A recent data leak sheds light on the massive compute costs shouldered by OpenAI and reveals a peculiar, two-way revenue-sharing agreement with Microsoft.
The documents explain how, even with colossal revenue growth, OpenAI is likely still operating at a loss.
The Peculiar Rev-Share Arrangement
The core of the partnership involves a bilateral revenue-sharing scheme: each side takes approximately 20% from different layers of the same AI stack.
- OpenAI Pays Microsoft: OpenAI gives Microsoft about 20% of its revenue. This covers their $13B partnership deal and, critically, their usage of the Azure cloud platform.
- Microsoft Pays OpenAI: Microsoft, in turn, returns roughly 20% of the revenue generated by its Bing and Azure OpenAI Service. This is because these services are powered directly by OpenAI's proprietary models.
The Money Trail
The leaked numbers reveal the net payouts to Microsoft after funds are returned to OpenAI:
- 2024: $493.8 Million
- Q1-Q3 2025: $865.8 Million
By reversing the 20% share calculation, we can infer OpenAI's massive revenue:
- 2024 Revenue: At least $2.5 Billion
- Q1-Q3 2025 Revenue: ~ $4.3 Billion
The Budget Killer: Inference Costs
While model training is largely covered by Microsoft credits, inference—the real-time operation of models (every user request, every chat session)—is a huge, immediate cash drain.
According to estimates by Zitron, the costs associated with inference alone are staggering:
- 2024 Inference Cost: ~ $3.8 Billion
- Q1-Q3 2025 Inference Cost: ~ $8.65 Billion
Previously reported figures for total compute costs:
- Total Compute Cost (2024): ~ $5.6 Billion
- Cost of Revenue (H1 2025): ~ $2.5 Billion
The Bottom Line
When the enormous real-time inference costs are compared against the explosive revenue growth, the conclusion is clear: OpenAI is likely still deep in the red. The cost of serving its models to the world outpaces its income.
A Multi-Cloud Hedge
To mitigate its reliance on Azure and gain leverage for better pricing, OpenAI is actively diversifying its infrastructure. It has signed deals with CoreWeave, Oracle, AWS, and Google Cloud, signaling a clear strategy to drive competition for hardware and secure a multi-cloud future.