OpenAI's Leaked Audited Financials: Revenue Up 3.5x in a Year, Losses Locked In by R&D and Compute

Summary

Independent journalist Ed Zitron obtained OpenAI’s audited financial documents, and the FT reviewed the same set. The numbers themselves are not shocking: revenue rose from $3.7B in 2024 to $13.07B in 2025, up 3.5x in a year, with monthly revenue near $2B by the end of 2025. But over the same period, operating losses widened from $8.78B to $20.92B, growing faster than revenue did.

Put revenue and cost on one table and the real signal appears. In 2025, R&D alone cost $19.18B, more than the entire $13.07B of annual revenue. OpenAI’s problem is not that revenue grows too slowly. It is that the company has pre-committed all the room that revenue growth could create to R&D and compute. So the question moves away from when OpenAI turns a profit toward two sharper ones: who keeps filling a roughly $20B annual operating gap before an IPO, and whether the pace of the loss narrowing from 237% to 160% of revenue is fast enough to convince public markets when it actually lists.

What happened

These are audited documents, a notch more credible than the various estimates that circulated before. From the Ars Technica and FT reporting, the main lines are clear.

On revenue: $3.7B in 2024, $13.07B in 2025. The FT adds that monthly revenue had reached nearly $2B by the end of 2025, meaning revenue accelerated through the year and the year-end run rate ran ahead of the annual average.

The cost side is the heart of the story. R&D grew from $7.81B in 2024 to $19.18B in 2025, of which about $10.59B went to Microsoft, mostly for training new models. Cost of revenue, the money spent producing and distributing the product itself (this line mainly reflects the inference compute used as models respond to a growing flood of user prompts), rose from $2.65B to $7.5B. Sales and marketing climbed from $1.11B to $5.73B. Together, those three lines swallow the fast revenue growth whole.

All told, the day-to-day operating loss widened from $8.78B in 2024 to $20.92B in 2025. One number needs separate handling: OpenAI’s 2025 net loss was nearly $39B, far above the operating loss, but Ars and the FT both note it includes a one-time accounting charge tied to the 2025 for-profit conversion and shifting investor valuations. That is not ongoing operations. When reading the statements, keep the operating loss (how much the business itself loses) apart from the net loss (which carries one-time items), or you will overstate the operational bleed rate.

A bright spot easy to miss sits in the ratio: operating loss as a share of revenue fell from 237% in 2024 to 160% in 2025. The absolute loss is growing while the burden relative to revenue is easing.

Context matters too. OpenAI is filing with the SEC to prepare for an IPO and tells investors it aims to be profitable by 2030. In March it closed a $122B funding round at an $852B valuation. ChatGPT has over 900M weekly active users, but only about 50M are paying subscribers. This year it also shut down its Sora video model, and applications chief Fidji Simo told staff to cut back on side quests and focus on core coding and business users.

Why it matters

Start with the question that deserves to be asked first: is this loss structural, or will it shrink on its own with scale? The answer is some of both, leaning structural.

It is structural because the two largest cost lines ($19.18B R&D, $7.5B inference compute) are not ones OpenAI can quickly cut on its own. About $10.59B of R&D is a compute commitment to Microsoft, and inference cost rises with user request volume. As long as the strategy stays anchored on training bigger models and serving more users, both lines climb alongside revenue. That is what a locked-in loss means here: every extra dollar of revenue arrives with a large chunk already spoken for by pre-committed compute and R&D.

But the loss is not fully immovable either. The 160% ratio is a clear improvement over 237%, proof that revenue growth is outrunning loss growth for now. The catch is that the narrowing depends on revenue more than tripling in a year, which is not a normal rate. Reaching profit requires continued high revenue growth and a clear slowdown in R&D growth at the same time. Ars points the same way: to turn losses into profit, OpenAI will eventually have to rein in costs, especially the rising R&D tied to model training.

In the IPO context, the importance gets concrete. For a company planning to list, public markets do not fixate on one quarter’s absolute loss. They look at the slope of margin improvement and the credibility of the path to profit. The move from 237% to 160% is a real slope that can be turned into a story. But going from 160% to break-even still means closing a gap worth 1.6x revenue. Until then, who fills the roughly $20B annual operating gap each year is the core variable for survival, and filling it depends entirely on continued access to financing. The $122B raised in March is exactly that gap-filler.

Builder impact

If your product is built on OpenAI, these numbers give you two actionable reads that point in opposite directions but both hold.

First, on price. Current API prices most likely contain a financing subsidy. The audited figures show inference compute (cost of revenue) rising from $2.65B to $7.5B in a year, while Ars notes enterprise customers are starting to balk at token-based pricing and demand a measurable return. Put those together: OpenAI is paying more and more for inference while facing customer pushback on price. It can sell below a long-term sustainable level to buy growth only because financing backs it. So builders should assume long-run upward pressure on API prices, and at minimum should not build a business model on prices falling forever. Concretely, stress-test your unit economics as if prices revert toward cost, and check whether margins still hold if token prices rise 30% to 50%.

Second, on dependence. Two misreads are worth killing here. Reading “OpenAI is losing a fortune” as “OpenAI is about to collapse” is wrong: it holds massive financing, revenue is growing fast, and the loss ratio is improving, with no sign of a near-term crash. But reading it as “the losses do not matter to me” is also wrong: OpenAI’s survival is tied to continued access to financing, and that access depends on capital-market sentiment toward AI and on the credibility of its own profit story, neither of which you control.

In architecture terms, that means one concrete move: treat dependence on a single frontier supplier as a risk to hedge, not a constant to ignore. This is not a call to drop OpenAI now. It is a call to avoid single-betting an irreplaceable core path on it. Keep an abstraction layer so the same requirements can move to another frontier model or open weights within a reasonable window, and for price-sensitive, high-volume paths, assess cost exposure if token prices rise. The Sora shutdown and Simo’s push to cut side quests are an extension of the same signal: under cost pressure, OpenAI will trim non-core product lines, so if you depend on one of its edge features, plan for it as a candidate to be cut.

What to ignore

The first read to throw out is treating the near-$39B net loss as a doomsday signal to pass around. That number includes a one-time accounting charge tied to the for-profit conversion and valuation changes, and does not reflect ongoing operations. To measure operational bleed, use the operating loss ($20.92B). Computing “how much OpenAI burns per day” from the net loss systematically overstates it and is a textbook misreading of the statements.

The second thing to ignore is lifting the absolute loss out of context as panic fuel. $20.92B is genuinely large, but it is meaningless apart from revenue. In the same documents, the trend of loss narrowing from 237% to 160% of revenue matters far more for judging whether OpenAI can reach an IPO than the absolute figure does. Shouting “out of control” at a $20B loss while omitting that the ratio is improving is reading only half the statement.

Finally, do not get pulled into guessing whether OpenAI hits profitability exactly in 2030. The audited numbers give you the facts that matter: the loss is largely structural, the narrowing rests on abnormal growth, and survival is tied to financing. Which specific year break-even lands depends on too many unpredictable variables (capital-market sentiment, the competitive field, compute prices). The actionable move for builders is not betting on a year, but hedging price pressure and supplier dependence ahead of time no matter which year it is.

Sources

1. Leaked financial docs show OpenAI is losing billions of dollars a year (Ars Technica) / news
2. Exclusive: OpenAI's audited financials (Ed Zitron) / blog

No official primary source available; this analysis is based on reliable secondary reporting (named outlets, cross-confirmed).