AWS and OpenAI Sign $38 Billion Multi-Year Compute Deal

Amazon Web Services and OpenAI have announced a multi-year strategic partnership worth $38 billion, giving OpenAI immediate access to AWS infrastructure to run and scale its core AI workloads.

The agreement, which the companies say will see continued growth over the next seven years, commits OpenAI to a massive expansion of compute capacity on AWS. OpenAI will start using AWS compute immediately, with all capacity targeted for deployment before the end of 2026 and the option to expand further into 2027 and beyond.

What AWS is providing

The infrastructure AWS is building for OpenAI features what the companies describe as a sophisticated architectural design optimized for maximum AI processing efficiency and performance. The deployment clusters NVIDIA GB200 and GB300 GPUs via Amazon EC2 UltraServers on the same network, enabling low-latency performance across interconnected systems.

The scale is large even by the standards of frontier AI. AWS will provide OpenAI with Amazon EC2 UltraServers featuring hundreds of thousands of chips, with the ability to scale to tens of millions of CPUs for advanced generative AI workloads. AWS says it has experience running AI infrastructure securely and reliably with clusters topping 500,000 chips.

The clusters are designed to handle a range of workloads. These span from serving inference for ChatGPT to training next-generation models, with flexibility built in to adapt to OpenAI's evolving needs — a notable detail, since it positions AWS as a supplier for both sides of OpenAI's compute budget: training frontier models and serving the millions of users who interact with ChatGPT.

Why it matters

The deal lands amid unprecedented demand for computing power. As frontier model providers push their models toward new levels of intelligence, they are increasingly turning to AWS for the performance, scale, and security it can provide, according to the announcement. A $38 billion commitment from OpenAI — one of the largest AI companies and, until now, most closely associated with Microsoft's Azure infrastructure — signals how the market for frontier compute has widened.

The agreement also extends an existing commercial relationship. Earlier this year, OpenAI's open weight foundation models became available on Amazon Bedrock, AWS's managed model service. OpenAI has quickly become one of the most popular publicly available model providers on Bedrock, with thousands of customers using its models. Named customers include Bystreet, Comscore, Peloton, Thomson Reuters, Triomics, and Verana Health, applying the models to agentic workflows, coding, scientific analysis, mathematical problem-solving, and other tasks.

What the executives said

"Scaling frontier AI requires massive, reliable compute," said Sam Altman, OpenAI co-founder and CEO. "Our partnership with AWS strengthens the broad compute ecosystem that will power this next era and bring advanced AI to everyone."

Matt Garman, CEO of AWS, framed the deal as proof of AWS's position in the AI infrastructure market. "As OpenAI continues to push the boundaries of what's possible, AWS's best-in-class infrastructure will serve as a backbone for their AI ambitions," Garman said. "The breadth and immediate availability of optimized compute demonstrates why AWS is uniquely positioned to support OpenAI's vast AI workloads."

The compute race, in numbers

The figures in the announcement sketch the scale of what frontier AI now requires. Hundreds of thousands of NVIDIA GPUs, clustered into Amazon EC2 UltraServers. A path to tens of millions of CPUs for agentic workloads. A $38 billion commitment spread over seven years of continued growth. Clusters exceeding 500,000 chips as a demonstrated capability.

For context on why the numbers matter: the rapid advancement of AI technology has created demand for computing power that the announcement characterizes as unprecedented. Frontier model providers are not competing only on algorithms and talent — they are competing on access to silicon, networking, and the operational expertise to run enormous clusters reliably. AWS's pitch is that its cloud infrastructure leadership, combined with OpenAI's generative AI work, will help millions of users continue to get value from ChatGPT.

The choice of hardware is specific. The GB200 and GB300 are NVIDIA's latest-generation GPUs, and clustering them on a single low-latency network via EC2 UltraServers is what allows OpenAI to run workloads with what AWS calls optimal performance. The design supports both inference and training, meaning the same infrastructure can serve ChatGPT traffic today and be repurposed for training next-generation models tomorrow.

The Bedrock connection

The partnership builds on groundwork laid earlier this year, when OpenAI open weight foundation models landed on Amazon Bedrock. That move brought OpenAI's models to millions of AWS customers and, by the companies' account, made OpenAI one of the fastest-rising model providers on the platform.

The named Bedrock customers illustrate the spread of OpenAI's models beyond consumer chat. Peloton, Thomson Reuters, and Comscore sit alongside healthcare and life-science firms Triomics and Verana Health and Bystreet, all using OpenAI models for work that ranges from agent-driven workflows to mathematical problem-solving. The new compute deal tightens that relationship at the infrastructure layer, making AWS simultaneously a distribution channel and a supplier for OpenAI.

What comes next

Deployment starts now. OpenAI is accessing AWS compute immediately under the agreement, with the full capacity target set for before the end of 2026 and room to grow into 2027 and beyond. The seven-year horizon of the commitment suggests both companies expect compute demand to keep climbing rather than plateau — and with the ability to scale to tens of millions of CPUs for agentic workloads, the partnership is sized for an era in which AI agents, not just chat interfaces, drive the bulk of inference demand.