AWS-NVIDIA GPU Expansion: 2M More Chips and Government Capacity

Three million GPUs, on top of what was already committed

AWS and announced that AWS will deploy 2 million additional NVIDIA GPUs in 2027 and 2028, on top of the more than 1 million GPUs AWS had already committed to deploy starting in 2026. That puts the combined commitment past 3 million GPUs across the two waves. [1]

Of the new 2 million, 100,000 GPUs are set aside for secure U.S. government workloads running on infrastructure classified at Impact Level 6 and above, one of the highest government security tiers. That is the tier used for national security and classified data, which points this carve-out at federal and defense customers rather than commercial AWS tenants. [2]

Three NVIDIA chip generations, spanning two years of roadmap

The deployment spans Blackwell Ultra, the current-generation NVIDIA accelerator that followed the original Blackwell chip, and Rubin and Rubin Ultra, the successor generation NVIDIA has previewed as its next architecture. Naming three generations across a single 2027-2028 commitment signals that AWS is contracting for NVIDIA's roadmap rather than for one fixed chip. [1]

What "agentic and physical AI infrastructure" means beyond the chips

The companies described specific pieces beyond the GPUs themselves: NVIDIA Vera CPUs paired with the accelerators for general-purpose compute, NVLink Fusion interconnect bringing custom high-bandwidth memory to AWS's own Trainium chips, and NVIDIA Spectrum networking for large-scale training clusters. AWS layers its Nitro System and Elastic Fabric Adapter on top for isolation and network reliability. [2]

"Physical AI" refers to NVIDIA's robotics and simulation stack, the Jetson edge-computing platform, Omniverse simulation software, and the Isaac robotics development platform, which the announcement ties to warehouse automation. "Agentic AI" covers the production-scale AI factories AWS is building for automated multi-step workloads, scientific computing, and enterprise workflows. Together the phrase describes a specific hardware and software stack, not a marketing label for the GPUs alone. [2]

The stock fell on the spending, not the demand

Atlas interpretation: A multi-year hardware commitment this size is future revenue for NVIDIA and future capital spending for Amazon, and the two readings diverged. Amazon's capital expenditures had already reached roughly $220 billion in 2026, and free cash flow had swung from a positive $18.2 billion to a negative $7.6 billion year over year. Piling another two years of GPU purchases onto that base is what moved the stock, not any doubt about whether the capacity would get used. [3]

The Motley Fool's reporting notes that AWS revenue growth had accelerated to 37 percent, its backlog was growing at a triple-digit rate, and AWS's CEO had said capacity, not customer demand, was the binding constraint through 2027. The sell-off reflected investors weighing whether a budget that large leaves little room for a wrong demand forecast, not skepticism that the demand exists. [3]