August 28, 2026 / SemiMedia / — Amazon Web Services and NVIDIA have expanded their AI infrastructure partnership, with AWS planning to deploy 2 million additional NVIDIA GPUs across its global infrastructure during 2027 and 2028.

The deployment will include NVIDIA Blackwell Ultra, Rubin and Rubin Ultra GPUs for AI model training and inference, scientific computing, enterprise automation, agentic AI and robotics workloads.

AWS announced in March that it would begin adding more than 1 million NVIDIA GPUs in 2026. The latest expansion brings the company’s publicly announced NVIDIA GPU deployment plans to more than 3 million units between 2026 and 2028.

AWS and NVIDIA will also deepen their collaboration across CPUs, high-speed interconnects, data processing, open models and robotics. AWS plans to introduce infrastructure based on NVIDIA Vera CPUs and further integrate NVLink Fusion with the AWS Nitro System and Elastic Fabric Adapter.

Neither company disclosed the financial terms of the 2 million-GPU deployment. The project covers multiple GPU architectures, server systems, networking equipment and supporting data center infrastructure, making retail chip prices an unreliable measure of its total value.

Amazon continues to invest in Annapurna Labs, which develops Graviton CPUs and Trainium AI accelerators for AWS. Future AWS custom silicon will also be more closely integrated with NVIDIA’s networking and memory technologies.

The expansion shows that AWS is pursuing custom accelerators and NVIDIA GPUs in parallel. Trainium can serve workloads optimized for AWS infrastructure and operating costs, while NVIDIA GPUs remain central to mainstream AI frameworks and large-scale model deployment.