AWS and NVIDIA plan 2 million more GPUs as their AI stack expands beyond accelerators

On August 26, 2026, AWS and NVIDIA announced that AWS plans to deploy two million additional NVIDIA GPUs across its global infrastructure in 2027 and 2028. The capacity will span Blackwell Ultra, Rubin and Rubin Ultra generations and comes on top of AWS's earlier plan to add more than one million NVIDIA GPUs beginning in 2026. The companies are also working to bring NVIDIA Vera CPU-based infrastructure to AWS, extend NVLink Fusion with NVIDIA high-bandwidth memory, integrate the NVIDIA platform with Nitro and Elastic Fabric Adapter, accelerate EMR and OpenSearch workloads with cuDF and cuVS, continue Nemotron availability through Bedrock and SageMaker, and build U.S. government AI factories including 100,000 GPUs on secure AWS infrastructure.

For builders, this is a capacity and architecture signal rather than a benchmark story. AWS is committing to another very large block of NVIDIA capacity while making more of the surrounding NVIDIA stack native to its cloud. That can broaden future managed access to Rubin-era infrastructure, reduce friction for heterogeneous CPU/GPU workloads, and make NVIDIA-accelerated data processing and vector search more available inside ordinary AWS services. It also reinforces a strategic concentration risk: teams that choose AWS for managed AI can gain unusually deep NVIDIA integration while becoming more exposed to the roadmap and pricing decisions of both vendors.

AWS says demand exceeded the expectations behind its earlier plan for more than one million NVIDIA GPUs. The new agreement adds two million more GPUs in 2027–2028 across Blackwell Ultra, Rubin and Rubin Ultra systems. The companies have not published customer pricing or a region-by-region availability schedule, so the commitment establishes future capacity direction rather than immediately purchasable inventory.

AWS and NVIDIA are working to bring Vera CPU-based infrastructure to AWS. NVIDIA positions Vera for CPU-heavy parts of agentic AI and reinforcement-learning systems such as code execution, tool use, sandboxing, analytics and orchestration. AWS says it has already received its first Vera CPU server and Vera Rubin GPU, but the announcement does not give a general-availability date for customer instances.

The expansion includes NVLink Fusion work, Nitro and EFA integration, Nemotron open models in Bedrock and SageMaker, and NVIDIA cuDF/cuVS acceleration for EMR and OpenSearch. The practical direction is a more integrated stack in which compute, networking, model access and data preparation are increasingly co-designed rather than exposed as isolated products.

Amazon Robotics is adopting NVIDIA's physical-AI platform, and the companies say they will build AI factories for U.S. government workloads including a 100,000-GPU secure AWS deployment. Those projects do not directly change ordinary developer access today, but they add large competing demand pools for the same accelerator ecosystem and show where AWS expects future AI infrastructure growth.