Building on our August 27 coverage, new reports fill in the plan and timeline for AWS and NVIDIA's multi-year GPU build-out inside Amazon's cloud.
According to multiple outlets, AWS will add roughly 2 million NVIDIA GPUs across its data centers during 2027 and 2028. The rollout reportedly includes Blackwell Ultra, Rubin and Rubin Ultra accelerators, plus dedicated, isolated "AI factory" deployments of about 100,000 GPUs for sensitive government workloads.
Why this matters: this is not a small capacity bump. Adding GPUs at this scale materially increases the raw compute available for training very large models and for running high-volume inference. That makes previously expensive workloads, large-scale simulation, distributed training, enterprise automation and more ambitious agentic systems, practical inside the cloud.
How it works, at a glance: think of AWS as a highway operator. Today many teams rent lanes from external providers. Installing millions of GPUs is like building new lanes and entire terminals designed for heavy AI traffic. The hardware is specialized, and training big models depends on many of these machines working together with fast networking.
What changes now: companies and researchers will have more cloud-native options to build and scale advanced models without relying solely on other vendors' hosted APIs. It also deepens AWS's reliance on NVIDIA even as Amazon develops its own chips. A practical limit to keep in mind: this is data-center infrastructure, not something a team can run on a workstation; it requires large-scale GPU fleets and specialized networking.
What's next: watch whether AWS exposes this capacity as affordable, self-serve products, or keeps most of it for its own services and government contracts. The coming GA dates, benchmarks and any regulatory scrutiny will tell us how widely this compute actually gets used.