Amazon Web Services (AWS) and Nvidia have agreed to add two million graphics processing units (GPUs) to AWS's cloud infrastructure by the years 2027 and 2028.

This expansion of their partnership comes in response to the increasing demand for computing resources dedicated to both agentic and physical artificial intelligence.

Building on Previous Commitments

The new initiative aims to deliver over two million GPUs to AWS's global infrastructure during 2027 and 2028. This is in addition to a prior commitment made at the GTC 2026 event, where AWS announced plans to deploy over one million Nvidia GPUs starting in 2026.

With this, AWS is set to create a multi-million GPU fleet intended for training and inference tasks, as well as supporting new workloads associated with autonomous AI agents and robots.

Integrating Multiple Nvidia Generations

AWS plans to integrate various generations of Nvidia platforms into its AI clusters, including Blackwell Ultra, Rubin, and Rubin Ultra. Additionally, the two companies are enhancing their collaboration on custom chips and memory solutions.

Specifically, AWS is preparing infrastructure based on the NVIDIA Vera CPU, which is tailored for tasks related to AI agents, such as tool coordination, code execution, data processing, and simulations. This will complement the GPUs and improve the efficiency of using accelerators in multi-step agent scenarios.

Another area of focus is the integration of Nvidia technologies with AWS's Trainium chips. Annapurna Labs, a subsidiary of Amazon, and Nvidia are expanding their previously announced support for NVLink Fusion, incorporating the use of Nvidia's high-speed memory and scalable interconnects in rack-level configurations.

New GPU Instances for AWS

AWS has also introduced the EC2 G7, which features cloud virtual servers (instances) equipped with NVIDIA RTX PRO 4500 Blackwell Server Edition GPUs. The company claims these are the first offerings of their kind among major cloud providers.

According to Nvidia, the G7 instances provide significant improvements over the previous G6 generation, including:

  • up to 4.6 times more efficient AI inference;
  • up to 2.1 times greater graphical performance.

GPU instances and systems based on Trainium will continue to leverage AWS's Nitro network infrastructure and Elastic Fabric Adapter, along with Nvidia's Spectrum networking technologies.

100,000 GPUs Allocated for U.S. Government

An additional aspect of the partnership focuses on government and defense clients.

AWS and Nvidia plan to establish specialized AI factories for the U.S. government, which will be equipped with approximately 100,000 Nvidia GPUs. This infrastructure will be designed to handle classified data and meet security requirements at Impact Level 6 and above.

Amazon's Investment in Physical AI

The partnership also extends into the physical realm, with Amazon Robotics standardizing the use of Nvidia platforms for physical AI, including Jetson, Omniverse, and Isaac.

These technologies will be utilized for warehouse automation, synthetic data generation, and simulating robotic system behavior.

In parallel, open models from NVIDIA Nemotron will remain accessible through Amazon Bedrock and SageMaker, available both as managed services and for custom deployment.

Nvidia remarked that the agreement with AWS reflects a trend in the competitive landscape of AI infrastructure, emphasizing that the focus is shifting from merely the number of GPUs to constructing entire computing factories that integrate accelerators, CPUs, memory, networks, software, and specialized systems for agents and robots.

In August, SpaceXAI announced plans to further leverage Nvidia's infrastructure for developing agent systems and Grok, including the deployment of Vera processors and the Vera Rubin platform in space.