Amazon Web Services and NVIDIA said on August 26 they will deploy 2 million additional NVIDIA GPUs across AWS’s global infrastructure in 2027-2028, on top of the more than 1 million units already committed for 2026, and will drop NVIDIA’s Vera CPUs, NVLink Fusion, Nemotron open models and physical-AI stack into the AWS platform.
2 million more Blackwell Ultra, Rubin and Rubin Ultra GPUs
The new commitment covers Blackwell Ultra, Rubin and Rubin Ultra parts across AWS regions and dedicated AI factories, and expands existing Blackwell capacity with NVIDIA RTX PRO 4500 Blackwell Server Edition GPUs on the new Amazon EC2 G7 instances — AWS says the first major cloud offering to ship that SKU. G7 delivers 4.6× the AI-inference performance and 2.1× the graphics performance of the previous G6 generation, and will be paired with NVIDIA Spectrum networking for large-scale training. AWS chief Matt Garman framed the deal around customer choice, while NVIDIA CEO Jensen Huang said “demand is running ahead of every forecast” for the pair’s decade-and-a-half joint stack.
Vera CPUs, NVLink Fusion and secure federal AI factories
Beyond raw GPU counts, AWS is bringing NVIDIA’s Vera CPU platform to its instance line-up and expanding the Trainium+NVLink Fusion tie-up first shown at re:Invent 2025 with NVIDIA’s new custom high-bandwidth memory (NVHBM). That would let AWS’s Annapurna Labs mix Trainium chips and NVIDIA GPUs inside a common rack-scale fabric with shared memory. The pair will also build federal AI factories running 100,000 GPUs on AWS’s secure infrastructure for U.S. government workloads certified at Impact Level 6 (IL6) and above.
Physical AI wires into Amazon Robotics
The most concrete robotics angle is a new collaboration between Amazon Robotics and NVIDIA covering the Jetson platform, Omniverse simulation libraries and the Isaac open robotics platform. AWS says the joint stack will span simulation, synthetic-data generation, robot training, route optimisation, functional safety and real-to-sim validation on GPU-accelerated EC2 instances, feeding the next generation of Amazon warehouse robots. Nemotron open models remain available on Bedrock and SageMaker, and cuDF/cuVS-accelerated data pipelines land on Amazon EMR and OpenSearch. The deal follows NVIDIA’s Cloverleaf U.S. AI-factory partnership, Anthropic’s Theseus infrastructure venture and the $500B SK Group–NVIDIA AI factory pact.
Reporting based on coverage from GlobeNewswire, AWS and NVIDIA.