AWS Boosts Nvidia GPU Fleet to 2M for AI Demand

Amazon Web Services is significantly expanding its alliance with Nvidia, committing to deploy an additional 2 million Nvidia GPUs across its global infrastructure by 2028. This aggressive move directly responds to AI computing demand far surpassing earlier projections, coming just five months after an initial plan to add over 1 million Nvidia GPUs starting in 2026.

The strategic commitment underscores a deepening 16-year partnership, extending beyond GPUs to encompass CPUs, networking, AI models, data processing, and robotics. AWS CEO Matt Garman emphasized providing customers with optimal tools for AI workloads, ensuring seamless integration.

Expanding GPU Fleet & AI Infrastructure

The new deployment will feature Nvidia’s latest GPU architectures, including Blackwell Ultra, Rubin, and Rubin Ultra. Nvidia CEO Jensen Huang highlighted accelerating demand, noting the partnership’s expansion across the full stack to advance “agentic and physical AI.”

• Nvidia’s Vera CPU-based infrastructure will be introduced to AWS, specifically designed for CPU-intensive AI tasks like agentic AI.
• AWS’s custom Trainium chips will connect with Nvidia technology using NVLink Fusion and NVHBM, allowing co-operation within the same rack-scale system.
• Networking capabilities will improve through collaboration on Nvidia Spectrum technology, boosting performance for large AI training workloads across extensive GPU clusters.

Broadening the AI Ecosystem

A significant aspect of this collaboration involves building robust AI infrastructure for the US government. This initiative plans to deliver Nvidia’s AI stack, including 100,000 GPUs, on secure AWS infrastructure for federal and national-security workloads classified at Impact Level 6 (IL6) and above.

This secure infrastructure will leverage AWS’s Nitro System and Elastic Fabric Adapter (EFA) for enhanced security, reliability, and high-speed networking. The partnership extends to making Nvidia’s Nemotron family of open AI models available through Amazon Bedrock and Amazon SageMaker.

Key Performance Improvements

• Nvidia’s cuDF library integrated with Amazon EMR promises up to 3.7 times faster processing and 30% better price performance.
• GPU-based vector indexing for Amazon OpenSearch is expected to deliver up to nine times faster indexing at a quarter of the cost.

Advancing Physical AI & Robotics

The collaboration is also making significant strides into physical AI. Amazon Robotics will integrate Nvidia’s Jetson platform, Omniverse libraries, and Isaac robotics platform for developing next-generation warehouse robots.

This integration covers critical areas such as simulation, synthetic data generation, robot training, route optimization, and safety testing. Robotics systems will now be trained and tested using GPU-powered AWS infrastructure before real-world deployment, reflecting a crucial evolution in AI applications beyond traditional large language models.