Startup Fortune reports that leaked OpenAI financials show the numbers behind the AI bubble finally cracking.

OpenAI burned nearly $21 billion last year to bring in $13 billion, and this week a fresh revenue correction hit chip stocks hard. The math investors were told to trust just got much harder to believe.

For months, the loudest warnings about an AI bubble were dismissed as the work of doomsayers who didn't understand the technology. This week, OpenAI's own numbers made their case for them. Leaked financial statements obtained by journalist Ed Zitron and independently verified by the Financial Times show that OpenAI's 2025 revenue reached $13.07 billion, up sharply from $3.7 billion in 2024. On its face, that is strong growth. But operating losses hit $20.92 billion, against total costs and expenses of roughly $34 billion. The company spent more than $2.50 for every dollar it brought in.

Then, on October 8 and 9, the second shoe dropped. Reuters and the Financial Times reported that OpenAI's annualized revenue as of late September was about $50 billion, not the $70 billion figure that had circulated among investors for months. Technology stocks sold off on the news, with chipmakers tied to the AI buildout, including Nvidia, taking the hit alongside the broader Nasdaq.

The $70 billion number was never really OpenAI's own revenue. The company's finance team had used a methodology that folded in gross revenue from partner companies, a approach that put OpenAI's growth rate on the same footing as Anthropic, which counts cloud-partner sales toward its own totals. Strip that out, and you get the $50 billion figure investors are now working with. OpenAI says it still expects to reach or exceed $70 billion in annualized revenue by the end of 2026, pointing to 77% run-rate growth across the business and 107% growth in its enterprise segment. That may well happen. But a company that is revising its headline number down by $20 billion, in the same period its audited books show a $21 billion operating loss, is not in a position to ask for the benefit of the doubt.

The spending breakdown in the leaked documents makes the bubble story concrete. Research and development alone cost OpenAI $19.18 billion in 2025. Sales and marketing added another $5.73 billion. A large share of this spending flowed straight back to Microsoft: OpenAI paid its biggest backer and infrastructure partner $17.2 billion across the year, split among R&D compute, cost-of-revenue charges, and other costs. That is not an ecosystem of independent winners. It is one company's growth numbers propping up another company's revenue line, and both propping up a stock market narrative.

The net loss, once the accounting effects of OpenAI's conversion from nonprofit to for-profit are included, swelled to $38.5 billion. That figure was driven by a $41.55 billion non-cash charge tied to changes in the fair value of convertible interests and warrant liabilities. Excluding that one-time accounting item, the adjusted net loss is closer to $8 billion. But the operating loss, the actual cash burned running the business, was real, at $20.92 billion. No accounting adjustment makes that number disappear.

This is what a bubble deflating looks like in real time: not a single crash, but a steady stream of disclosures that force a market-wide repricing. In June, leaked books revealed the scale of the burn. In October, the revenue figure turned out to be smaller than advertised. Each revelation could perhaps be explained on its own. Taken together, they describe a company whose growth story was told in the most favorable light possible, and whose actual financial position does not support the valuations built on top of it.

OpenAI's most recent funding round, closed in March, valued the company at $852 billion. Justifying that figure requires believing the company can grow revenue fast enough, and cut losses fast enough, to eventually earn a multiple that most software companies never approach, even at maturity. A $13 billion revenue base that burns $21 billion a year, built substantially on payments to a single infrastructure partner, is a shaky foundation for that belief. The problem is not that OpenAI's product doesn't work. ChatGPT has real and enormous usage. The problem is that the price the market has been asked to pay for that usage may never be supported by the cash the usage actually generates.

Nvidia, Oracle, and the rest of the AI infrastructure stack have spent two years telling investors that demand for compute is effectively limitless. OpenAI's own books are now the clearest evidence yet that demand for compute and demand for a company's output are two very different things, and that the gap between them is being paid for by investors, not customers.