Leaked OpenAI financials show $38.5B loss and compute burn
Leaked financials suggest OpenAI lost $38.5 billion in 2025, but commenters note that roughly $30 billion of that stems from a one-time, non-cash accounting charge tied to its conversion to a for‑profit structure. Stripping that out, OpenAI appears to have about $13 billion in revenue against $7.5 billion in cost of revenue and roughly $8 billion in underlying losses, implying that inference may already be profitable while massive R&D and sales spending drive the red ink. The leak fuels debate over whether such a business should be allowed to IPO at a near‑trillion‑dollar valuation, how sustainable its compute-heavy model is amid rising competition and open-weight alternatives, and whether projected growth can justify current investor optimism.
IPO, public markets, and index funds
- Some argue firms with multi‑billion losses shouldn’t be allowed to IPO, to avoid “bag‑holding” by the public and retirement savers.
- Others counter that IPOs are voluntary purchases; if a company meets existing regulations it should be free to list, with markets pricing risk/reward.
- Index inclusion is raised as a complication: once a big loss‑making tech stock joins major indices, many investors are indirectly exposed.
- Several comments note that index rule changes, not IPOs themselves, are the real problem.
Financials: losses, accounting, and cost structure
- The headline
$38.5B 2025 loss is widely noted as mostly a one‑time, non‑cash accounting charge ($30B) from converting prior investor rights into equity. - Stripping that out, commenters cite an ~$8B operational loss, with 2025 revenue around $13B.
- Reported spending mix (approx): $7.5B cost of revenue, ~$19B R&D, ~$5.7B sales/marketing, ~$1.6B G&A.
- Some see the adjusted numbers as “better than expected” and compatible with eventual profitability; others remain alarmed by the burn rate.
Inference economics and “subsidy” debate
- A major thread: revenue (
$13B) exceeds cost of revenue ($7.5B), which many take as evidence that inference tokens are not sold below cost. - Skeptics respond that ignoring training/R&D is like ignoring depreciation on a factory; “profitable tokens” don’t mean a profitable business.
- There is disagreement about how much compute is subsidized (e.g., via partners or governments); precise subsidies remain unclear.
R&D, training, and business sustainability
- Heavy R&D (~$19B) is seen as both a moat (scale, frontier models) and a structural burden (models have short competitive lifetimes).
- Some argue OpenAI could be profitable if it froze R&D, but most say that would quickly erode its position versus rivals and open‑source.
- Questions persist over what exactly is in “R&D” versus cost of revenue, especially training vs ongoing engineering.
Moat, competition, and valuation
- Views diverge sharply: some see strong product–market fit, fast revenue growth, and a plausible path to break‑even at $25–30B revenue.
- Others argue there’s no durable moat: strong competitors (Anthropic, Google, Chinese labs, open‑weights) and easy switching for enterprises.
- Valuation debates hinge on whether massive future AI “labor replacement” and market dominance are realistic, or bubble thinking.