Newly reported internal documents are putting a sharper frame around one of the central questions in artificial intelligence: how expensive is it to keep major AI models running after they are built?
According to Techcrunch and blogger Ed Zitron, the documents describe large payments between OpenAI and Microsoft. The numbers have not been officially confirmed, but they suggest that OpenAI inference costs are a major pressure point for the company.
What the leaked figures suggest
Citing internal documents, AI critic Ed Zitron says OpenAI paid Microsoft about 493.8 million dollars in 2024 under a 20 percent revenue-share arrangement. He also says another 865.9 million dollars followed in the first three quarters of 2025.
Those payments matter because they can be used to estimate the minimum revenue behind them. With a 20 percent cut, the reported Microsoft share implies minimum revenue of about 2.47 billion dollars in 2024 and 4.33 billion dollars in the first three quarters of 2025.
That does not mean those numbers are the full OpenAI revenue picture. Techcrunch reports that the revenue share works both ways. Microsoft sends roughly 20 percent of the revenue from Bing and the Azure OpenAI service back to OpenAI, so the leaked payments are described as Microsoft’s net share after subtracting what it returns.
In plain terms, the payments cited by Zitron may reflect only part of a broader commercial relationship. Actual revenue is likely higher because Microsoft’s payments back to OpenAI are not included in those implied minimums.
Why inference is the pressure point
The more striking part of the report is not only the revenue estimate. It is the reported cost of running the models.
Zitron says OpenAI’s inference costs reached 3.77 billion dollars in 2024 and 8.67 billion dollars in the first nine months of 2025. Inference means the compute and operating expense required when users and customers actually run the models.
That distinction is important. Training creates new models. Inference keeps those models available for everyday use. According to Techcrunch, Microsoft credits cover much of the training cost for new models, while inference spending is mostly cash.
If those figures are accurate, the economics look demanding. The reported inference costs are larger than the minimum revenue implied by Microsoft’s share in both periods cited in the source. That does not prove the full business is upside down, because the minimum revenue numbers are not the same as total revenue. But it does show why inference spending is central to any discussion of AI profitability.
The Microsoft relationship is more complex than one payment stream
The source describes a two-way revenue-share structure. OpenAI pays Microsoft a share, while Microsoft also sends OpenAI a share of revenue from Bing and the Azure OpenAI service.
That makes the leaked figures harder to read than a simple bill. They represent Microsoft’s net share, not a complete ledger of every dollar moving between the companies.
Still, the reported numbers help explain why the OpenAI and Microsoft relationship is so important to the economics of large language models. The arrangement connects model usage, infrastructure, Bing, the Azure OpenAI service, and revenue sharing into one financial picture.
Sam Altman has previously mentioned projecting 20 billion dollars in annualized revenue by the end of 2025. The leaked figures do not confirm that target. They instead show the scale of payments and costs that may sit underneath OpenAI revenue growth.
Profitability still looks distant
The reported inference spending is only one part of the cost base. The source also points to training costs, substantial staffing expenses, and other operational spending.
That is why the leaked documents, if accurate, do not simply describe a fast-growing AI company. They describe a company where growth may require very large cash outlays just to serve demand.
For AI businesses, this is the hard part of the model. More usage can mean more revenue, but it can also mean more inference. If the cost of serving that usage rises alongside revenue, profitability becomes harder to reach.
The source makes clear that neither OpenAI nor Microsoft has commented on the reports. The figures are therefore best treated as reported and unconfirmed, not as official company disclosures.
What this means for the AI market
The broader implication is straightforward. The cost of large AI systems is not limited to research or training. Once models are in use, inference can become a recurring expense at enormous scale.
That makes OpenAI inference costs a key measure to watch, alongside OpenAI revenue, Microsoft revenue share payments, and revenue from products connected to Bing and the Azure OpenAI service.
The leaked numbers do not settle the question of whether OpenAI can become profitable. They do show why that question is more complicated than headline revenue growth. If the reported costs are close to reality, the path to profit depends not just on selling more AI services, but on whether the cost of running those services can be brought under control.