.
Analysts have a name for this, borrowed with a bit of flourish from Lewis Carroll. A “Red Queen’s race,” where you sprint at full speed just to hold your position. Every training cycle costs more, the edge it buys is temporary, and rivals like Moonshot’s Kimi can close the gap for a fraction of the price through distillation. Revenue will rise with each new model, Ho concedes. The question is by how much. With all the debt these labs are carrying, he said, “if you miscalculate even by just a very fine amount, that can be the difference between life or death for a company.”
Why the self-improvement dream leaves him cold
Many of his former peers stay calm because they believe a breakthrough is close. They call it recursive self-improvement, or RSI: an AI good enough at AI research to run its own experiments and train its own successors, each version bootstrapping a slightly better one until the curve goes parabolic. Ho doesn’t buy it. Research, he argues, isn’t bottlenecked by raw intelligence but by “research taste.” Models can’t reliably propose good experiments or tell which results matter. “You can’t reason your way to phenomena,” he said.
There’s a pattern to what AI does well, and it isn’t random. Progress has been fastest wherever answers are easy to check. A proof holds or it doesn’t. Code compiles or it doesn’t. That’s where the billions have gone, and it has worked. Everything harder to verify remains stuck, which is exactly the gap Ho’s day-old, still-unnamed company means to fill by building datasets for judgment-heavy work, starting with long-horizon scientific reasoning and statistical analysis.
Who actually wins the race
Ho’s view puts him against podcaster Dwarkesh Patel, who argued this week that RSI could make compute 10x more expensive and the labs 10x more revenue. It puts him closer to Wall Street, which knocked 10% off Meta on Wednesday and 8% off Google the week before, spooked that the AI buildout won’t pay for itself. The clear winners, in his telling, are Nvidia and Micron, selling chips to everyone in the race. Labs have only two exits, both brutal: build their own silicon to break Nvidia’s pricing power, or climb into the application layer and sell products directly, the way Meta bought Cursor. Right now they capture a sliver of the value their models create, a gap Ho calls “very undercapitalized at the moment.”
For now he’s fielding heavy interest in the new venture and trying not to stare at the locked equity that makes up a huge share of his net worth. His advice to everyone still inside is oddly gentle. “You’re going to work,” he said. “Don’t think too hard about these questions. Which is maybe for the best.”
For more coverage of frontier AI economics, visit Mylistingo.
Source: Fortune
