Leaked audited financial statements, first published by blogger Ed Zitron (Where's Your Ed At) and independently verified by the Financial Times, show OpenAI growing revenue from $3.7 billion in 2024 to $13.07 billion in 2025 while incurring steep losses. Zitron reports a GAAP net loss attributable to OpenAI of $38.53 billion in 2025 - up from $5.09 billion in 2024 - which includes approximately $41.55 billion in non-cash charges from OpenAI's nonprofit-to-for-profit conversion, including changes in fair value of convertible interests and warrant liability (Benzinga). Fortune reports an operating loss of roughly $21 billion - revenue minus ~$34B in total costs - while the Financial Times estimates an adjusted loss of roughly $8 billion after stripping those one-time items (Gizmodo, Ars Technica). Key 2025 expense drivers include R&D of $19.18 billion (up from $7.81B in 2024) and $10.59 billion in payments to Microsoft, per Ars Technica.
Revenue and Costs
Leaked audited financial statements, first published by blogger Ed Zitron on Where's Your Ed At and independently verified by the Financial Times, show OpenAI reporting $13.07 billion in 2025 revenue - more than tripling from $3.7 billion in 2024. Total costs and expenses reached approximately $34 billion in 2025, driven by $19.18 billion in research and development and $5.73 billion in sales and marketing (Where's Your Ed At, Ars Technica). Revenue came in ahead of OpenAI's internally forecast target of $10 billion.
Three Loss Figures - Why They Differ
Coverage of these statements references three distinct loss measures that should not be conflated. Fortune reports an operating loss of roughly $21 billion - what you get by subtracting ~$34B in operating costs from $13B in revenue. Ed Zitron's Where's Your Ed At, confirmed by Benzinga and QZ, reports a GAAP net loss attributable to OpenAI of $38.53 billion; the ~$17.5B gap between the two figures reflects approximately $41.55 billion in non-cash charges tied to OpenAI's conversion from nonprofit to for-profit status, including changes in the fair value of convertible interests and warrant liability. A Financial Times source, cited by Ars Technica and Gizmodo, estimates that after stripping out those one-time, non-cash items, the adjusted 2025 loss is closer to $8 billion.
Key Expense Drivers
Ars Technica's line-item reporting identifies R&D as the dominant cost center, growing from $7.81 billion in 2024 to $19.18 billion in 2025. Ars Technica also reports $10.59 billion in R&D-related payments made to Microsoft in 2025 - a significant intercompany item that materially affects how the R&D and cost-of-revenue figures read on the income statement. Cost of revenue rose to $7.5 billion, consistent with expanded model inference and cloud hosting activity.
Prior Forecasts and Context
Internal management documents previously reported by The Information and summarized by Yahoo Finance projected a $14 billion loss for 2026 and a path to very large revenues by 2029. Those figures predated the leaked 2025 audited statements and were forward-looking estimates, not audited results. The revenue trajectory - $3.7B to $13B in one year - is the clearest signal of rapid commercial scaling, even as operating and cash losses remain large relative to revenue.
What to Watch
Formal commentary from OpenAI and any regulatory filings tied to a potential IPO will be the authoritative reference points for reconciling these figures. Observers should track disclosures around the nonprofit-to-for-profit conversion charges, which are the primary driver separating the $38.5B GAAP headline loss from the ~$21B operating loss and the ~$8B adjusted loss. Microsoft-related intercompany payments are also worth watching as a key variable in reported R&D expenditure.
Scoring Rationale #
Leaked audited financials for the world's most prominent AI company, corroborated by multiple major outlets, materially affect investor, partner, and practitioner assessments of AI sector economics. The three distinct loss measures ($21B operating, $38.5B GAAP, ~$8B adjusted) require careful interpretation and the story warrants a major rating, just below a frontier model release or landmark regulation.
Practice interview problems based on real data
1,500+ SQL & Python problems across 15 industry datasets — the exact type of data you work with.