Amazon Web Services announced it would deploy an additional 2 million NVIDIA GPUs across its global cloud infrastructure by 2028 expanding an earlier plan to add more than 1 million chips beginning in 2026. The move was presented as a continuation of a collaboration spanning 16 years and was described as targeting global customer demand for accelerated computing.
Last update: 31 August 2026. The expansion aimed to support a wide range of workloads from large language models and agentic AI to robotics simulation and sensitive government deployments. AWS said the program would combine GPU scale with new CPU and interconnect technologies to improve
Scale and performance commitments for 2026–2028
The plan covered deployments through 2027 and 2028 and built on an initial commitment starting in 2026. AWS stated that the hardware mix would include the latest Blackwell Ultra, Rubin and Rubin Ultra GPU families and that the rollout would enable Amazon EC2 G7 instances featuring RTX PRO 4500 Blackwell Server Edition GPUs. Those instances were presented as delivering up to 4.6 times faster inference and 2.1 times stronger graphics performance compared with prior-generation G6 instances. AWS also reserved 100,000 chips for sensitive government and defense computing needs.
Infrastructure integration: Vera CPUs, NVLink Fusion and NVHBM
The announcement described integration of NVIDIA Vera CPUs into AWS offerings to support agentic AI workloads that required sustained CPU throughput alongside accelerated GPU compute. AWS framed Vera CPUs as a host and standalone option to keep inference and training loops fed. The companies also planned to adopt NVIDIA NVLink Fusion high-speed chip interconnect technology and custom high-bandwidth memory (NVHBM) in next-generation trainable systems, enabling rack-scale heterogeneous architectures that combined Trainium chips with GPUs.
Those upgrades were presented as delivering a roughly 30% improvement in price performance over configurations relying solely on standard chips, and as enabling faster analytics using NVIDIA cuDF and cuVS libraries. AWS highlighted that vector index construction and analytics pipelines could complete significantly faster on GPU-accelerated nodes.
Software, models and robotics integration
AWS planned to make NVIDIA’s Nemotron family of open models available through Amazon Bedrock and SageMaker and to integrate NVIDIA’s physical AI stack into Amazon’s robotics roadmap. The physical AI suite referenced included OmniverseCosmos and Isaac tools AWS said would accelerate simulation-driven training for next-generation warehouse robots and other automation systems.
The companies said analytics services using NVIDIA’s libraries could run up to 3.7 times faster than standard configurations and that vector search index construction on Amazon’s search services could be roughly nine times quicker when executed on GPUs. The expansion was positioned to serve startups, enterprises and public-sector customers with a mix of accelerated compute and model hosting.
Implications for customers and the market
AWS framed the hardware commitment as a response to demand that had outpaced earlier projections and as a step to make agentic and physical AI more accessible at scale. The firms noted that recent hardware upgrades already delivered up to 4.6 times faster inference and 2.1 times stronger graphics output versus the previous generation, and that deploying millions of GPUs would support scientific discovery, enterprise automation and robotics innovation.
Deployments were to be staged globally across AWS infrastructure through 2027 and 2028. The companies emphasized existing collaboration over 16 years as the basis for the expanded plan and signaled continued investment in both accelerated GPU capacity and complementary CPU and interconnect technologies to meet evolving AI workloads.
