OpenAI Signs $38B AWS Deal to Scale Agentic AI Workloads

OpenAI announced a multi-year, $38 billion agreement with Amazon Web Services (AWS) to access large-scale compute for training and running advanced AI models. The deal gives OpenAI immediate use of AWS capacity—including “hundreds of thousands” of NVIDIA GPUs via Amazon EC2 UltraServers and the ability to scale to tens of millions of CPUs—with deployment of the contracted capacity targeted before the end of 2026 and room to expand thereafter. 

What OpenAI And AWS Leaders Said

Key facts from the announcement 

• Value & term: $38 billion over a multi-year (seven-year) strategic partnership. 

• Compute profile: Access to “hundreds of thousands” of state-of-the-art NVIDIA GPUs (GB200/GB300 referenced in AWS materials) and the option to scale to tens of millions of conventional CPUs for agentic workloads. 

• Timing: OpenAI will begin using AWS compute immediately; full capacity is targeted to be online by the end of 2026 with potential expansion into 2027 and beyond. 

• Official lines: “Scaling frontier AI requires massive, reliable compute,” Sam Altman said. “Our partnership with AWS strengthens the broad compute ecosystem that will power this next era and bring advanced AI to everyone.” Matt Garman of AWS framed the deal as a demonstration of AWS’s ability to run large-scale AI infrastructure securely and reliably. 

Significant for Enterprises and AI builders

Large, long-dated infrastructure commitments reshuffle how providers and customers plan AI roadmaps. For enterprises that embed models into products or workflows, the AWS deal signals wider availability of high-performance inference and training capacity across the cloud ecosystem—potentially improving redundancy and choice for model deployment. At the same time, the agreement is a reminder of the enormous and recurring operating cost of frontier AI; OpenAI’s compute commitments this year are part of a wider set of multi-hundred-billion-dollar partnerships with multiple vendors. 

AWS says the infrastructure will use Amazon EC2 UltraServers configured to interconnect large GPU clusters (optimised for low latency and high throughput) while supporting both training and inference workloads. That architecture aims to let OpenAI run everything from real-time ChatGPT inference to next-generation model training on the same interlinked fabric—an efficiency and performance design that matters at this scale. 

Strategic implications and market context

• Diversification of cloud suppliers: The agreement follows OpenAI’s broader move to work with multiple cloud and chip vendors after its restructuring; it reduces exclusive reliance on any single hyperscaler and increases negotiating latitude for compute procurement. Several outlets view the AWS pact as both strategic and symbolic in that context.

• Competitive ripple effects: Hyperscalers supplying GPU clusters (AWS, Microsoft Azure, Google Cloud) and large chipmakers (NVIDIA, AMD) are central to the AI supply chain. Long-term contracts of this size reshape capacity planning, inventory, and pricing dynamics across the industry. 

The companies disclosed high-level capacity and timing but did not publish detailed economic terms (beyond the $38B headline), margin impacts, or the precise mix of training versus inference commitments. Practitioners will want to know how much capacity targets real-time inference for millions of users versus model training and what the agreement means for OpenAI’s cost structure and unit economics. Independent metrics on service reliability, latency, and regional availability will determine business impact for global customers.