AWS and NVIDIA Plan 2 Million More GPUs for Agentic and Physical AI

AWS and NVIDIA expand their AI infrastructure roadmap

AWS and NVIDIA announced on August 26, 2026 that they plan to deploy 2 million additional NVIDIA GPUs across AWS's global infrastructure during 2027 and 2028.

The agreement extends well beyond raw GPU capacity. The companies also plan deeper integration around NVIDIA Vera CPUs, NVLink Fusion and custom high-bandwidth memory, secure U.S. government AI factories, Nemotron open models, GPU-accelerated data processing and Amazon Robotics.

The timing matters: the 2 million GPU deployment is a future infrastructure plan, not capacity already installed today.

The additional fleet spans Blackwell Ultra, Rubin and Rubin Ultra

NVIDIA says AWS plans to deploy additional Blackwell Ultra, Rubin and Rubin Ultra GPUs across its global infrastructure, including future AI factories.

AWS had already announced plans at GTC 2026 to add more than 1 million NVIDIA GPUs starting in 2026. The new announcement adds another 2 million GPUs for the 2027–2028 period.

NVIDIA also says AWS will expand Blackwell capacity with RTX PRO 4500 Blackwell Server Edition GPUs for Amazon EC2 G7 instances.

Vera CPU and NVHBM deepen the hardware integration

AWS and NVIDIA are working to bring NVIDIA Vera CPU-based infrastructure to AWS for agentic AI workloads that need strong CPU performance alongside accelerators.

The companies are also extending NVIDIA NVLink Fusion work around future Trainium systems to include NVIDIA custom high-bandwidth memory (NVHBM). The idea is to give Amazon's Annapurna Labs another route to combine Trainium accelerators, NVIDIA interconnect technology and custom memory within rack-scale AI systems.

These are planned integrations, not proof that every listed configuration is broadly available in AWS today.

U.S. government AI factories are part of the plan

The collaboration also includes plans 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 target includes workloads classified at Impact Level 6 and above.

This is a significant public-sector infrastructure commitment, but the announcement describes a planned deployment. It should not be reported as though 100,000 federal GPUs are already operational.

Some integrations are already available today

The press release also describes capabilities that are already in market.

NVIDIA's Nemotron open models are available through Amazon Bedrock and SageMaker. AWS and NVIDIA also say they are collaborating on GPU-accelerated data processing with cuDF on Amazon EMR and GPU-accelerated vector indexing for Amazon OpenSearch.

NVIDIA reports up to 3.7x faster processing with 30% better price performance for the described EMR configuration, and up to 9x faster vector index building at one quarter of the cost for the described OpenSearch approach.

Those are company-reported performance claims tied to specific configurations, not universal guarantees for all workloads.

Amazon Robotics is adopting more of NVIDIA's physical AI stack

Amazon Robotics is collaborating with NVIDIA around Jetson, Omniverse and Isaac technologies for simulation, synthetic-data generation, robot training, route optimization, functional safety and real-to-sim validation.

That makes the partnership relevant not only to foundation-model infrastructure but also to warehouse automation and physical AI.

Why this matters for the AI infrastructure market

The scale of the announcement shows how cloud providers and accelerator companies are planning several hardware generations ahead.

The deal spans GPUs, CPUs, memory, networking, open models, data engines, government infrastructure and robotics. It therefore illustrates the shift from buying individual accelerators toward building vertically integrated AI factories.

Planned for 2027–2028: 2 million additional NVIDIA GPUs across AWS and associated next-generation infrastructure.

Planned public-sector expansion: 100,000 GPUs for U.S. government AI factories.

Already available or actively integrated today: existing NVIDIA GPU and Trainium EC2 infrastructure, Nemotron access through Bedrock and SageMaker, and current data-processing and vector-indexing integrations described by the companies.

Keeping those categories separate is essential because the headline capacity is a forward commitment rather than deployed compute.