AWS and NVIDIA say they plan to deploy 2 million additional NVIDIA GPUs across AWS’s global infrastructure in 2027 and 2028, expanding an AI infrastructure partnership that now spans GPUs, CPUs, networking, memory, open models, data processing, vector indexing, robotics, and secure government workloads.
The companies say the new capacity will use NVIDIA Blackwell Ultra, Rubin, and Rubin Ultra GPUs across AWS Global Infrastructure, including AI factories. AWS had already announced plans at NVIDIA GTC 2026 to add more than 1 million NVIDIA GPUs starting in 2026. The new announcement says demand exceeded those expectations.
The headline number matters, but the broader stack is the story. AWS and NVIDIA are not describing a simple accelerator purchase. They are packaging cloud capacity around the kinds of workloads that agent platforms create: long-running inference, retrieval, data preparation, simulation, robotics, and secure public-sector deployments.
The expansion is a full-stack cloud bet
NVIDIA says the expanded collaboration includes work to bring Vera CPU-based infrastructure to AWS, extend NVIDIA NVLink Fusion with custom NVIDIA high-bandwidth memory, and integrate NVIDIA systems with the AWS Nitro System and Elastic Fabric Adapter.
The companies also plan AI factories for the U.S. government, including 100,000 GPUs on secure AWS infrastructure for federal and national-security workloads. NVIDIA says those workloads are intended for environments classified at Impact Level 6 and above.
That government detail is important because AI infrastructure demand is no longer only a hyperscaler and frontier-lab story. If agencies, enterprises, and robotics operators move production workloads into agentic systems, capacity has to ship with compliance, reliability, isolation, and procurement paths.
NVIDIA and AWS also say NVIDIA Nemotron open models will continue to be available on Amazon Bedrock and Amazon SageMaker. That keeps the announcement tied to model distribution, not just hardware supply.
Retrieval and robotics are part of the capacity pitch
The release also names data processing and vector indexing. AWS and NVIDIA say they are collaborating on GPU-accelerated data processing in Amazon EMR using EC2 G7 instances and NVIDIA cuDF. For vector search, the release says GPU-accelerated indexing on Amazon OpenSearch can offload index construction onto dedicated GPUs.
Those details show where cloud AI bottlenecks are moving. Agentic systems often spend as much time gathering context, searching records, transforming data, and calling tools as they spend generating text. If the retrieval and data layers lag, bigger models do not automatically create better workflows.
Physical AI is another named lane. Amazon Robotics is working with NVIDIA’s physical AI platform, including Jetson, Omniverse libraries, and Isaac, for simulation, synthetic data generation, robot training, route optimization, functional safety, and real-to-sim validation.
Capacity promises still need delivery dates
The announcement is a plan, not delivered capacity. The 2 million additional GPUs are targeted for 2027-2028. That leaves execution risk around supply chains, memory, data-center construction, power, networking, and customer demand.
For AI buyers, the useful takeaway is to watch what AWS exposes as actual products. GPU counts are not directly comparable to usable capacity if the instance types, regional availability, reservation terms, networking topology, managed-model surfaces, and data services are not ready when workloads need them.
