The Lockup That Concentrates the Mind
Andrew Ho spent eight months at OpenAI, left on Wednesday, and by 5 a.m. Thursday was posting unsolicited financial advice to the colleagues he'd left behind. The message was direct: take liquidity now, if you can.
'I would strongly recommend taking liquidity if you're eligible for tender offers,' Ho wrote on X. '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%.'
Ho told Fortune he holds roughly $700,000 in OpenAI shares that he is legally barred from selling until after the IPO and its lockup period expires. 'So I'm just stuck,' he said.
The posts landed in the middle of a tech selloff that had pushed the Nasdaq 100 into correction the day before, and drew hundreds of thousands of views. 'It was kind of remarkable,' Ho said.
The Red Queen's Race
Ho's core concern is structural. He believes demand for AI inference will rise dramatically and that a compute crunch is coming — the data-center buildout, in his view, is real and necessary. But competition from cheaper models is forcing frontier labs onto an accelerating spending treadmill, and revenue, he argues, will not keep pace.
Analysts call it a 'Red Queen's race': run as fast as you can just to stay in place. Each training cycle costs more than the last, the competitive advantage it buys is temporary, and rivals such as Moonshot's Kimi can close capability gaps through distillation at a fraction of the cost. 'If you miscalculate even by just a very fine amount,' Ho said, 'that can be the difference between life or death for a company.'
Why He Doesn't Buy the RSI Thesis
Most investors and researchers at frontier labs remain optimistic, Ho acknowledged, because they believe recursive self-improvement — RSI, the idea that an AI system can improve itself by running its own experiments and training its own successors — is imminent. If each model version builds a slightly better successor, the intelligence curve, and the compute bill, could go parabolic.
Ho is skeptical. Research, he argues, is not limited by raw intelligence but by 'research taste' — the judgment to propose the right experiments and recognize which results matter. 'You can't reason your way to phenomena,' he said. Models have advanced fastest where outputs are easy to verify: a proof is valid or it isn't, code compiles or it doesn't. Everything harder to verify is, in his words, still stuck.
That gap is exactly the business he is building: selling high-end reinforcement learning datasets to frontier labs for judgment-heavy tasks, starting with long-horizon scientific reasoning and statistical analysis.
As for who is winning the current race, Ho is blunt: 'The people making out really happily here are Nvidia and Micron' — the chip sellers, not the labs doing the buying.
CEO Times Read
Ho's warning is a data point Wall Street has already started pricing in. Meta dropped 10% and Google shed 8% last week on fears that the companies funding the AI buildout will not recover their capital. The market has already voted, at least provisionally.
For free-enterprise readers, the episode illustrates a durable principle: capital locked behind regulatory structures — in this case IPO lockup agreements — cannot respond to new information. Ho cannot sell; the market cannot clear. When the IPO does arrive, that pent-up supply will arrive all at once. Employees who believed the headline valuation without stress-testing the underlying revenue math may discover that the most expensive lesson in finance is the one you can't exit.
