Amazon announced it will expand its collaboration with NVIDIA to deploy up to three million NVIDIA AI GPUs in its AWS cloud infrastructure, tripling the one‑million‑GPU commitment disclosed in March.

The initial agreement, announced at the GTC conference in March, set the target at one million GPUs and introduced support for NVIDIA’s custom Trainium chips and the upcoming Vera CPU architecture. Under the new expansion, the partnership will add two million additional GPUs, including 100,000 units allocated to AI factories designed for U.S. government use at Impact Level 6 security classification. Amazon also plans to integrate NVIDIA’s NVHBM memory technology, which the company says offers 30 % greater bandwidth and 15 % higher power efficiency than HBM4E.

The expanded deployment supports workloads for agentic AI, automation and physical AI, and complements AWS’s existing Trainium processors. NVIDIA chief executive Jensen Huang stated that demand for the combined platform is outpacing forecasts and that the partnership now spans the full stack of GPUs, CPUs, networking, models and software. Amazon said the arrangement allows customers to select the hardware that best fits their AI computing needs.

The deal underscores the growing scale of AI infrastructure investments within cloud services and may shape future competition among major cloud providers.