Leaked internal documents, analyzed in a TechCrunch led series of reports, reveal the scale of OpenAI payments to Microsoft under a revenue share arrangement and detail inference spend on Azure. Public summaries and independently circulated copies place OpenAI inference costs in the billions for 2024 to 2025, with some coverage citing a figure near 12.4B. These numbers expose the raw infrastructure bills behind large language models and shift the debate over AI cloud cost analysis, pricing and margins.

Why infrastructure economics matter

AI platforms incur two distinct kinds of compute spend. Training costs are episodic and capital intensive. Inference costs are ongoing and accrue each time a deployed model answers a user query. For commercial AI providers, inference can become the dominant recurring expense as usage scales. The leaked documents put a spotlight on those recurring bills and on how close ties to a cloud provider can reshape unit economics.

Key findings

• Inference spend: Leaks and public summaries estimate OpenAI inference costs in the billions for 2024 to 2025, with some reporting an aggregate near 12.4B. Inference is the compute used whenever a model generates text or an image for a user.
• Revenue share transfers: Files show substantial transfers from OpenAI to Microsoft under their revenue share agreement, indicating Microsoft receives a material slice of revenue tied to Azure hosted inference.
• Source and timing: The TechCrunch led analysis, published November 15, 2025, aggregated internal slides, billing summaries and blogged data from the leaks.
• Strategic positioning: The documents underscore Microsoft as both a distributor of OpenAI powered products and the principal infrastructure partner hosting inference workloads on Azure.

Implications for margins pricing and strategy

These findings matter for industry players customers and investors. If inference bills reach multi billion levels then gross margins on AI services may be thinner than subscription prices imply. Companies will need to:

• Improve inference efficiency through model engineering and runtime optimizations.
• Explore hardware bargains and cost management tools to reduce AI infrastructure costs.
• Adopt tiered pricing rate limits or premium features for heavy usage customers to preserve unit economics.

For cloud providers like Microsoft the revenue share arrangement aligns incentives and secures recurring cloud revenue. For model makers the deal reduces capital expenditure but increases dependency on a single provider.

Industry context

This reporting aligns with broader trends in automation and cloud economics where platform scale players consolidate infrastructure to capture performance and pricing advantages. Rising attention to per inference costs is prompting engineers to optimize models switch to more efficient runtimes and experiment with on premise or hybrid architectures. Conversations around AI compute pricing cloud billing for AI workloads and cost optimization are accelerating across the sector.

What to watch next

• Will cloud providers sign similar revenue share deals with other model makers and how will that affect competition in cloud computing pricing?
• Will model providers push harder for inference efficiency to lower per query costs?
• Will regulators scrutinize deep ties between cloud vendors and AI platform companies for market power concerns?

Quick FAQ

How much did OpenAI spend on Azure inference

Public summaries of the leaks estimate OpenAI inference costs in the billions for 2024 to 2025 with some reports citing roughly 12.4B for that period.

What is inference cost

Inference cost is the compute consumed each time a deployed model responds to a user request. It is a recurring operational expense that scales with usage.

How does this affect pricing for customers

If inference bills are high providers may adopt tiered pricing pay as you go models or rate limits to manage costs and preserve margins.

Conclusion

The leaks do more than reveal dollar amounts. They force a reckoning with the business model behind modern AI: models create demand but the compute required to serve that demand creates substantial recurring costs. For businesses building or buying AI solutions the takeaway is practical. Evaluate model performance alongside infrastructure arrangements and cloud billing terms that shape long term margins. Watch OpenAI and Microsoft for any public clarifications and for competitive moves that could reshape AI compute pricing and cloud cost management.