Audited statements obtained by Ed Zitron and verified by the Financial Times expose the real economics behind the world’s most valuable startup, weeks before its IPO
OpenAI lost $38.53 billion in 2025 on $13.07 billion in revenue, according to audited financials obtained by journalist Ed Zitron and independently verified by the Financial Times. Most of that headline loss is a one-time accounting charge, not cash. The figure that should worry investors is the $20.92 billion operating loss, where research and development alone outran every dollar of revenue.
The numbers landed at the worst possible moment for Sam Altman. OpenAI filed a confidential S-1 with the SEC on June 8, targeting a valuation of up to $1 trillion. The leak strips management of the chance to frame its own story before the prospectus goes public, and the story underneath is messier than the brand suggests.
What Happened
Two separate leaks have now exposed OpenAI’s books, and it is worth keeping them apart because their reliability differs.
The first, in early April, was a reconstructed cap table posted on X by investor Sheel Mohnot and later detailed by Forbes. It was assembled from public filings and secondary-market data, so treat it as a well-sourced estimate rather than gospel.
The second, this week, is the serious one. Zitron obtained what he describes as audited financial statements, and the FT verified them. For 2025, the documents show:
| Line item | 2025 figure |
|---|---|
| Revenue | $13.07 billion |
| Total costs and expenses | ~$34 billion |
| Research and development | $19.18 billion |
| Sales and marketing | $5.73 billion |
| Cost of revenue | $7.5 billion |
| Operating loss | ~$20.92 billion |
| Net loss | $38.53 billion |
Revenue roughly tripled from about $3.7 billion in 2024. R&D more than doubled. Sales and marketing grew nearly fivefold, a signal that customer acquisition is getting more expensive, not less.
The Backstory
The eye-watering $38.53 billion net loss needs an asterisk, and skipping it is the fastest way to misread this story.
Roughly $41.55 billion of charges flowed from OpenAI’s 2025 conversion from nonprofit to for-profit. When early investor rights and warrants converted into equity, the rising value of those instruments created a non-cash fair-value charge. A person familiar with the matter told the FT the increase in OpenAI’s worth alone produced a charge of about $30 billion. No money left the building. It is an accounting event that inflates the loss on paper and grows larger precisely because the company is becoming more valuable.
Strip the non-cash items, the stock-based compensation, and the in-kind Microsoft compute credits, and one source pegged OpenAI’s true out-of-pocket cash burn at roughly $8 billion for the year, against a cash reserve of about $25 billion. By that math the company is not in imminent danger. The danger is structural, and it lives in the operating loss.
The Plan
OpenAI’s pitch to public markets rests on a simple bet: revenue growth outruns the cost of frontier AI before the cash runs out. Monthly revenue reached about $2 billion by late 2025 and held that pace into 2026, with roughly 900 million weekly users. The company has confidentially filed for its IPO, with Goldman Sachs and Morgan Stanley leading, and a listing possible as early as September.
There is genuine progress buried in the losses. In 2024, OpenAI spent $2.37 to generate each dollar of revenue. In 2025 that ratio fell to $1.60. The slope is moving the right way. The question is whether it moves far enough, fast enough, before 2030, the year most reporting suggests OpenAI finally turns cash-positive.
The Business Model Angle
Here is the contrarian read, and it cuts against both the bulls and the bears.
The bears scream “$38 billion loss” and miss that most of it is accounting theater. The bulls point to tripling revenue and miss the part that actually matters: OpenAI does not have software economics. Its gross margin sat around 33% in 2025, down from roughly 40% a year earlier. Margins compressed while revenue nearly tripled. That is the opposite of how great software companies behave, where each new customer costs almost nothing to serve and margins widen with scale.
The reason is compute. R&D of $19.18 billion exceeded total revenue, and OpenAI paid Microsoft north of $10 billion for R&D compute alone. Every query costs real money to answer, and the cost does not vanish as the user base grows. This is why the OpenAI business model is so much harder to value than a classic SaaS franchise. It is closer to a capital-intensive infrastructure business wearing a software multiple.
Then there is the moat problem. OpenAI is reportedly weighing steep token price cuts to fend off Anthropic, a move that would shred already-thin margins. When two products are near-interchangeable, price becomes the battlefield, and there is no winning a price war when both sides lose money on compute. For context on the field, our breakdown of OpenAI’s competitors lays out how crowded this race has become.
The Cap Table Angle
The reconstructed cap table tells its own story, and the most striking line belongs to the man running the company.
Sam Altman is listed at 0% ownership, marked “None/Pending.” His equity grant is tied to the public benefit corporation conversion, and he is reportedly negotiating for around 7%, which at current marks would be worth roughly $59 billion. Until that closes, the CEO of an $852 billion company holds no shares.
The rest of the estimated table, at the $852 billion post-money valuation from March’s $122 billion round:
| Holder | Stake | Notes |
|---|---|---|
| Strategic corporates (combined) | 46.58% | Microsoft, SoftBank, Amazon, NVIDIA |
| Microsoft | ~26.79% | ~$13B invested, valued near $228B, about 17.6x |
| OpenAI Foundation (nonprofit) | 25.80% | Zero cost basis, controls 100% of board appointments |
| Current and former employees | ~19.35% | |
| SoftBank | ~11.66% | Committed $64.6B, up an estimated $50B on paper |
| VC and institutional | 7.83% | |
| Amazon | 4.66% | Committed up to $50B in March 2026 |
| NVIDIA | 3.47% | $30B in the 2026 round |
| Retail and secondary | 0.40% |
Two facts deserve attention. First, four corporate balance sheets control nearly half of a company founded to keep AI out of concentrated hands. This was not built by Sand Hill Road. The entire venture block owns less than SoftBank alone. Second, the nonprofit foundation’s 25.8% is the largest single overhang on the table. It cannot be sold or leveraged, it controls every board appointment, and an IPO that puts a nonprofit holding company atop a public benefit corporation is genuinely unprecedented territory for the IRS and state regulators.
The Risk
The path to profitability depends on variables OpenAI does not fully control. Inference costs have to keep falling per query, and if hardware efficiency gains slow while demand climbs, the 2030 breakeven slips and the cumulative burn grows.
Competition is the sharper threat. Anthropic has reportedly posted a small operating profit, surpassed OpenAI in private valuation, and is forcing the price pressure that threatens OpenAI’s margins. The Microsoft relationship is also a flagged risk in OpenAI’s own investor materials, given revenue-sharing terms and the scale of compute OpenAI buys back from a company that also owns a quarter of it.
And the in-kind compute story hides a slow-moving problem. Much of OpenAI’s compute has come as prepaid Azure credits rather than cash out the door. When those run down, more of that cost starts hitting the books as real cash, and the comfortable $8 billion burn figure gets less comfortable fast.
Quick Questions
How much did OpenAI actually lose in 2025? The audited net loss was $38.53 billion, but roughly $41.55 billion of charges were non-cash, tied to its nonprofit-to-for-profit conversion. The operating loss of about $20.92 billion is the more meaningful figure, and true cash burn was estimated near $8 billion.
Why does Sam Altman own zero equity? His grant is pending the public benefit corporation conversion. He is reportedly negotiating for about 7%, a stake that could be worth around $59 billion, though nothing has been finalized.
Are the cap table and the financials the same leak? No. The cap table surfaced in April as a reconstructed estimate from public and secondary data. The financials surfaced this week as audited statements verified by the Financial Times. Different sources, different reliability.
When does OpenAI expect to make money? Most reporting points to around 2030 for positive cash flow, contingent on falling compute costs and continued revenue growth.
The Business Model Analyst Take
The most useful thing a founder can take from this leak is not the loss figure. It is the reminder that accounting losses and economic losses are different animals, and that confusing them is how you get fooled in both directions.
OpenAI’s real problem is not the $38 billion headline. It is that the business does not yet behave like software. Margins fell as revenue tripled, R&D outran the entire top line, and the cost of serving each user refuses to disappear with scale. That is a capital-intensive infrastructure business, and infrastructure businesses do not earn 35x revenue multiples on faith. The cap table makes the same point from another angle: this company was funded by corporate balance sheets writing ten-figure checks, not by venture math, because venture math does not work at this burn rate.
The bet that makes a $1 trillion IPO rational is that compute gets dramatically cheaper before the capital gets nervous. That bet might pay off. But it is a bet on the cost curve, not on the product, and any operator pricing OpenAI should be honest about which one they are actually buying.
Sources: Financial Times, Fortune, and reporting by Ed Zitron on the audited 2025 financials; the reconstructed cap table posted by Sheel Mohnot and detailed by Forbes; Sacra and CNBC for valuation and margin context.
