Cloud giant Amazon Web Services and chipmaker Nvidia have signed a new large-scale agreement, under which an additional two million graphics processors will appear in AWS's global infrastructure in 2027–2028. This is not just another capacity expansion, but a strategic response to explosive demand from the agentic and physical AI segments, which require fundamentally different computing architectures.

Millions of accelerators: from training to robotics

The new tranche of deliveries supplements the previously announced plan at GTC 2026 to deploy over 1 million GPUs starting this year. Thus, the total fleet of Nvidia accelerators in AWS will reach multi-million scale. These resources will be used not only for classic model training and inference, but also for fundamentally new workloads—autonomous agents executing multi-step scenarios and robotic systems operating in real physical environments.

Hardware evolution: Blackwell, Rubin, and custom solutions

As part of the partnership, AWS will integrate several generations of Nvidia platforms into its clusters, including Blackwell Ultra and the upcoming Rubin and Rubin Ultra. A key element is the use of the NVIDIA Vera CPU, which is designed specifically to coordinate tools, execute code, and run simulations in agentic scenarios. This offloads GPUs from routine tasks and improves overall computing efficiency.

A separate focus is deep integration with AWS Trainium chips. Amazon's division, Annapurna Labs, together with Nvidia, is expanding support for NVLink Fusion technology, including high-speed memory and scalable rack-level interconnects. This hybrid approach gives customers flexibility in choosing the optimal architecture for specific tasks.

New instances and performance

AWS has also introduced EC2 G7 cloud instances based on NVIDIA RTX PRO 4500 Blackwell Server Edition GPUs. This is the first such offering among major cloud providers. According to my data, compared to the previous G6 generation, the new instances deliver:

  • up to 4.6 times higher AI inference efficiency;
  • up to 2.1 times graphics performance improvement.

At the same time, the infrastructure remains compatible with AWS Nitro and Elastic Fabric Adapter networking solutions, as well as NVIDIA Spectrum technologies.

Government orders and physical AI

A separate part of the agreement concerns U.S. government and defense customers. AWS and Nvidia will create specialized AI factories for processing classified data with an Impact Level 6 security level and above, which will receive about 100,000 GPUs. This signals that AI infrastructure is becoming a critical element of national security.

The partnership extends far beyond data centers. Amazon Robotics is standardizing the use of Nvidia's physical AI platform—Jetson, Omniverse, and Isaac. These technologies will be used for warehouse automation, synthetic data generation, and robot behavior simulation. NVIDIA Nemotron open models will remain available through Amazon Bedrock and SageMaker.

My view: This deal finally cements a paradigm shift in the industry. Competition is moving from simply increasing GPU counts to building holistic computing factories where accelerators, CPUs, memory, and networking work as a single organism. Companies that can offer such an integrated approach will gain a decisive advantage in the race for dominance in the agentic AI era. Investors and developers should closely watch how this integration affects computing costs and capacity availability over the next two to three years.