OpenAI lost $20.92 billion on $13 billion in revenue in 2025, and its 2026 burn rate is tracking toward $28 billion , but the real beneficiaries of that spending aren't building models at all.
The documents weren't supposed to get out. Audited financial statements obtained by blogger Ed Zitron and independently verified by the Financial Times landed on June 16, and the numbers confirmed what many in the industry had suspected but nobody had seen in black and white. OpenAI spent $34 billion last year to generate $13 billion in revenue. Its operating loss came to $20.92 billion. In the first quarter of 2026, it recorded $5.7 billion in sales against an operating margin of negative 122 percent, implying a quarterly loss approaching $7 billion and an annualized run rate of roughly $28 billion in the red.
The company filed confidentially with the SEC on June 8 for an IPO targeting a valuation above $1 trillion, with Goldman Sachs and Morgan Stanley leading the process. What the leaked financials do is make that aspiration considerably harder to square with conventional valuation math. Skeptics have a clean line now: no company has ever earned a trillion-dollar public market price tag while burning $27 billion a year, with projections suggesting losses climb to $63 billion by 2027 and the company doesn't hit cash-flow positive until 2030.
Here's the thing, though. The losses aren't going to waste. They're going somewhere very specific, and the destination tells you more about where the real AI money is than anything in OpenAI's own S-1 will.
Of OpenAI's $34 billion in 2025 spending, $17.2 billion went directly to Microsoft. That breaks down as $10.59 billion in research and development expenses, $6.047 billion in cost-of-revenue charges, and the remainder in sales and marketing. Microsoft isn't just a partner here, it's essentially OpenAI's landlord, cloud provider, and compute supplier wrapped into one, and the payments flow in a direction that keeps Azure utilization elevated regardless of whether GPT-5 ever turns a profit. OpenAI already represents 45 percent of Microsoft's $625 billion revenue backlog, a figure that gives you some sense of how structurally dependent the two companies have become on each other.
Nvidia's position is, if anything, more straightforward. OpenAI has committed to spending on Nvidia chips across at least 10 gigawatts of AI data centers, and the company recently struck a deal to pay Oracle $60 billion annually for five years starting 2027 for infrastructure built substantially on Nvidia hardware. The first deployments on Nvidia's Vera Rubin platform begin in the second half of 2026. Nvidia also disclosed it has invested up to $100 billion in OpenAI for non-voting shares, so it collects on the upside of the very revenue it generates by selling the chips OpenAI buys. That is a tidy arrangement.
As Yahoo Finance noted in its coverage of the leaked documents, OpenAI's massive losses strengthen the bull case for these two stocks precisely because the losses are, from Nvidia's and Microsoft's perspective, revenue. The distinction sounds obvious once you say it, but the market commentary around OpenAI's financials keeps treating the burn as a signal about the AI sector broadly, when the more useful signal is about where the margin in AI actually accumulates.
The structural ceiling no one wants to name #
What the unit economics in the leak make plain is that frontier model development may not have a path to standalone profitability at scale. OpenAI's cost efficiency did improve: in 2024 it spent $2.37 for every dollar it took in; by 2025 that ratio had narrowed to $1.60. Progress. But the absolute losses widened nearly eightfold year on year because revenue growth, as fast as it has been, is still chasing a cost base that expands faster than it compresses. Training the next frontier model doesn't get cheaper just because you're better at it.
That's not an OpenAI-specific problem. Anthropic is burning at a similar rate with similar dynamics. The companies that can afford to keep playing are the ones with patient capital and a strategic rationale beyond the model itself: a cloud business that benefits from utilization, a chip business that benefits from orders, or a consumer platform that can monetize at the application layer. Pure-play model labs, the ones with no hardware moat and no cloud moat, are racing toward a ceiling that the OpenAI documents now put a number on.
For investors in venture or growth equity, the implication isn't subtle. The picks-and-shovels thesis that's been easy to say for two years now has actual financials behind it. OpenAI spent $34 billion last year and took in $13 billion. The gap is, in the most literal sense, other companies' profits. Nvidia's data center segment has been the single largest beneficiary of the AI capex supercycle, and the OpenAI leak just quantified a meaningful slice of why. Microsoft's cloud backlog, inflated by a company burning $7 billion a quarter, tells the same story from a different angle. None of this means OpenAI fails. Revenue more than tripled from $3.7 billion in 2024 to $13 billion in 2025, and by February 2026 its annualized run rate had reportedly crossed $25 billion. The company is growing fast enough that a path to break-even exists on paper, if the cost curve bends and the revenue curve doesn't. But the structure of the AI economy revealed in this leak is that the model builders are, for now, the biggest and most reliable customers the infrastructure layer has ever had. That's who's winning. They just don't make headlines the same way a leaked loss does.
Also read: Researchers have finally worked out why AI models keep inventing the same fake names • Japan's chip equipment giants are losing China and betting on AI to fill the gap • OpenRouter's Fusion API bets the future of AI belongs to panels, not single frontier models