‘I‘m just stuck’: Meet the former OpenAI researcher sitting on $700K of equity that he says is overvalued

Andrew Ho spent eight months at OpenAI before walking out the door on a Wednesday to start his own company selling reinforcement learning datasets to frontier labs. Then, around 5 a.m. the next morning, he did something that stunned his former colleagues: he told them to sell their shares while they still could. Not exactly the parting gift most people leave behind.

His timing was almost cruel. The posts landed just as the Nasdaq 100 slid into correction territory, and within hours hundreds of thousands of people were reading his warning that frontier lab valuations look, in his words, unlikely to double after IPO but very plausibly capable of dropping by half. Ho himself is holding roughly $700,000 in OpenAI equity he legally can't touch until after the IPO and lockup expire, which makes his own advice a little painful to give. He knows it. "I'm just stuck," he told Fortune.

His real argument isn't that AI is a bubble about to pop overnight — he still expects inference demand to explode and compute to get scarce. What worries him is the economics underneath: labs are spending more each training cycle for gains that evaporate fast, as rivals like Moonshot's Kimi distill their way to comparable performance for pennies on the dollar. Revenue climbs with every new model, sure, but nobody knows if it climbs fast enough to cover debt-fueled spending that keeps accelerating. Miscalculate the timing even slightly, he says, and that gap can sink a company.

The optimists at his old lab are banking on recursive self-improvement — AI systems smart enough to design their own successors, sending capability curves parabolic. Ho isn't a believer. He argues the bottleneck was never raw intelligence; it's "research taste," the judgment calls about which experiments matter and which results are noise. Models excel where answers are cleanly verifiable, like math proofs or compiling code, because that's where billions in spending have been aimed. Everything murkier and more judgment-based remains stubbornly hard, which is precisely the gap his unnamed new startup wants to fill with datasets for long-horizon scientific reasoning.

That puts him at odds with bulls like podcaster Dwarkesh Patel, who's been arguing compute could get ten times pricier while revenue rises just as much thanks to RSI. Ho instead sides with the Wall Street skeptics who dumped Meta and Google shares recently over AI spending fears. In his view, the real winners of this race are Nvidia and Micron, who get paid regardless of who wins the model war, while labs themselves face two brutal options: build their own chips to break Nvidia's grip, or push up into applications and start capturing more value directly, the way Meta grabbed Cursor.