AWS Adds 2 Million Nvidia GPUs—Its AI Bet Triples in Five Months

AWS and Nvidia unveiled their August 26 plan to deploy 2 million additional GPUs across AWS’s global infrastructure in 2027 and 2028. The future capacity will use Nvidia Blackwell Ultra, Rubin and Rubin Ultra products; the announcement does not mean those accelerators have already been delivered or entered service.

The expansion follows AWS’s earlier commitment to add more than 1 million Nvidia GPUs across its global cloud regions starting in 2026. The two programs therefore have different deployment windows, and the later commitment supplements rather than replaces the first.

The capacity ledger shows a threefold expansion—with a caveat

The earlier plan established a lower bound, not an exact order size: AWS used “more than 1 million.” The new increment is 2 million, so adding the published figures produces a combined lower bound above 3 million GPUs if both programs proceed as described.

That arithmetic supports the headline’s “triples” shorthand, but it does not establish an exact procurement ratio. Because the original commitment could be higher than its stated minimum, the precise baseline and exact multiple remain unknown.

TechCrunch’s same-day account of the threefold expansion placed the new commitment roughly five months after the first and noted that neither company disclosed financial terms. Calling the arrangement an “order” is convenient shorthand; the companies’ formal language is that AWS plans to deploy the capacity.

The additional GPUs are scheduled for 2027–2028, not immediate delivery

The two-year window is a deployment period rather than a single arrival date. AWS has not published a region-by-region schedule, the allocation among Blackwell Ultra, Rubin and Rubin Ultra, or dates when customer-facing EC2 capacity using the additional fleet will become available.

Deployment at this scale can involve manufacturing, delivery, data-center preparation, cluster integration and service validation. The public commitment establishes the intended volume and broad timing, but it does not show that every GPU will arrive at the beginning of 2027 or become available simultaneously across AWS regions.

Nvidia’s outline of the collaboration likewise describes a planned deployment across global AWS infrastructure. It does not provide shipment milestones, procurement structure or ownership details for the installed systems.

The commitment extends beyond the GPU count

The expansion is also a systems-integration program involving processors, memory links and networking. AWS and Nvidia plan to bring Vera CPU-based infrastructure to the cloud provider, extend NVLink Fusion work with Nvidia high-bandwidth memory and continue developing Spectrum networking for large GPU clusters.

The architecture connects Nvidia components with AWS technology. Nvidia GPU-based and AWS Trainium-based EC2 instances use the AWS Nitro System and Elastic Fabric Adapter, while Nvidia and Amazon’s Annapurna Labs are extending their work on rack-scale systems. The companies have not specified how many Vera CPUs will be deployed or which configurations and regions will receive them.

At the software layer, the collaboration covers Nemotron open models through Amazon Bedrock and SageMaker, GPU-accelerated data processing on Amazon EMR with cuDF, and vector indexing on Amazon OpenSearch with cuVS. These product integrations have their own availability and performance conditions; they are not evidence that the newly planned GPU capacity is already online.

Physical AI is another part of the agreement. An Associated Press report on the AWS–Nvidia plan independently recorded both the additional capacity and the extension of Nvidia technology into Amazon’s warehouse-robotics operations.

Amazon Robotics is set to use Nvidia’s Jetson, Omniverse and Isaac technologies for work spanning simulation, synthetic-data generation, robot training, route optimization, functional safety and validation. The broader program also includes planned AI factories for the U.S. government, with 100,000 GPUs on secure AWS infrastructure for federal and national-security workloads.

Price, regional allocation and customer availability remain unknown

No purchase price, payment schedule, unit cost or expected revenue contribution has been made public for the additional capacity. Applying a retail GPU price to the headline quantity would not yield a reliable deal value because large infrastructure arrangements can combine accelerators, servers, networking, memory, software and support under negotiated terms.

The available disclosures also do not establish how the government allocation relates to the broader deployment, how capacity will be divided among GPU generations, or which AWS instance families will expose it to customers. Those details determine when the commitment turns into usable cloud capacity rather than planned infrastructure.

As of August 28, the confirmed story is therefore a multiyear expansion: the earlier program begins in 2026, while the additional fleet is planned for 2027–2028. Shipment milestones, named regions, instance launches, model allocations and any later financial disclosure will provide the next evidence of execution; until then, the combined figure describes planned additions, not AWS’s current installed inventory.

Also read:

• Nvidia Is Pulling Off the Most Sophisticated Financial Loop in Tech History
• Is AI a Bubble? Five Indicators Matter More Than the Headline
• Peter Thiel Leads $140 Million Bet on Floating AI Data Centers Powered by Ocean Waves