An Ex-OpenAI Engineer Says a $1 Trillion AI Lab Is Mathematically Impossible. He Left to Prove It.

Andrew Ho joined OpenAI's technical team in late 2025 and left eight months later. He walked away with roughly $700,000 in OpenAI equity that he cannot currently sell, and is now publicly telling his former colleagues to do the opposite of what he did: take the next tender offer rather than wait for the IPO. The qbitai piece that surfaced his posts and Fortune's English-language writeup both carry his blunt message to OpenAI staff who still hold vested grants, in the original Chinese: "能参加要约收购的，赶紧把钱拿到手，别等IPO了." Translation: take the next tender offer while you can, do not wait for the IPO.

OpenAI ran a roughly $6.6 billion employee tender offer in October 2025, with more than 600 staff selling shares and many becoming paper millionaires on the spot, according to Sina Finance's retrospective on the round. A tender offer is a private secondary: a buyer or group of buyers agrees to purchase a set of shares at a set price on a set date, and existing employees choose whether to sell. The trade-off is that the buyer prices the deal at a discount to the most recent primary round, and the employee accepts that haircut in exchange for liquidity today. Ho's case is that the post-IPO alternative is worse than the tender discount. After listing, employee shares are typically locked up for months. When the lockup expires, a large pool of newly liquid sellers hits the market at the same moment, and the public market prices that overhang in early. The first day the lockup drops is often the first day the stock has to absorb a coordinated, concentrated seller.

His warning is built on specific arithmetic, and it is bullish on the technology while bearish on the public valuation of frontier labs. The math, as carried in the Chinese coverage, runs like this. Take an 80% gross margin, the number the labs like to point to, and a 20x price-to-earnings multiple, the number the public market usually applies to high-quality software businesses. Multiply them by a $1 trillion equity value and the implied annual revenue requirement is between $100 billion and $200 billion. No frontier lab has a credible path to that number, because the cost of training the next generation of models is rising faster than the revenue the current generation is producing, and because open-weight and cheaper competitors keep pushing API prices down. The math is the news.

The labs cannot stop training, because the moment one of them pauses, Anthropic or a cheap open-weight model catches up and the leader's pricing power evaporates. Every generation of model costs more to train than the last. At the same time, a third lane of well-funded open-weight competitors keeps pushing API prices down, so the revenue the current generation earns is itself under pressure. The result is an industry that has to keep spending more to keep earning roughly what it earned before. That is the Red Queen problem in cash-flow language.

After leaving OpenAI, Ho started a company that sells reinforcement-learning training data to frontier labs. The bet is that the next competitive bottleneck is not compute, where the labs are already over-funded, but high-quality post-training data. He projects that frontier labs will collectively spend more than $100 billion on that data over the next several years. Short the public-market valuation of the labs, long the inputs they will still need. The two positions only look contradictory if you think the public market is already pricing in the input demand.

Ho's posts landed the same week the Nasdaq 100 entered a technical correction, with Meta down more than 10% in a single session and Google down more than 8% across a week. The private-market valuation of OpenAI has been the test case for the broader category, and OpenAI's path to a public listing now runs through a Musk lawsuit, congressional and governance scrutiny, and the Microsoft-Nvidia stakeholder dynamics catalogued in TradingKey's Q4 2026 IPO analysis.

The next test of Ho's argument is the next OpenAI secondary. If the round closes at a price that is steady or up versus the October 2025 deal, Ho's warning is just an ex-employee's view. If it prices flat or down, or if the company widens the discount to clear the book, the trade he is recommending starts to look like the right one, and the public-market window he is warning about gets a date.