AWS and NVIDIA to Deliver 2 Million Additional GPUs and Next-Generation Infrastructure for Agentic and Physical AI
AI-generated image: synthetic visual, not an actual depiction of events, people, or locations.
Amazon Web Services (AWS) and NVIDIA (Nasdaq: NVDA) announced an expansion of their strategic partnership today, committing to deploy an additional 2 million NVIDIA GPUs across AWS global infrastructure throughout 2027 and 2028.
The multi-year agreement builds on previous commitments, including AWS's plan announced at NVIDIA GTC 2026 to add more than 1 million GPUs starting in 2026. The accelerated rollout reflects surging enterprise, sovereign, and frontier lab demand for accelerated compute across training, inference, and physical AI workloads.
Silicon Expansion and Heterogeneous Architecture
Under the expanded roadmap, AWS will integrate NVIDIA's next-generation Blackwell Ultra, Rubin, and Rubin Ultra architectures into its AI factory clusters.
The technical collaboration spans several key hardware and silicon co-engineering initiatives:
Vera CPU Deployment: AWS is working to bring NVIDIA Vera CPU-based infrastructure to its cloud platform, delivering dedicated high-performance CPU architectures tailored specifically for agentic AI orchestration alongside accelerated clusters.
Annapurna Labs and NVLink Fusion: Building on their re:Invent 2025 announcement supporting NVLink Fusion in next-generation Trainium chips, NVIDIA and Amazon's Annapurna Labs are extending integration to support NVIDIA's custom high-bandwidth memory (NVHBM). The architecture allows Trainium silicon to leverage NVIDIA memory subsystems and scale-up interconnects within unified rack-scale configurations.
Blackwell Server Instances: AWS is introducing Amazon EC2 G7 instances powered by NVIDIA RTX PRO 4500 Blackwell Server Edition GPUs, becoming the first major cloud provider to offer the accelerator. The G7 instances deliver 4.6x the AI inference performance and 2.1x the graphics performance of previous-generation G6 systems.
Networking and Security Backbone: All GPU-accelerated and Trainium-based EC2 instances will continue to be anchored by the AWS Nitro System and interconnected via AWS Elastic Fabric Adapter (EFA) and NVIDIA Spectrum networking.
100,000 GPU Sovereign Federal AI Deployment
Addressing public sector and defense computing requirements, AWS and NVIDIA confirmed plans to construct dedicated AI factories for the United States government.
The initiative will deploy 100,000 NVIDIA GPUs across secure AWS infrastructure engineered to process classified national security workloads at Impact Level 6 (IL6) and above.
Data Acceleration, Open Models, and Physical AI
Beyond silicon deployment, the expanded agreement targets data pipeline throughput and robotics automation across Amazon's ecosystem:
Accelerated Analytics: AWS is implementing NVIDIA cuDF libraries on Amazon EMR to achieve up to 3.7x faster processing speeds and a 30% price-performance gain over CPU setups. Additionally, GPU-accelerated vector indexing on Amazon OpenSearch Service via cuVS delivers up to 9x faster vector indexing at a quarter of previous costs.
Amazon Robotics: Amazon Robotics is standardizing on NVIDIA's full-stack physical AI platform, utilizing NVIDIA Jetson, Omniverse libraries, and the Isaac robotics platform for warehouse automation, synthetic training data generation, and fleet simulation.
Nemotron Model Availability: NVIDIA's Nemotron open-model family will remain available as managed serverless offerings on Amazon Bedrock and for customized deployment on Amazon SageMaker.
