OpenAI's audited 2025: the Microsoft line is the one to read

OpenAI's 2025 audited financial statements leaked yesterday, days after the company filed its confidential S-1. Revenue was $13.07bn. Operating loss was $20.92bn. Net loss was $38.53bn. The company is pursuing a $1trn listing on these numbers. The line in the statements that actually matters, though, is a different one: OpenAI paid Microsoft $17.2bn in 2025, and Microsoft paid OpenAI $303m back.

The documents were obtained by Ed Zitron and verified independently by the Financial Times.¹ I spent the morning with the secondary reporting and the comparable Anthropic disclosures, because the primary statements have not been published in full, and the structural story is legible from what has been disclosed.

The headline net loss is mostly an accounting artefact. The $38.53bn figure is inflated by a non-cash charge tied to OpenAI's 2025 conversion from a capped-profit structure to a full for-profit entity.² Strip that out and the adjusted net loss is roughly $8bn on $13bn of revenue. Still large. Still a company that loses money on a vast scale. But $8bn and $38.5bn tell quite different IPO stories, and the framing of risk-factor disclosures will turn on which one anchors the road show.

The operating loss is the real number, and it has trajectory. Revenue grew from $3.7bn to $13.07bn, roughly 253% year-on-year. Operating loss grew from $8.78bn to $20.92bn. So losses widened in absolute terms but compressed as a share of revenue, from about 237% to about 160%.² That ratio compression is the single data point in these statements that supports a path-to-margin narrative. Everything else in the disclosure cuts the other way.

The Microsoft line

This is the disclosure that ought to dominate the S-1 risk factors. OpenAI's total cost base for 2025 was approximately $34bn. Of that, $17.2bn, just over half, went to a single counterparty.¹ That counterparty is Microsoft, which also holds equity in OpenAI, holds IP licensing rights to OpenAI's model weights, and distributes OpenAI's models through Azure as a competing first-party offering.

The reverse flow was $303m. So net, $16.9bn left OpenAI for Microsoft in one fiscal year. This is not a vendor relationship in any normal sense. It is the structural fact of the business.

A public-markets investor reading the S-1 will want three things from the Microsoft section. The contractual structure of the Azure compute commitment — minimum spend, term, renegotiation triggers. The IP carve-outs, particularly around model weights and what happens to Microsoft's rights if OpenAI achieves AGI as internally defined. And the cost-share architecture: how much of the $17.2bn is genuinely usage-based inference spend versus committed capacity OpenAI is paying for whether it uses it or not.

None of this has been disclosed in the leaked materials. All of it is going to be in the S-1, and all of it will be picked apart.

Inference economics, tested

The cleanest read of the cost-of-revenue line is unfavourable. Cost of revenue was $7.5bn in 2025, roughly triple the prior year.² Revenue grew 253% while direct cost grew about 200%. That is some operating leverage, gross-cost ratio improved, but not the order-of-magnitude leverage that the inference-economics frame predicts as a strong version: cost per query falling materially faster than revenue per query.

What the disclosure looks like is a business where compute cost scales roughly with usage and the unit economics are slowly improving but not transforming. R&D was $19.18bn, larger than revenue. The implicit per-query cost line is the one the S-1 risk factors will be re-read on, and it does not yet show the inflection that would justify pricing for margin expansion.

Anthropic's awkward comparison

The narrative problem for OpenAI's IPO pricing is what Anthropic is doing on the next desk. Anthropic has separately reported a $47bn ARR (annual recurring revenue — the run-rate of subscription revenue) run-rate and a $559m operating profit in Q2 2026.³ On its face, the smaller lab is the disciplined one and the larger lab is the one burning capital at scale.

The comparison is not quite apples-to-apples — ARR run-rate is not audited annual revenue, Anthropic's cost base is structured differently, and the period ends do not match. But the public-markets question is whether OpenAI can present a credible reason that its losses are investment and Anthropic's profit is constraint, rather than the other way round. That is a harder pitch in mid-2026 than it was in 2024.

The $1trn question

OpenAI is targeting a listing at roughly 77x trailing revenue. The valuation does not price earnings; it prices positioning, optionality, and the thesis that AI capex commitments are self-fulfilling because nobody else can plausibly match the spend. This is the AI performativity case, and the leaked numbers are consistent with management running the company that way: R&D above revenue, compute commitments dwarfing every other cost line, losses absorbed as the price of holding the frontier.

The public-markets test for that thesis is different from the private-rounds test. Private capital priced positioning. Public capital prices disclosure, sensitivities, and the question of what happens when one counterparty controls half the cost base. The S-1 will have to argue both that the Microsoft relationship is durable and that it is renegotiable — durable enough to underwrite scale, flexible enough not to be a structural margin cap. Those two things are in tension.

The audited numbers tell us that ratio compression is real, that adjusted losses are not catastrophic for a company growing 253%, and that the inference-cost line is moving in the right direction but slowly. They also tell us that OpenAI's largest commercial relationship is with a company that is simultaneously its shareholder, its IP licensee, its distributor, and its competitor. The IPO turns on how that sentence reads in a risk-factor section.

Footnotes

¶