OpenAI just rewrote the rules on who powers its AI — and how much it’s willing to spend to stay ahead.

By the numbers: OpenAI has spread what looks like a $600 billion compute commitment across major cloud providers. That includes a new $38 billion, 7-year deal with AWS for access to high-performance GPU clusters and millions of CPUs. longside that, reported allocations to Microsoft and Oracle remain massive — about $250 billion to Microsoft and $300 billion to Oracle under various long-term arrangements.

What’s changed

• OpenAI no longer relies exclusively on Microsoft Azure. The AWS deal marks a visible pivot away from exclusivity.
• The company is spreading its bets — diversifying risk, securing more capacity, and gaining bargaining leverage.
• Some of the capacity isn’t immediately available: full deployment for the AWS components (for example) isn’t expected until end of 2026 or later.

Why it matters

• AI workloads demand massive infrastructure. Training and running large-scale models isn’t something you can spin up overnight. OpenAI’s spending reflects that reality.
• This is becoming less about “cloud as utility” and more like “infrastructure capex under long-term contract.” OpenAI is securing access to next-generation AI hardware (NVIDIA chips like the GB200/GB300) and building redundancy across providers.
• The shift signals changing dynamics among the big cloud providers. Microsoft may still have deep ties with OpenAI — but AWS and Oracle now have a stake in powering its future models. Competition for AI infrastructure is heating up.

The Big Picture

OpenAI is treating compute resources like core infrastructure — not just utilities. It’s locking in capacity years ahead, across multiple clouds, to ensure scale and flexibility. For other AI companies and enterprises watching closely, the lesson is clear: if you want to compete at the frontier, you’ll need more than just software. You’ll need a multi-cloud strategy, long-term contracts, and probably billions (or hundreds of billions?) of dollars of committed compute capacity.