Amazon, OpenAI ink $38b cloud deal to supply Nvidia GPUs

OpenAI and Amazon Web Services (AWS) have signed a multi-year agreement worth US$38 billion that will give OpenAI access to AWS’s cloud infrastructure for advanced AI workloads.

The deal allows OpenAI to use AWS compute resources, including hundreds of thousands of Nvidia GPUs and the potential to scale to tens of millions of CPUs, with deployment set to complete by the end of 2026, and options to expand further.

AWS will provide its EC2 UltraServers to support OpenAI’s generative AI workloads, such as training and running large models like ChatGPT.

The partnership is intended to help OpenAI scale its AI systems as demand for computing power rises.

OpenAI’s models have already been made available earlier this year on Amazon Bedrock, AWS’s platform for foundation models, where it is used by customers in various sectors.

Both companies said the partnership will support continued development and deployment of advanced AI technology.

🔗 Source: Amazon

🧠 Food for thought

Implications, context, and why it matters.

OpenAI shifts to multi-cloud for portability and less lock-in

OpenAI signed an AWS deal after the October 2025 restructuring removed Microsoft’s right of first refusal 1. It can spread compute across AWS ($38B) 2 and Microsoft Azure ($250B commitment) 1. Oracle Cloud Infrastructure ($300B reported) and Google Cloud join the list 2. Cloud contracts near $600B against about $13B revenue 3.
The restructuring lets OpenAI work with third parties to build AI products and ship open‑weight models (downloadable models whose parameters are publicly available) 2. Amazon added OpenAI’s first open‑weight models to Amazon Bedrock (AWS’s managed service for foundation models) and SageMaker (AWS’s machine learning platform) under an open‑source license, bypassing prior API exclusivity 1.
Microsoft keeps IP rights to OpenAI models and products through 2032 1. Only Azure among major clouds can offer OpenAI tech through the Azure OpenAI Service 1. OpenAI chooses where to run based on availability, performance, and cost.

Platform and infrastructure leaders see an opening in multi-cloud GPU orchestration and workload management

Enterprises will copy this pattern. Rafay is a Kubernetes management platform for operating applications across data centers and clouds, while Run:ai is a GPU orchestration platform 4 5. These platforms enable GPU‑as‑a‑service (on‑demand GPU capacity delivered as a managed service) and placement based on availability, cost, and performance 4 5.
Most AI teams mix spot instances (lower-cost, interruptible cloud capacity) for training, dedicated capacity for inference, and development instances for experimentation 6. Automation to handle interruptions and placement is missing. Platforms that blend spot capacity with orchestration cut costs by up to 90% 6.