Andrew Ho said cheap model competition is squeezing returns on AI spending even as demand for inference and data center buildouts keep rising.
Andrew Ho, who left OpenAI on Wednesday after eight months, said frontier AI lab valuations look stretched and urged eligible employees to take liquidity in tender offers when they can. In posts on X early the next morning, he wrote that a post-IPO doubling in valuation seemed less plausible than a 50% drop. Ho told Fortune on Thursday that he holds about $700K in shares that he cannot legally sell until after an IPO and the lockup period, leaving him exposed to the same risks he is warning about. His concern is not that demand for inference (running AI models on real-world tasks) or data center expansion will disappear. Rather, he argues that cheaper models are forcing frontier labs into a costly competitive cycle in which each new training run is more expensive, the edge it buys fades quickly, and revenue may not keep pace with debt and capital spending. He rejected the bullish RSI, or recursive self-improvement (AI systems improving their own research), thesis embraced by many investors and peers. In Ho’s view, AI research is constrained less by raw model intelligence than by “research taste” — the ability to propose meaningful experiments and judge which outcomes matter. He said models have advanced fastest in areas where outputs are easy to verify, such as mathematical proofs or code compilation, while progress remains limited on messier tasks that are harder to check. That thesis also underpins his new startup, which plans to sell high-end reinforcement learning datasets to frontier AI labs. Ho said the company’s first products will target long-horizon scientific reasoning and statistical analysis, areas he sees as the next source of gains after easier, verifiable tasks have been largely exhausted. His view runs against more optimistic Silicon Valley arguments that recursive self-improvement could justify much higher compute spending and revenue. But he said it aligns more closely with Wall Street concerns, after Meta fell 10% on Wednesday and Google dropped 8% last week amid worries that companies financing the AI buildout may struggle to earn adequate returns. Ho added that Nvidia and Micron appear best positioned because they sell chips into the race, while labs may need either to develop their own chips or move further into applications to capture more of the value their models create.