OpenAI lost $38.53 billion in 2025. The figure is not an analyst estimate or a leaked deck slide — it is the net loss attributable to the company, taken from audited 2025 financial documents viewed by Ed Zitron’s Where’s Your Ed At and independently verified by the Financial Times. The numbers land nine days after OpenAI’s confidential S-1 filing on June 8 — and they make the IPO math significantly harder than the filing will.
🔍 THE BOTTOM LINE
In 2024, OpenAI lost $5.09B on $3.7B in revenue. In 2025, it lost $38.53B on $13.07B in revenue. Revenue grew 3.5x; the net loss attributable to the company grew 7.6x. The unadjusted 2025 net loss — before the $17.87B reclassified as “net loss attributable to noncontrolling members capital” and the $3.95B as “net loss attributable to redeemable noncontrolling interests” — was $60.35B, swollen by a $41.55B one-time charge when the company converted from non-profit to for-profit. The audited numbers turn a story the IPO roadshow was going to tell (“we are scaling fast”) into a story the roadshow will have to answer (“we are scaling fast and burning faster”).
The Numbers, Plain
The 2025 figures, per Zitron’s leaked documents:
- Revenue: $13.07 billion (3.5x the $3.7B in 2024)
- Cost of revenue: $7.5 billion
- Research and development: $19.18 billion
- Sales and marketing: $5.73 billion
- General and administrative: $1.57 billion
- Total costs and expenses: $34 billion
- Loss from operations: $20.92 billion
- Net loss attributable to the company: $38.53 billion
- Total unadjusted net loss (including the $41.55B non-profit conversion charge): $60.35 billion
- Year-end assets: ~$50 billion, almost half in cash
The most striking single line is the gap between revenue growth and loss growth. A 3.5x revenue jump would normally be a win. A 7.6x loss jump makes the revenue jump a vanity metric. The unadjusted 2025 loss is, on its own, larger than the GDP of most countries.
Compute Is the Real Story
The single largest number inside the 2025 cost base is the $17.2 billion OpenAI paid Microsoft across R&D, cost of revenue, sales and marketing, and G&A line items. The biggest sub-line — $10.59 billion for “research and development” expenses paid to Microsoft — almost certainly represents the cost of training OpenAI’s models on Microsoft’s Azure infrastructure. OpenAI ended 2025 with $3.64 billion in liabilities to Microsoft on the books, plus another $79M in accrued and non-current liabilities.
This is a structural dependency, not a line item. OpenAI’s ability to ship the next-generation model is, in dollar terms, tied to Microsoft’s willingness to keep providing compute. CFO Sarah Friar’s April warning to staff about roughly $600B in compute commitments already signalled the trajectory; the audited numbers show that 2025 spend on Microsoft alone is now 2.9% of that total — and the runway extends for years.
The IPO Context
OpenAI’s June 8 confidential S-1 came one week after Anthropic’s June 1 filing, kicking off the largest IPO race in AI history. The audited financials are exactly the kind of disclosure a public-company prospectus will eventually force into the open — and the SEC’s S-1 requirements for audited financials mean these numbers, or close variants of them, will appear in the eventual prospectus.
What the S-1 will not be able to do is explain them away. Anthropic’s Stratechery-analyzed “safety superpower” positioning — the public-company marketing story that justifies enterprise pricing — is doing real work for Anthropic. OpenAI’s equivalent marketing story has historically been “AGI is coming and we’ll be first.” With $38.53B in losses on the ledger, the AGI story is the only thing keeping the valuation from collapsing on the first earnings call.
Capital Inflow — Still Robust, For Now
OpenAI ended 2025 with $50B in assets, almost half in cash. That liquidity is being maintained by capital injections from a small number of counterparties. In 2025, SoftBank paid OpenAI $867 million and Microsoft paid OpenAI $303 million — the latter likely representing revenue-share or service credits flowing the other direction, but on the books it counts as cash in. The total cash position reflects the cumulative effect of the March 2026 $122B funding round, not 2025 operating cash flow. 2025 operating cash flow was deeply negative.
The risk: if any one of the major counterparties (Microsoft, SoftBank, the lead institutional investors in the March round) reduces exposure, the cash runway shortens dramatically. With $34B in 2025 costs and an unadjusted $60B loss, OpenAI needs capital inflows at roughly $50B+ per year to maintain runway. That math is the reason the IPO timing is not optional.
The NZ Angle: A Sovereign Question
New Zealand does not appear directly in OpenAI’s audited 2025 financials. The country is a small customer, not a capital provider or a strategic counterparty. But the math still matters here. The KiwiSaver default providers hold equity in the listed AI-adjacent names (Microsoft, NVIDIA, Oracle, Amazon) that are also exposed to OpenAI’s burn rate through their contracts and equity stakes. A write-down of OpenAI’s valuation in the IPO or in subsequent earnings would propagate, in slow motion, into the default KiwiSaver funds.
The more immediate NZ question is procurement. Government agencies, universities, and a small number of enterprise customers are using OpenAI’s APIs or ChatGPT Team/Enterprise under multi-year contracts. If OpenAI has to repricing aggressively post-IPO to defend revenue growth (a pattern we already saw in Anthropic’s enterprise ramp), the cost basis for NZ public-sector AI contracts shifts. EU-style federated compute efforts like EUROMESH — which explicitly reduce dependency on a single hyperscaler — start to look less like academic exercises and more like procurement hedge.
❓ FAQ
Q: Is the $38.53B loss number real? A: It is the net loss attributable to the company, taken from audited 2025 financial documents. Ed Zitron’s Where’s Your Ed At viewed the documents; the Financial Times independently verified them. The unadjusted loss (before the $17.87B reclassification to noncontrolling interests and the $3.95B to redeemable noncontrolling interests) was $60.35B.
Q: Why are 2024 and 2025 so different — a 7.6x jump in losses? A: Three factors: (1) the R&D line tripled to $19.18B, mostly compute; (2) S&M jumped to $5.73B as OpenAI defended enterprise share against Anthropic; (3) the for-profit conversion triggered a $41.55B fair-value adjustment on convertible interests and warrant liability. The first two are ongoing; the third is one-time.
Q: Does this change the IPO valuation? A: The S-1 confidential filing already exists, and the SEC requires audited financials in the eventual prospectus. The numbers will appear. The question is whether the IPO can clear at a higher valuation than the March 2026 private round ($852B post-money) given the burn rate. Ed Zitron’s read is that the financials are “deeply concerning” and the losses are “far higher than most believed.”
Q: What is the role of the $17.2B paid to Microsoft? A: It is the single largest counterparty line in the cost base. The $10.59B in R&D payments to Microsoft almost certainly represents the cost of training OpenAI’s models on Azure. OpenAI’s ability to ship its next-generation model is, in dollar terms, tied to Microsoft’s compute.
Q: What does this mean for the broader AI bubble? A: It is the first audited, not-modelled, not-projected, primary-source data point on what one of the two frontier labs actually costs to run. Anthropic’s audited numbers (when they appear in the S-1) will be the second. The gap between these numbers and the revenue growth story in the AI industry’s marketing is now an open question, not a debate.
🔍 THE BOTTOM LINE
The leaked audited financials are the first time the public has seen a primary-source look at what the AI boom actually costs to operate. The numbers do not contradict OpenAI’s revenue story — $13.07B is real, growing, and significant. They do contradict the margin story. A 7.6x jump in net loss alongside a 3.5x jump in revenue is not a company on a path to profitability; it is a company on a path to a larger IPO. The market will decide in 2026 whether the S-1 can carry a valuation that prices in 2030 returns. The audited 2025 numbers are the case the bears will keep pointing to until OpenAI’s prospectus either confirms them or — for the first time — contradicts them.
