Amazon and OpenAI announced a multi-year strategic agreement pairing Amazon's cloud infrastructure and proprietary chips with OpenAI's models and enterprise deployment software. Under the financial terms, Amazon will invest $50 billion in OpenAI, starting with an initial $15 billion outlay and an additional $35 billion planned in subsequent months once specified conditions are met.

Alongside the capital injection, the companies are expanding an existing $38 billion commercial agreement by $100 billion over eight years. Under this expansion, OpenAI has committed to consume approximately 2 gigawatts of AWS computing capacity powered by Amazon's custom Trainium silicon.

Runtime Integration and Cloud Distribution

For engineering teams operating on AWS, the partnership outlines software integrations across Amazon Bedrock and OpenAI's product lines:

  • Exclusive third-party distribution for OpenAI Frontier: AWS will serve as the exclusive third-party cloud distributor for Frontier, OpenAI's enterprise platform designed to build, deploy, and manage multi-agent systems with unified governance and security.
  • Stateful Runtime Environment on Amazon Bedrock: The two companies are co-developing a stateful runtime powered by OpenAI models. According to Amazon, the environment is engineered to preserve context, remember prior workflow state, access compute, and interface with software tools and data sources. Amazon stated it expects the runtime to launch in the coming months, integrated directly with Amazon Bedrock AgentCore and AWS infrastructure services.
  • Custom internal models: OpenAI and Amazon will also develop customized models for Amazon's internal developers to power customer-facing applications, complementing the company's existing Nova model family.

Staged Hardware Delivery and Trainium Capacity

The 2 GW compute agreement ties OpenAI's operational scaling directly to Amazon's in-house silicon roadmap across both Trainium3 and next-generation Trainium4 processors. The capacity is planned to support the Stateful Runtime Environment, Frontier deployments, and other advanced workloads.

However, the compute footprint depends on multi-year hardware staging. While the agreement encompasses Trainium3, Amazon stated that Trainium4 is not expected to begin delivery until 2027. According to Amazon, Trainium4 is planned to deliver higher FP4 compute performance, higher memory bandwidth, and expanded high-bandwidth memory (HBM) capacity. The 2 GW figure is a multi-year capacity commitment spanning both generations; it does not establish how much capacity was already operational when the agreement was announced.