OpenAI signs $38B, seven-year compute deal with AWS

What the $38 billion actually buys

The headline number is $38 billion over seven years, but the specifics matter more than the total. AWS is providing OpenAI with hundreds of thousands of NVIDIA GPUs—explicitly named as GB200 and GB300 chips—clustered via Amazon EC2 UltraServers, plus the ability to scale to tens of millions of CPUs for agentic workloads.

The deployment timeline is aggressive. OpenAI starts using AWS compute immediately, with all contracted capacity targeted to be deployed before the end of 2026, and room to expand further into 2027 and beyond. That is a compressed build-out for infrastructure of this magnitude.

AWS describes the architecture as GPUs clustered on the same network to enable low-latency performance across interconnected systems. The clusters are designed to handle both ChatGPT inference serving and training of next-generation models—not one or the other.

The GPU-plus-CPU split points at agents

One detail is easy to skim past: the GPUs handle generative AI, but the deal specifically calls out the ability to expand to tens of millions of CPUs to 'rapidly scale agentic workloads.'

That distinction reflects how agent systems actually run. The model call is GPU-bound, but the surrounding orchestration—tool execution, retrieval, coordination, the many non-inference steps an agent takes—is CPU work. Provisioning that headroom separately signals OpenAI expects agentic traffic to grow in a way that ordinary GPU capacity planning would not capture.

OpenAI is spreading its compute across providers

Sam Altman's framing is the tell here. He describes the deal as strengthening 'the broad compute ecosystem'—language that treats AWS as one supplier among several rather than the exclusive backbone.

Scaling frontier AI requires massive, reliable compute. Our partnership with AWS strengthens the broad compute ecosystem that will power this next era and bring advanced AI to everyone.Montana Labs

AWS, for its part, leans on operational credibility: it cites experience running clusters topping 500,000 chips. The subtext is that at frontier scale, reliability at the cluster level is the differentiator, not just chip availability.

The announcement also notes an existing relationship: OpenAI's open weight models are already on Amazon Bedrock, used by named customers including Peloton, Thomson Reuters, Comscore, and Verana Health for coding, scientific analysis, and agentic workflows. So this compute deal sits on top of an existing distribution channel, not a cold start.

What a single-vendor timeline means for teams building on OpenAI

For anyone building products on OpenAI's APIs, the specific implication is capacity concreteness. A commitment with named chips, a named deployment deadline (end of 2026), and named cloud infrastructure is a stronger signal about near-term availability than vague scaling promises.

It also reinforces that OpenAI's capacity now flows through multiple clouds. Teams that assumed OpenAI was a single-cloud dependency should update that model—the compute serving your inference calls may increasingly run on AWS hardware, whatever cloud your own stack uses.

The practical read: frontier labs are competing on secured, dated compute contracts as much as on model architecture. When capacity is the binding constraint, a $38 billion supply agreement with a firm deployment schedule is itself a product roadmap statement.