Amazon Web Services and NVIDIA said on August 26, 2026 that they will deploy 2 million additional NVIDIA GPUs across AWS’s global infrastructure in 2027 to 2028. The extra capacity is meant to cover agentic AI, scientific discovery, enterprise automation, and robotics after demand ran past the plan AWS announced at NVIDIA GTC 2026 to add more than 1 million GPUs starting this year.

The companies also said they will bring NVIDIA Vera CPU infrastructure to AWS, extend NVLink Fusion with custom NVIDIA high-bandwidth memory for Trainium, and keep Nemotron open models on Amazon Bedrock and SageMaker. A parallel AWS note framed the GPU order as a complement to Amazon’s own silicon, not a replacement for it.

Federal factories and a wider stack

AWS and NVIDIA plan AI factories for the U.S. government, including 100,000 GPUs on secure AWS infrastructure for federal and national-security work at Impact Level 6 and above. They said NVIDIA GPU and Trainium instances will stay on the Nitro System and scale out through Elastic Fabric Adapter.

On the data path, the companies pointed to GPU-accelerated processing on Amazon EMR with the cuDF library, which they said can run up to 3.7 times faster and 30 percent better on price performance than CPU setups. GPU-accelerated vector indexing on Amazon OpenSearch is described as up to 9 times faster at a quarter of the cost. Amazon Robotics is also adopting NVIDIA’s physical AI stack, including Jetson, Omniverse, and Isaac, for warehouse automation.

“NVIDIA and AWS have built one of the great growth engines of the AI era, and demand is running ahead of every forecast.” Jensen Huang, founder and CEO, NVIDIA

AWS CEO Matt Garman said customers want to pick the best tool for a workload and still have the pieces work together. The named GPU generations in the extra 2 million are Blackwell Ultra, Rubin, and Rubin Ultra. AWS also said it will expand Blackwell capacity, including RTX PRO 4500 Blackwell Server Edition GPUs on EC2 G7 instances, and that it is the first major cloud to offer those cards.

Decoded Take

This is a demand letter written as a partnership. AWS already promised more than a million NVIDIA GPUs at GTC. Coming back months later with 2 million more for 2027 to 2028 tells customers the earlier number was a floor, not a ceiling. The interesting part is not the GPU headline. It is Trainium getting NVLink Fusion and custom NVIDIA memory, and a 100,000-GPU federal factory sitting next to the commercial fleet. That pairing lets AWS sell choice (NVIDIA, Trainium, or both in one rack) while locking government work into the same security story. Watch whether the 2027 shipments show up as available instances or as reserved AI-factory clusters that never hit the public price list. If Rubin and Vera land on AWS first, other clouds will have to match the menu. If they slip, the 2 million figure becomes a slide from last August.