Amazon Web Services (AWS) and Nvidia have agreed to deploy an additional two million GPUs in AWS’s cloud infrastructure during 2027–2028.
The companies are expanding their partnership amid growing demand for computing power for agentic and physical AI.
In Addition to One Million GPUs
The new plan involves supplying over 2 million GPUs for AWS’s global infrastructure in 2027 and 2028. This is in addition to previous agreements: at GTC 2026, AWS announced plans to install over 1 million Nvidia GPUs starting in 2026.
This expansion involves a multi-million GPU fleet that AWS intends to use for training and inference of models, as well as new workloads related to autonomous AI agents and robots.
From Blackwell to Rubin
AWS will integrate several generations of Nvidia platforms into its AI clusters, including Blackwell Ultra, Rubin, and Rubin Ultra. The companies are also expanding collaboration on custom chips and memory.
Specifically, AWS is preparing infrastructure based on the NVIDIA Vera CPU, designed for tasks involving AI agents: tool coordination, code execution, data processing, and simulations. This aims to complement GPUs and enable more efficient use of accelerators in multi-step agent scenarios.
Another focus is integrating Nvidia technologies with AWS Trainium chips. Annapurna Labs, an Amazon subsidiary, and Nvidia are expanding previously announced NVLink Fusion support, including the use of Nvidia’s high-speed memory and scalable interconnects in rack-level configurations.
AWS to Receive New GPU Instances
AWS also introduced EC2 G7 — cloud virtual servers (instances) with NVIDIA RTX PRO 4500 Blackwell Server Edition GPUs. The company describes them as the first of their kind among major cloud providers.
According to Nvidia, G7 offers improvements over the previous G6 generation, including:
- AI inference efficiency up to 4.6 times;
- graphics performance up to 2.1 times.
Meanwhile, GPU instances and Trainium-based systems will continue to use AWS Nitro network infrastructure and Elastic Fabric Adapter alongside NVIDIA Spectrum networking technologies.
100,000 GPUs for the US Government
A separate aspect of the partnership involves government and defense clients.
AWS and Nvidia plan to create specialized AI factories for the US government, which will receive about 100,000 Nvidia GPUs. The infrastructure will be designed to handle classified data and security requirements at Impact Level 6 and above.
Amazon Bets on Physical AI
The partnership also extends into the physical realm. Amazon Robotics is standardizing the use of Nvidia’s platform for physical AI, including Jetson, Omniverse, and Isaac.
These technologies will be used for warehouse automation, synthetic data creation, and modeling the behavior of robotic systems.
Simultaneously, open models NVIDIA Nemotron will remain available through Amazon Bedrock and SageMaker — both as managed services and for custom deployment.
Nvidia noted that the deal with AWS underscores a trend in AI infrastructure competition: it’s no longer just about the number of GPUs, but about building entire computing factories that integrate accelerators, CPUs, memory, networks, software, and specialized systems for agents and robots.
In August, SpaceXAI announced an expansion of its use of Nvidia infrastructure for developing agentic systems and Grok. Plans include deploying Vera processors and the Vera Rubin platform in space.
