AWS and NVIDIA Plan 2 Million Additional GPUs for 2027–2028 AI Expansion

AWS and NVIDIA announce a major future expansion of AI compute

AWS and NVIDIA announced on August 26, 2026 that they plan to deploy 2 million additional NVIDIA GPUs across AWS global infrastructure during 2027–2028. The companies say the expansion will include NVIDIA Blackwell Ultra, Rubin and Rubin Ultra systems and will extend their collaboration across CPUs, networking, memory, open models, data processing and robotics.

The status matters: this is an announced future deployment plan, not 2 million GPUs already installed or generally available today. The companies describe the capacity as a 2027–2028 expansion on top of AWS's previously announced plans to add more than 1 million NVIDIA GPUs starting in 2026.

The GPU plan spans Blackwell Ultra through Rubin Ultra

NVIDIA says the additional capacity will be deployed across AWS Global Infrastructure, including AI-factory environments, for workloads such as agentic AI, scientific computing, enterprise automation and physical AI. AWS also plans to expand current Blackwell capacity, including RTX PRO 4500 Blackwell Server Edition acceleration for Amazon EC2 G7 instances.

NVIDIA states that G7 delivers 4.6× AI inference performance and 2.1× graphics performance compared with the previous G6 generation. Those are vendor-reported comparisons for the specified EC2 generation, not universal performance gains across arbitrary AI workloads.

Vera CPUs are planned for AWS

The partnership also includes work to bring NVIDIA Vera CPU-based infrastructure to AWS. Vera is positioned as high-performance CPU compute designed to sit alongside accelerated AI infrastructure, giving customers another option for agentic workloads that need substantial CPU resources in addition to GPUs or custom accelerators.

Again, the announcement describes this as work the companies are undertaking; it should not be interpreted as evidence that Vera CPU instances are already broadly available in AWS regions.

NVLink Fusion expands toward NVIDIA custom high-bandwidth memory

AWS previously announced support for NVIDIA NVLink Fusion in future Trainium systems. The new agreement extends that collaboration toward NVIDIA custom high-bandwidth memory, or NVHBM, through Amazon's Annapurna Labs and memory suppliers. The goal is to let future Trainium-based infrastructure tap NVIDIA memory and scale-up technologies while operating in a common rack-scale architecture with GPUs.

This is strategically important because AI infrastructure competition is no longer only about accelerator chips. Memory bandwidth, scale-up interconnects, rack design, networking and software increasingly determine how efficiently large models can train and serve at scale.

Federal AI factories and 100,000-GPU plans

AWS and NVIDIA also say they plan to build AI factories for the U.S. government, including 100,000 GPUs on secure AWS infrastructure for federal and national-security workloads. NVIDIA says the collaboration is intended to support workloads classified at Impact Level 6 and above.

This element is a planned government-infrastructure program rather than a general commercial-region feature, and the announcement does not mean all of that capacity is operational today.

Data processing, vector indexing and open models are part of the same stack

The expanded partnership goes beyond raw GPU capacity. NVIDIA's Nemotron open models continue to be supported through Amazon Bedrock and SageMaker. The companies are also working on GPU-accelerated data processing with cuDF on Amazon EMR and GPU-accelerated vector indexing on Amazon OpenSearch Service using NVIDIA cuVS.

NVIDIA reports up to 3.7× faster EMR processing with 30% better price performance for the cited configurations, and up to 9× faster vector indexing at one-quarter the cost for the cited OpenSearch setup. These are vendor-reported workload-specific results and should not be generalized to unrelated deployments without testing.

Amazon Robotics deepens its NVIDIA physical-AI work

The announcement also expands collaboration between Amazon Robotics and NVIDIA's physical-AI stack, including Jetson, Omniverse and Isaac technologies. The stated areas include simulation, synthetic-data generation, robot training, route optimization, functional safety and real-to-sim validation on GPU-accelerated EC2 infrastructure.

That makes the agreement broader than a cloud-capacity purchase: it links future data-center scale to model serving, data pipelines, custom silicon, government AI and warehouse robotics.

Why this matters

A planned addition of 2 million GPUs is a major signal about how much compute AWS and NVIDIA expect future agentic, multimodal and physical-AI workloads to consume. It also illustrates the increasingly heterogeneous shape of AI infrastructure: NVIDIA GPUs, Vera CPUs, AWS Trainium, NVLink Fusion, custom high-bandwidth memory, Nitro, EFA and accelerated data services are being designed to work as a coordinated stack.

The correct status is: AWS and NVIDIA announced the additional 2 million-GPU deployment for 2027–2028. The related Vera, NVHBM, federal AI-factory and robotics work includes future commitments and integrations, so none should be described as fully deployed today unless separately verified.