AWS plans two million additional NVIDIA GPUs across 2027 and 2028

AWS and NVIDIA have outlined a large expansion of accelerated computing capacity. AWS plans to deploy two million additional NVIDIA GPUs during 2027 and 2028 across its global infrastructure. The named generations are Blackwell Ultra, Rubin and Rubin Ultra.

The word “additional” is important because this capacity sits on top of earlier plans rather than describing the provider's total installed fleet. It is also a forward deployment schedule, not two million systems already available to rent. Customers should separate the strategic capacity commitment from region-by-region instance availability and delivery dates.

The collaboration extends beyond GPUs. AWS and NVIDIA plan Vera CPU-based infrastructure, continued support for Nemotron open models through Amazon Bedrock and SageMaker, and further integration across networking, data processing and robotics. The companies also describe secure infrastructure containing 100,000 GPUs for U.S. government workloads.

This breadth shows an effort to sell a system stack: compute, CPU, network, software and managed model access. Each layer has a different readiness path. A capacity plan can be firm at the partnership level while individual services remain subject to previews, regional launches and customer qualification.

The central number is two million additional GPUs across 2027-2028. Within the government work, the companies identify 100,000 GPUs on secure AWS infrastructure. For a more immediate instance comparison, NVIDIA says EC2 G7 with RTX PRO 4500 delivers 4.6x AI inference performance and 2.1x graphics performance versus G6.

Those performance ratios are vendor claims for specified instance generations. They should not be applied to every model or graphics workload. Buyers need price, availability, memory, network and software details to translate a speed ratio into cost per completed job.

Capacity at this scale can influence where model developers train and serve workloads. A cloud provider with a large multi-generation GPU pipeline can offer customers migration paths without requiring them to build data centers. The inclusion of CPUs and managed-model services also aims to keep more of the workflow inside one cloud.

That concentration has tradeoffs. It can simplify procurement and integration, while increasing dependence on one provider's regions, pricing and service interfaces. The announcement demonstrates planned supply and technical collaboration; it does not establish future utilization, customer allocation or the total energy and infrastructure required to operate the fleet.

Frontier labs, enterprises, governments and robotics developers are the stated demand groups. Smaller customers may feel the effect indirectly through instance availability or managed services rather than through dedicated clusters. Hardware generations also matter: Blackwell Ultra, Rubin and Rubin Ultra will not present identical performance, memory or software characteristics.

Teams planning multi-year capacity should map workloads to the generation they expect to use and retain portability where practical. A reservation or architectural decision made before detailed instance specifications arrive carries schedule and performance risk.

The release does not provide a region-by-region delivery calendar, instance prices, power requirements or customer allocation rules for the two million GPUs. It also does not state how much capacity will be available in each of 2027 and 2028.

Government infrastructure is described as secure, but the public announcement does not detail accreditation boundaries or workload eligibility. The confirmed facts are the planned quantities, generations and collaboration areas. Actual service availability must be verified from AWS product pages as individual offerings reach production.