OpenAI Lost $38.5B in 2025. Your Data Pays the Gap.

The Leak: Audited Documents, Two Independent Eyeballs, Three Tier-1 Outlets

Ryan Merket reported at RuntimeWire on June 16, 2026 at 4:36 pm CT that OpenAI lost $38.53 billion attributable to the company in 2025, according to audited financial documents viewed by Ed Zitron and independently verified by the Financial Times.^[1] The documents themselves are not public. The figures reported from them are specific enough to cut through the abstraction that has defined the AI financing cycle: $13.07 billion in 2025 revenue, $34 billion in total costs and expenses, $20.92 billion loss from operations, and a $38.53 billion net loss attributable to OpenAI after a $60.35 billion group net loss was reduced by losses attributed to noncontrolling interests.^[1]

Merket's piece was the primary report. Two tier-1 outlets followed the same day. Ars Technica (Benj Edwards, June 16, 2026) ran "Leaked financial docs show OpenAI is losing billions of dollars a year," confirming the figures with the same attributed-to-OpenAI framing and the same $20.92 billion operating-loss anchor.^[2] Fortune (June 16, 2026, 23:54 UTC) ran "OpenAI financials leaked: losses, revenue, profit," with the tier-1 editorial pickup that lifted the story from a niche disclosure to a structural event.^[3] The Hacker News discussion thread for the RuntimeWire piece sat at 197 points and 207 comments as of the 07:35 UTC scan on June 17, 2026, up from 7 points at the 02:09 UTC morning scan: a 28x compound in roughly five and a half hours, the biggest engagement compound on a state-aligned-power-network story of the day, and the third-biggest engagement compound on any story in the 06:00 UTC to 08:00 UTC window.^[9]

The provenance is what makes the disclosure structural. Ed Zitron of "Where's Your Ed At" is the journalist who has been most consistently on the OpenAI financial-record story through 2025 and 2026, and Zitron's name on the viewed-documents line is the editorial endorsement the figures needed to graduate from "leaked documents" framing to "audited disclosed financials." The Financial Times independent verification is the second endorsement. Tier-1 outlet pickup is the third.^[1][2][3] The graduation is what makes the gap visible to readers who do not follow the AI financing beat. The graduation is also what makes the gap an accountability event for OpenAI's first-party disclosures.

The Numbers: A Revenue Ramp That Did Not Catch the Cost Base

The 2025 figures, in the audited documents viewed by Zitron and verified by the Financial Times:^[1]

• Revenue: $13.07 billion in 2025, up from $3.7 billion in 2024. Roughly 3.5x growth.^[1]
• Total costs and expenses: $34 billion in 2025, up from $12.48 billion in 2024. Roughly 2.7x growth.^[1]
• Operating loss: $20.92 billion in 2025, up from $8.78 billion in 2024. Roughly 2.4x growth.^[1]
• Net loss attributable to OpenAI: $38.53 billion. Includes the $41.55 billion loss related to changes in the fair value of convertible interests and warrant liability, offset by the noncontrolling interest share of the larger $60.35 billion group net loss.^[1]
• Cost-of-revenue line: $7.5 billion. The line that pays for the inference serving capacity, the data-center leases, and the GPU depreciation.^[1]
• Research and development: $19.18 billion. 56% of the cost base. Compute is the line item.^[1]
• Sales and marketing: $5.73 billion.^[1]
• General and administrative: $1.57 billion.^[1]
• Microsoft paid: $17.2 billion ($10.59B booked as R&D, $6.047B cost of revenue, $527M sales and marketing, $42M G&A).^[1]
• Microsoft liabilities at year-end: $3.64 billion.^[1]

The 3.5x revenue growth versus 2.4x operating-loss expansion is the "scale argument" Altman has been selling since the company's founding. The scale argument is that compute, distribution, capital, and consumer habit compound faster than the cost base. The audited 2025 numbers say the opposite for one more year. The cost base expanded faster than revenue would have if the company had been profitable. The expansion is what the company is asking capital markets to fund through the IPO window.^[1]

The $41.55 billion fair-value adjustment on convertible interests and warrant liability is the accounting item that needs its own paragraph. The figure represents the change in value of the convertible instruments and warrants tied to OpenAI's restructuring and recapitalization. The figure does not represent cash burn. The figure represents the mark-to-market on the paper instruments OpenAI issued to raise the $122 billion in committed capital at the $852 billion post-money valuation OpenAI announced in March 2026. The figure is large because OpenAI's valuation increased during 2025 and the warrants became worth more on paper. The figure is real in the sense that it is a real loss for the equity holders; the figure is not real in the sense that no cash left the building. The $20.92 billion operating loss is the cash-flow number. The $38.53 billion attributable net loss is the GAAP number. The two are different framings of the same year.^[1]

The Microsoft Line: The Single Largest Infrastructure Dependency

The $17.2 billion OpenAI paid Microsoft in 2025 is the line item that makes the audited financials matter for the surveillance beat. Microsoft is simultaneously an equity holder (roughly 27% of OpenAI Group as of the closing of OpenAI's 2025 recapitalization), a commercial partner (the Azure OpenAI Service relationship), the cloud supplier (the compute behind ChatGPT, the API, and the enterprise deployments), and the infrastructure dependency (the GPUs, the datacenters, the energy contracts).^[1][4] Microsoft is not just another investor on OpenAI's cap table. Microsoft is the vendor relationship that consumes tens of billions of dollars of OpenAI's annual cash burn, and Microsoft is the vendor relationship that, on the equity side, gets the upside when the IPO lands.^[1][4]

The $3.64 billion in liabilities to Microsoft at year-end is the second Microsoft line. The liability figure is what OpenAI owes Microsoft for compute delivered but not yet paid for at the December 31, 2025 balance-sheet date. The liability figure is the working-capital snapshot of the vendor relationship. The liability figure is also the structural advantage Microsoft holds over OpenAI's pricing for the next round of capacity purchases. The relationship is not symmetric. Microsoft has alternatives to OpenAI. Microsoft has its own Maia chips, its own Azure ML stack, its own partnership with Mistral, its own internal model training program. OpenAI does not have alternatives to Microsoft at the scale OpenAI needs. OpenAI's alternatives are AWS (Anthropic's primary), Google Cloud (Anthropic and Gemini), Oracle Cloud (xAI), and CoreWeave (a Microsoft-partnered neocloud). The alternatives exist. The alternatives are not equivalent at the scale OpenAI needs.^[1][4]

The structural privacy read on the Microsoft line is the surveillance-as-revenue dimension. Senator Wyden's June 2026 letters to Anthropic, Google, OpenAI, and xAI asked whether the companies would allow the government to use their AI products to surveil Americans through bulk data collection. Anthropic and Google replied. OpenAI and xAI stayed silent. The silence is the answer the Microsoft relationship makes structural: OpenAI's largest infrastructure dependency is also a vendor that has contracts with ICE, contracts with the Pentagon, contracts with every federal agency that has a cloud budget. OpenAI cannot reply to Wyden's letter with a categorical "we won't help surveil Americans" because OpenAI's largest infrastructure dependency is also the vendor that does help surveil Americans. The dependency is structural. The dependency is what the $17.2 billion pays for.^[10][11]

The IPO Window: Why the Audited Disclosure Is Timed Where It Is

OpenAI said on June 8, 2026 that it had confidentially submitted a draft S-1 to the SEC. The disclosure added that the company had "expected the filing to leak and had not decided when to go public."^[1][12] The draft S-1 is the filing that will become the IPO prospectus when OpenAI decides to go public. The audited 2025 financials are the numbers that will be in the prospectus. The audited 2025 financials are now in the public domain eight days before the public-filings window opens through the SEC's review process.^[1][12]

The timing matters because OpenAI has spent the first half of 2026 selling investors the scale argument. In March 2026, OpenAI said it had closed $122 billion in committed capital at an $852 billion post-money valuation. The round was anchored by Amazon, Nvidia, SoftBank, and Microsoft. In the same announcement, OpenAI said it was generating $2 billion in revenue per month, had more than 900 million weekly active ChatGPT users, and had more than 50 million paying subscribers. The audited 2025 financials do not contradict the $2 billion per month figure. The audited 2025 financials do contradict the implicit assumption that the $2 billion per month is enough to cover the compute bill. The annualization of $2 billion per month is $24 billion in revenue against $34 billion in costs. The annualization gap is $10 billion per year, before any fair-value adjustments, before any noncontrolling interest accounting, before any growth in compute spend.^[1][13]

The IPO window is the moment when the audited disclosure graduates from being a leak to being a public filing. The public filing is the moment when retail investors, pension funds, and index funds will see the gap between revenue and compute burn in the same document they see the equity story. The public filing is the moment when the cost-base trajectory becomes a valuation question. The valuation question is the question the company is asking capital markets to answer. The capital markets are the entity that funds the gap. The gap is what the user's data is supposed to close.^[1][12][13]

The Data Extraction Surface: What the Gap Is For

The structural privacy read on the audited disclosure is what the gap between $13 billion in revenue and $34 billion in costs is supposed to close. The gap is closed by four classes of inflow. The first class is capital raises: the $122 billion in committed capital at the $852 billion post-money valuation is the most recent. The second class is government contracts: the Pentagon deal, the ICE-related Azure dependence, the federal AI deployment contracts that flow through Microsoft. The third class is infrastructure subsidies: the Stargate / Missouri data-center thread, the Federal Data Center Enhancement Act expiration on September 30, 2026 that opens the door to a less-restricted federal data-center spending posture. The fourth class is the long-term extraction surface: the consumer-product surface (ChatGPT's 900 million weekly active users), the enterprise-product surface (the API, the Azure OpenAI Service), the government-product surface (the Pentagon contract, the federal agency deployments), and the developer-product surface (the API customers, the OpenAI Platform users).^{[1][5][6][7][8]}

The four classes of inflow are not separate. They are the same surface viewed from four sides. The capital raises fund the compute. The compute trains on the data extracted through the consumer-product, enterprise-product, government-product, and developer-product surfaces. The government contracts expand the data-extraction surface into the federal data layer. The infrastructure subsidies pay for the physical compute that the extraction surface trains on. The cycle is: capital in, data out, compute in, model out, contract in, surface out, capital in. The audited 2025 financials make the cycle visible. The cycle is what the gap between revenue and compute is for.^{[1][5][6][7][8]}

The Wyden letter is the same surface on the policy side. The Wyden letter asked whether OpenAI would help the government surveil Americans. OpenAI did not reply. The structural read is that OpenAI's largest infrastructure dependency (Microsoft) is also the vendor that helps the government surveil Americans (through ICE, through the Pentagon, through the federal cloud contracts). OpenAI cannot reply to the Wyden letter with a categorical no because OpenAI's products run on Microsoft's cloud, and Microsoft's cloud runs the federal surveillance infrastructure. The non-reply is the structural answer. The non-reply is what the $17.2 billion paid to Microsoft in 2025 pays for.^[10][11]

The Anthropic Fable 5 export-control directive is the same surface on the geopolitical side. The directive, issued June 12, 2026, suspended Anthropic's Fable 5 and Mythos 5 access for foreign nationals. The directive was triggered, per reporting, by a routine "fix this code" defensive-cybersecurity prompt. The directive's net effect was that both models were disabled for all customers, because the foreign-national cutoff was structurally inseparable from the all-customers cutoff. The structural read is that the export-control surface is the same surface that the data-extraction surface sits on. The export-control posture is the data-extraction posture, at a different layer.^[14][15]

The DOJ xAI "vital" national security filing in the NAACP lawsuit is the same surface on the military side. The DOJ memo names Grok as one of four AI models on Secret and Top-Secret classified networks. The DoD CDAO Cameron Stanley declaration says the military used Grok Gov in strikes against Iran. The structural read is that the AI-as-infrastructure dimension (the Pentagon, the intelligence agencies, the classified networks) is the same dimension the data-extraction surface is being prepared for. The dimension is not a separate beat. The dimension is the same beat at a different layer.^[16]

The SpaceX-Cursor consolidation is the same surface on the capital side. SpaceX bought Anysphere (the maker of Cursor) for $60 billion in June 2026. A single corporate entity now controls rockets, global satellite internet, a major social-media platform, a frontier AI lab, and the dominant AI coding tool. The structural read is that the capital-concentration dimension is the same dimension the data-extraction surface is being prepared for. The concentration is not a separate beat. The concentration is the same beat at a different layer.^[17]

What It Means for You Today

Three groups are affected, and the consequences diverge.

If you are a ChatGPT user, an OpenAI API customer, or a developer building on OpenAI's models. The structural read is that your data is part of the extraction surface the company is building to close the $21 billion annual gap between revenue and compute burn. The data includes the chat logs, the API request and response payloads, the fine-tuning datasets, the embedding inputs, the evaluation harness outputs, and the chat history used to improve the model. OpenAI's terms of service have varied on the question of whether chat logs are used for training. The audited 2025 financials make the question more concrete: the gap is large enough that the company has a structural incentive to use every available data point to close it. The structural incentive is the privacy story.

If you are a federal agency, a state government, or a public-sector entity evaluating AI procurement. The structural read is that the vendor you are evaluating is one of two things. It is either a vendor whose largest infrastructure dependency is a vendor (Microsoft) that has federal contracts with ICE, the Pentagon, and the intelligence agencies. Or it is a vendor that runs on the same cloud as the federal surveillance infrastructure. The Wyden letter is the surface of that question. The Wyden letter went unanswered by OpenAI and xAI. The unanswered letter is the structural signal. The signal is what the procurement decision has to price in.

If you are a retail investor, a pension-fund manager, or an index-fund allocator waiting for the OpenAI S-1. The structural read is that the audited 2025 financials are the numbers that will be in the prospectus. The numbers show a company whose cost base is expanding faster than its revenue, whose largest infrastructure dependency is a vendor that has its own model program, and whose fair-value adjustments on convertible instruments and warrants add $41.55 billion of paper loss to a $20.92 billion cash operating loss. The $852 billion post-money valuation from the March 2026 round is what the gap is supposed to close. The IPO is the moment the gap meets the public market. The public market is the entity that decides whether the gap is fundable.

The Bottom Line

OpenAI lost $38.53 billion attributable to the company in 2025 on $13.07 billion in revenue against $34 billion in costs, with a $20.92 billion operating loss and a $41.55 billion fair-value adjustment on convertible interests and warrants, according to audited documents viewed by Ed Zitron and independently verified by the Financial Times, reported by RuntimeWire on June 16, 2026 and picked up by Ars Technica and Fortune the same day.^[1][2][3] OpenAI paid Microsoft $17.2 billion of the $34 billion in costs. OpenAI owes Microsoft $3.64 billion at year-end. OpenAI's draft S-1 has been with the SEC since June 8, 2026. OpenAI has not decided when to go public.^[1][12]

The structural privacy read is what the gap between revenue and compute is for. The gap is closed by capital raises, government contracts, infrastructure subsidies, and the long-term data-extraction surface the company is building on its consumer-product, enterprise-product, government-product, and developer-product layers. The user's data is part of the surface. The surface is what is supposed to close the gap before the IPO window does. The audited disclosure makes the gap visible. The gap is what the next round of capital raises is going to fund.^{[1][5][6][7][8]}

Watch for five things over the next 30 days. First, OpenAI first-party response to the $38.53 billion attributable loss disclosure. Second, OpenAI first-party response to the $41.55 billion fair-value adjustment detail. Third, the first SEC S-1 amendment that bakes the audited figures into the public filing. Fourth, the first tier-1 outlet (Reuters, WSJ, NYT, Bloomberg) confirmation beyond the FT, Ars Technica, and Fortune pickups already on the record. Fifth, the first Wyden follow-up letter or congressional AI-surveillance hearing that asks OpenAI to reply to the original letter it ignored.

Sources

1. RuntimeWire: "Leaked OpenAI financials put a hard number on Altman's compute burn" (Ryan Merket, June 16, 2026, 4:36 pm CT, primary report on the audited 2025 financials, the $13.07B revenue, $34B costs, $20.92B operating loss, $38.53B attributable loss, $60.35B group loss, $41.55B fair-value adjustment, the $17.2B paid to Microsoft, the $3.64B Microsoft liabilities, the February 2026 funding announcement, the OpenAI Group ownership structure, the S-1 confidential filing on June 8, and the Ed Zitron viewed-documents + Financial Times independent-verification provenance)
2. Ars Technica: "Leaked financial docs show OpenAI is losing billions of dollars a year" (Benj Edwards, June 16, 2026, 16:52 UTC, the tier-1 editorial confirmation of the RuntimeWire piece, the same $20.92B operating-loss anchor, and the same attributed-to-OpenAI framing)
3. Fortune: "OpenAI financials leaked: losses, revenue, profit" (June 16, 2026, 23:54 UTC, the second tier-1 editorial confirmation, the lift from niche disclosure to structural event, and the editorial framing on the IPO-timeline timing)
4. Where's Your Ed At: Ed Zitron's prior OpenAI financial reporting (the journalist who viewed the audited 2025 documents, the prior "OpenAI is a sustained bubble" coverage, and the structural-critique anchor for the leak provenance)
5. Financial Times: "OpenAI spending hit $34bn last year ahead of planned IPO" (June 16, 2026, the FT independent verification of the $34B cost figure, the IPO timing framing, and the tier-1 editorial confirmation)
6. Senator Elizabeth Warren: Pentagon AI Deals Investigation (the Warren Senate inquiry into the Pentagon's Anthropic and OpenAI deals, the surveillance-as-revenue dimension, and the regulatory parallel)
7. Senator Ron Wyden: AI Surveillance Letter to Anthropic, Google, OpenAI, xAI (June 2026, the unanswered OpenAI letter, the surveillance-as-revenue dimension, and the structural-read anchor for the Microsoft dependency)
8. Reuters: "SpaceX to buy Anysphere for $60 billion" (June 16, 2026, the SpaceX-Cursor consolidation, the capital-concentration dimension, and the structural beat)
9. Hacker News thread for RuntimeWire OpenAI financials piece (id 48565130, 197p / 207c at 07:35 UTC scan on June 17, 2026, the engagement-justification primary source for the +2,671% compound from 7p at 02:09 UTC morning cycle, the biggest engagement growth of day)
10. State of Surveillance: "Congress Asked AI Companies If They'd Help Spy on Americans" (the Wyden letter, the unanswered OpenAI silence, the surveillance-as-revenue dimension, and the structural-read anchor for the Microsoft dependency)
11. State of Surveillance: "Warren Opens Investigation Into Pentagon's AI Deals" (the Warren inquiry, the regulatory dimension, and the parallel to the Wyden silence)
12. OpenAI: "Confidential submission of draft S-1 to the SEC" (June 8, 2026, the draft S-1 filing, the expected-leak statement, and the IPO-timeline anchor)
13. OpenAI: "OpenAI raises $122 billion to accelerate the next phase of AI" (March 2026, the $852B post-money valuation, the Amazon / Nvidia / SoftBank / Microsoft anchor investors, the 900M weekly active users, the 50M+ paying subscribers, and the $2B per month revenue framing)
14. State of Surveillance: "Anthropic Fable 5 + Mythos 5 Suspended by US Export Control" (the June 12, 2026 export-control directive, the "fix this code" trigger, and the compute-export dimension of the data-extraction surface)
15. State of Surveillance: "Anthropic Fable 5 + Mythos 5 Day-5: Stratechery Safety Superpower" (the structural critique anchor for the Anthropic thread, the Day-5 update, and the compute-export dimension)
16. State of Surveillance: "DOJ Says xAI's Grok Is 'Vital' National Security Infrastructure" (the DOJ memo, the Grok on Secret and Top-Secret networks, the CDAO Cameron Stanley declaration, and the AI-as-infrastructure dimension)
17. State of Surveillance: "SpaceX Buys Cursor for $60 Billion" (the SpaceX-Anysphere consolidation, the single-corporate-entity control of rockets + satellite internet + social media + frontier AI + AI coding tool, and the capital-concentration dimension)

Related Coverage

• Congress Asked AI Companies If They'd Help Spy on Americans. The Wyden letter that OpenAI and xAI did not answer. The unanswered letter is the structural signal that the Microsoft $17.2B dependency is also the vendor relationship that does help surveil Americans. The dependency is what the $17.2B pays for.
• Warren Opens Investigation Into Pentagon's AI Deals. The Senate inquiry into the Pentagon's Anthropic and OpenAI deals. The surveillance-as-revenue dimension is the same dimension the audited 2025 financials make visible. The dimension is the gap between revenue and compute.
• OpenAI Reworks Pentagon Deal After Surveillance Backlash. The contract language that banned "commercially acquired" data use but left the "intentionally" qualifier and the public-data loophole. The deal is the same surface the $34B compute burn funds.
• EFF Calls Out OpenAI's "Weasel Words". Why digital-rights advocates say the amended Pentagon contract will not stop mass surveillance. The weasel-words critique is the same critique the audited 2025 financials make operational. The gap is what the contract fills.
• Anthropic Fable 5 + Mythos 5 Suspended by US Export Control. The June 12, 2026 directive. The compute-export dimension of the data-extraction surface. The directive is what the gap between revenue and compute burn has to be defended against.
• Anthropic Fable 5 + Mythos 5 Day-5: Stratechery Safety Superpower. The structural critique anchor for the Anthropic thread. The safety-superpower framing is the same framing that the audited 2025 financials make operational on the OpenAI side.
• DOJ Says xAI's Grok Is 'Vital' National Security Infrastructure. The DOJ memo that names Grok on Secret and Top-Secret networks. The AI-as-infrastructure dimension of the data-extraction surface. The dimension is the classified-network layer of the gap.
• SpaceX Buys Cursor for $60 Billion. The SpaceX-Anysphere consolidation. The single corporate entity now controls rockets, satellite internet, social media, frontier AI, and AI coding tools. The capital-concentration dimension of the data-extraction surface. The concentration is the gap at the cap-table layer.
• 42 Attorneys General Gave AI Companies 37 Days. The Deadline Passed.. The bipartisan coalition of state attorneys general probing AI companies on safety. The regulatory dimension of the data-extraction surface. The dimension is the state-law layer of the gap.