Audited financial documents reviewed by Ed Zitron, and independently verified by the Financial Times, show OpenAI lost about $38.5 billion in 2025. That’s on $13.07 billion in revenue against $34 billion in total costs. That’s roughly 7.5 times its $5.09 billion loss in 2024. The headline number is real, but it’s inflated by a one-time $41.55 billion non-cash accounting charge from OpenAI’s nonprofit-to-for-profit conversion. Strip that out and the operating loss was $20.92 billion. Strip out other non-cash items and the FT puts the “true” cash loss closer to $8 billion.
So there are three numbers: $38.5 billion (the reported net loss), $20.92 billion (the operating loss), and roughly $8 billion (the adjusted cash burn). The bulls will quote the smallest. The bears will quote the largest. The honest read is that all three matter. The company is filing to go public at a valuation north of $1 trillion. And it told investors it plans to commit roughly $600 billion to infrastructure through 2030. Revenue tripled. Costs grew faster.
Best for anyone weighing the OpenAI IPO, AI-industry watchers, and people tired of both the doom and the hype. Not ideal for readers who want a single clean number, because the whole point is that there isn’t one.
Today the internet learned that OpenAI lost $38.5 billion last year.
By tonight, half of it will be screaming that number, and the other half will be explaining why it doesn’t count.
They’re both right, and that’s the problem.
What the Documents Actually Show
Audited financial documents reviewed by Ed Zitron and independently verified by the Financial Times lay out OpenAI’s 2025 in hard numbers. Revenue was $13.07 billion, up from $3.7 billion the year before. That’s strong growth by any measure. The problem is the other column.
Total costs and expenses hit $34 billion. Research and development alone was $19.18 billion. Sales and marketing was $5.73 billion. Cost of revenue, the actual expense of running the models people use, was $7.5 billion. Stack revenue against costs and you get a loss from operations of $20.92 billion.
Then the accounting gets strange. 2025 was the year OpenAI converted from a nonprofit to a for-profit entity. That triggered a $41.55 billion loss tied to changes in the fair value of convertible interests and warrant liabilities. That charge pushed the total net loss all the way to $60.35 billion. OpenAI then subtracted $17.87 billion, plus another $3.95 billion, through line items labeled net loss attributable to noncontrolling interests. That lands at the figure everyone is quoting today: a net loss attributable to OpenAI of $38.53 billion.
For comparison, the 2024 number was $5.09 billion. So the reported loss grew nearly eightfold in a single year.
The Three Numbers, and Why Each One Lies a Little
Here’s where most coverage will pick a side and run. The honest move is to hold all three numbers at once.
The $38.5 Billion Number
This is the reported net loss, and it’s the one in every headline. It’s technically accurate. It’s also inflated by the $41.55 billion conversion charge. That charge is non-cash and, per the FT’s sources, not expected to recur after the restructuring. Quoting $38.5 billion as if OpenAI set fire to that much actual money is misleading. The company didn’t write a check for $38.5 billion. A big chunk of that is an accounting event tied to how its investors’ rights got revalued as the company’s worth ballooned.
The $8 Billion Number
This is the one OpenAI’s defenders will quote. Strip out the conversion charge, the stock comp, the Microsoft computing credits, and the other non-cash items. The FT reports the underlying loss was about $8 billion. This is the most flattering framing, and it has a real point. $8 billion is the closest thing to “how much cash did the core business actually burn.” But it’s still an enormous operating hole. Presenting it as the comforting number requires waving away a lot.
The $20.92 Billion Number
This is the one almost nobody will quote, and it might be the most honest. It’s the loss from operations, before the accounting theater and before the generous adjustments. It’s what the actual business did: $13 billion in, $34 billion out. It strips the one-time conversion charge that unfairly bloats the $38.5 billion figure. But it doesn’t let the company off the hook with credits and stock comp the way the $8 billion figure does. If you want one number for how the business performed in 2025, $20.92 billion in operating losses is the fairest one.
So when someone tells you OpenAI lost X last year, the right response is which X.
Why the Gap Matters Right Now
This isn’t an academic accounting debate. The timing is the whole point.
OpenAI filed confidentially for an IPO this month, one that could value the company north of $1 trillion. It raised $122 billion earlier this year at a $730 billion valuation. SpaceX is listing this week, Anthropic is next, and OpenAI is behind them. All three are fighting for the same pool of institutional money in the same quarter. Into that environment drops a leaked set of audited financials showing the reported loss grew eightfold.
The number you choose to believe shapes how the IPO gets priced. If the market anchors on $38.5 billion, OpenAI looks like a company hemorrhaging money at a catastrophic rate. If it anchors on $8 billion, OpenAI looks like a high-growth business with a manageable burn and a clear path to scale. The truth sits in the middle, at the $20.92 billion operating loss. And the company has every incentive to push the conversation toward the smallest figure before the roadshow.
That’s why the leak matters more than the number. Someone handed these documents to a journalist known for being the AI industry’s loudest skeptic, weeks before the most-anticipated tech IPO in years. The financials would have come out eventually in the S-1. They came out now, framed by a bear, on someone else’s schedule.
The Number That Should Actually Worry You
Forget which loss figure wins the argument. The structural problem is in the trend, not the total.
Revenue tripled from $3.7 billion to $13.07 billion. That’s the kind of growth that justifies a huge valuation. But costs grew faster, and they grew in the worst place. R&D more than doubled to $19.18 billion. Cost of revenue, the price of simply serving the models, nearly tripled to $7.5 billion. The thing about frontier AI is that each more capable model costs more to train and more to run. So scaling the product scales the bill underneath it. Growth doesn’t relieve the pressure. It adds to it.
And OpenAI has told investors it expects to commit roughly $600 billion to AI infrastructure through 2030. The $34 billion it spent last year is an early installment on that. A company losing somewhere between $8 billion and $21 billion a year, while promising to spend $600 billion over five years, is making one bet. That revenue will eventually outrun a cost base that is, so far, outrunning it.
This is the same tension every frontier lab is living. We covered it when Anthropic raised $65 billion at a $965 billion valuation and when the SpaceX S-1 revealed Anthropic is paying $1.25 billion a month just for compute. The whole sector is running the same play: raise enormous amounts, spend even more, and bet that scale eventually flips the math. Nobody has proven it flips. They’re all just early enough that the bet hasn’t been called yet.
What To Watch
The S-1 is coming. When OpenAI files its actual public offering document, these numbers get audited, standardized, and stripped of the framing both sides are applying today. That’s the version that matters, and it’s the one institutional investors will price against. Watch whether the official operating loss lines up with the $20.92 billion in this leak. Or whether the public filing finds a way to present something gentler.
Watch the order of the IPOs too. SpaceX prices first this week and sets the comparable. If it lands strong, OpenAI and Anthropic ride the benchmark. If it stumbles, the whole AI IPO wave meets a more skeptical market. A leaked $38.5 billion loss number becomes a much heavier weight to carry into a roadshow.
And watch who keeps leaking. These documents came out early, framed by a critic, on a schedule that wasn’t OpenAI’s. In the weeks before a trillion-dollar listing, information is a weapon, and someone just fired the first shot.
The honest bottom line is the boring one. OpenAI did not literally lose $38.5 billion in cash, and it did not comfortably lose only $8 billion either. It lost about $21 billion running the actual business. The cost curve is steeper than the revenue curve. And it’s about to ask the public markets for a trillion-dollar valuation. The number is bad. The spin in both directions is worse. And the only figure that really counts hasn’t been filed yet.
