NVIDIA and AWS Plan 2 Million More GPUs for AI

NVIDIA and AWS plan to add 2 million more NVIDIA GPUs across AWS infrastructure in 2027 and 2028, expanding a 16-year collaboration from cloud instances into CPUs, networking, open models, data processing, and robotics. The announcement is a capacity bet — and a signal that the next phase of AI competition will be decided by complete infrastructure stacks, not chips alone.

Two Million GPUs Is a Capacity Commitment

In its August 26 announcement, NVIDIA said AWS will deploy the additional GPUs across its global infrastructure, including AI factories. AI Magazine’s coverage describes the move as a response to demand moving from pilots into production across agentic AI, scientific discovery, enterprise automation, and physical AI.

The scale is notable, but the timing matters just as much. The companies had already announced plans to add more than 1 million NVIDIA GPUs beginning in 2026; NVIDIA now says demand exceeded that expectation. This is not a promise that every GPU will be online at once or fully utilised. It is a procurement and deployment plan designed to give AWS enough capacity for customers whose workloads are becoming larger, more continuous, and more inference-heavy.

The Partnership Is Moving Up and Down the Stack

The deal extends beyond NVIDIA’s core accelerators. AWS will work to bring NVIDIA Vera CPU-based infrastructure to its cloud, supporting high-performance CPU work around agentic systems. The companies will also expand their earlier collaboration across networking, software, and deployment.

That stack approach is strategically important for NVIDIA’s data-centre business. The companies are also extending NVLink Fusion and custom high-bandwidth memory. Buyers care about system-level results: memory movement, network latency, security, orchestration, and the cost of serving a model after training. A stronger position across those layers can make NVIDIA’s platform harder to displace, even when cloud providers continue developing custom silicon.

Open Models and Networking Become Competitive Levers

The expanded collaboration supports NVIDIA Nemotron open models through Amazon Bedrock and Amazon SageMaker. Nemotron gives developers another model family to evaluate, while Amazon Bedrock and SageMaker provide managed routes into enterprise applications.

This model choice is more than a customer-convenience feature. Open and third-party models can reduce dependence on a single provider and let teams tune economics by workload. For AWS, model choice helps keep customers on its infrastructure even when the model layer changes. For NVIDIA, it extends the platform story from accelerated training into software and inference decisions that determine recurring compute demand.

Security and Government Demand Add a Second Track

AWS and NVIDIA plan to build AI factories for the U.S. government, including 100,000 GPUs on secure AWS infrastructure for federal and national-security workloads. The announcement also names the AWS Nitro System and Elastic Fabric Adapter as parts of the integration for security and reliability. Nitro is therefore not a side detail: trusted isolation and high-speed communication become procurement requirements when AI runs inside sensitive environments.

Government demand can provide a durable anchor for infrastructure, but it also raises the standard for auditability, access controls, supply assurance, and operational resilience. Capacity alone will not turn a general cloud platform into a national-security system; the surrounding controls and accreditation work carry much of the value.

Physical AI Makes the Roadmap Broader

The companies are also extending the relationship into robotics through Amazon Robotics and NVIDIA’s physical AI platform, spanning simulation, synthetic data, robot training, route optimisation, and real-to-sim validation. Our coverage of NVIDIA Omniverse and physical AI shows why this matters: robots require a loop connecting digital environments, perception, planning, and real-world feedback.

For investors and enterprise buyers, the key question is utilisation. AI infrastructure financing and large-scale compute investment can fund expansion, but the economics depend on customers turning capacity into production workloads. Competitors are expanding too, while model specialisation may change where that capacity is needed. NVIDIA and AWS have made a very large supply-side bet; deployment quality and sustained demand will decide its return.