The number that everyone is citing from OpenAI’s leaked 2025 financials is $38.5 billion. That’s the net loss. It’s large, it’s shocking, and it is also — according to financial analysts who have read the actual documents — heavily distorted by a one-time accounting charge related to the company’s restructuring from nonprofit to for-profit.
The number that nobody is really talking about is $17.2 billion.
That’s what OpenAI paid Microsoft in 2025 alone. Not a general operating cost. Not an infrastructure expense split across multiple vendors. A single payment to a single company: $10.59 billion classified as research and development expenses (mostly compute) and $6.047 billion in cost-of-revenue charges. The company that is supposed to be building the future of artificial intelligence is, structurally, a very expensive customer of Microsoft Azure.
What OpenAI’s 2025 Financials Actually Say
The figures below come from audited financial documents published by journalist Ed Zitron and subsequently verified by the Financial Times, which obtained independent confirmation of the data.
| Metric | 2024 | 2025 |
|---|---|---|
| Revenue | $3.7B | $13.07B |
| Total Costs & Expenses | $12.48B | $34B |
| Operating Loss | $8.78B | $20.92B |
| Net Loss | — | $38.5B |
| Paid to Microsoft | — | $17.2B |
Revenue tripled year-on-year. That’s real growth. The problem is costs grew faster — from $12.48 billion to $34 billion — and the company’s dependency on Microsoft infrastructure grew with it.
The $38.5 billion net loss headline figure is misleading. The Financial Times and multiple financial analysts confirmed that approximately $41.55 billion of that number is a non-cash accounting charge: a revaluation of convertible interests and warrant liabilities triggered by the nonprofit-to-for-profit conversion. Remove the accounting event, and you’re left with the $20.92 billion operating loss, which is the number that reflects what it actually costs OpenAI to run its business day-to-day.
That number is still eye-watering. But there’s important context: the loss ratio is improving. In 2024, OpenAI spent $2.37 to generate every dollar of revenue. In 2025, that ratio declined to $1.60. The direction is right. The distance to profitability is not.
At $13.07 billion in revenue and $34 billion in costs, OpenAI needs to roughly triple revenue again — while holding costs flat — to approach break-even. The compute costs baked into the Microsoft relationship make holding costs flat structurally difficult.
The Microsoft Dependency Nobody Mapped
OpenAI’s relationship with Microsoft is not a partnership in any conventional sense of the word. Microsoft has invested roughly $13 billion into OpenAI across multiple tranches. In exchange, OpenAI runs almost exclusively on Azure infrastructure, and Microsoft receives a substantial share of OpenAI’s profits until it recoups its investment — after which Microsoft takes a 49% equity stake.
The $17.2 billion OpenAI paid Microsoft in 2025 breaks down as:
- $10.59 billion: R&D expenses, primarily compute costs for training and running models
- $6.047 billion: Cost of revenue — the compute cost of serving every ChatGPT query to every user
Every time someone asks ChatGPT a question, money flows to Microsoft. The investor is also the landlord, the infrastructure provider, and the primary cost center. There is no scenario where OpenAI cuts its Microsoft bill without either moving infrastructure — an enormously complex and expensive migration — or dramatically reducing the scale of its model operations.
This matters for the AI sustainability debate in a way that the headline loss figure doesn’t capture. OpenAI’s path to profitability isn’t just about getting more subscribers. It’s about whether it can ever structurally reduce the cost of serving those subscribers without Microsoft’s blessing.
What Your Subscription Is Actually Costing Them
Research firm SemiAnalysis published an analysis of OpenAI’s subscription economics that makes the Microsoft dependency look like a rounding error.
The $200-per-month ChatGPT Pro plan — the one marketed to heavy users — can cost OpenAI up to $14,000 per month per user in API-equivalent compute when a subscriber fully maximizes the plan’s capabilities through agentic AI workflows.
Agentic workflows are the key term here. Traditional ChatGPT usage — typing a question, receiving an answer — consumes a modest amount of compute. Agentic tasks are different: the model autonomously plans, uses tools, browses the web, writes and executes code, iterates, and runs continuously for extended periods. Each of those actions burns tokens. Token burn translates directly into compute cost.
The subsidy ratio SemiAnalysis identified: a power user can consume approximately 70 times the compute value of their $200 subscription.
| Subscription | Monthly Price | Max Compute Cost (Power User) | Subsidy Ratio |
|---|---|---|---|
| ChatGPT Plus | ~$25 | ~$1,750 | 70x |
| ChatGPT Pro | $200 | ~$14,000 | 70x |
The model only works financially because most subscribers are not power users. The light users subsidize the heavy users, in the same way that most gym memberships fund facilities for the 6 AM crowd.
The problem — as SemiAnalysis and OpenAI’s own executives have acknowledged — is that agentic AI use is growing. As AI workflows become standard in professional settings, the ratio of power users to casual users shifts. The buffer of passive subscribers shrinks. The per-user subsidy cost rises.
OpenAI CFO Sarah Friar acknowledged in an interview with the Financial Times that the $200 Pro plan was deliberately priced to attract users rather than generate margin. The company was, in plain terms, buying market share with investor capital.
The 11% Rule and When the Math Breaks
SemiAnalysis identified a specific threshold: once a subscriber’s usage exceeds 11% of the plan’s stated value, OpenAI begins losing money on that subscriber.
For the $200 Pro plan, 11% utilization represents $22 worth of compute — roughly what a moderate daily user generates. This means the break-even point isn’t reserved for extreme power users. Any professional who uses ChatGPT as a consistent daily work tool is likely already past it.
The $14,000 figure — a subscriber who runs continuous agentic workflows around the clock — is the theoretical ceiling, not the typical case. But the direction is the concern. SemiAnalysis projects that as agentic use increases from a niche workflow to a standard professional tool, the percentage of subscribers operating above the break-even threshold rises from a manageable minority toward a majority.
At that point, the subscription pricing model breaks. OpenAI would need to either raise prices dramatically — SemiAnalysis suggested the actual sustainable price for a power-user plan may need to approach $3,000 per month — or shift to usage-based billing that directly tracks token consumption.
Neither option is painless. A price hike of that magnitude on a consumer product would drive significant churn. Usage-based billing removes the predictable revenue that makes subscription businesses attractive to investors.
Why This Is a Structural Problem, Not a Growing Pain
OpenAI isn’t the only company in this position. Every major AI subscription product — Anthropic’s Claude, Google’s Gemini Advanced, Microsoft’s Copilot — faces a version of the same economics. The compute cost of running frontier models doesn’t scale linearly with capability. As models get better, they get more expensive to run, often by a multiplier that outpaces the price increases companies can reasonably charge users.
The Netflix comparison that circulates in these discussions — early losses justified by subscriber growth, eventual profitability as the user base matures — misses a structural difference. Netflix’s cost to serve a subscriber does not increase as the subscriber watches more content. The bandwidth cost is marginal. The content library is a fixed cost that amortizes across users.
OpenAI’s cost to serve scales directly with usage. More usage equals more compute equals more money to Microsoft. The more valuable ChatGPT becomes to a user, the more it costs OpenAI to retain that user.
This is the dynamic that makes OpenAI’s improving loss ratio — from $2.37 per revenue dollar to $1.60 — difficult to read as unambiguously positive. Yes, the ratio is heading in the right direction. But the absolute loss is widening at the same time: from $8.78 billion operating loss to $20.92 billion. Revenue grew faster than before. Costs grew faster than revenue.
The question investors will eventually ask — and based on the IPO S-1 filing timeline, that conversation is approaching — is whether there is a plausible financial model in which OpenAI generates enough revenue, from enough subscribers, at prices high enough to cover infrastructure costs concentrated in a single vendor relationship, without losing the users it needs to pay those bills.
OpenAI’s own public statements do not answer that question. The leaked financials confirm it hasn’t been answered yet.
Sources
- Ed Zitron, Where’s Your Ed At (Substack) — initial publication of 2025 OpenAI financial data
- Financial Times — independent verification of OpenAI 2025 audited financials
- SemiAnalysis — compute cost analysis of ChatGPT Pro subscription economics
- OpenAI CFO Sarah Friar — Financial Times interview on Pro plan pricing strategy
