Amazon triples its Nvidia GPU order for AWS

Amazon is adding 2 million Nvidia GPUs, a mix of Blackwell Ultra, Rubin, and Rubin Ultra chips plus Vera CPUs, to AWS data centers for deployment across 2027 and 2028, the companies said August 26. Neither side disclosed exact terms, but the order is worth “tens of billions of dollars” based on prevailing per-unit GPU costs. It triples a prior order of more than 1 million GPUs that Amazon placed just five months earlier and which began deploying in 2026. Nvidia said in a statement that “demand has exceeded those expectations” since that first order. The expanded deal extends beyond raw chip purchases to include networking hardware, software, and integration with Nvidia’s robotics platform.

What it means for you

AWS builds custom AI chips of its own, so a tripled Nvidia order from Amazon specifically is a stronger demand signal than a similar move from a pure Nvidia reseller: it means AWS’s own internal silicon isn’t covering the appetite it’s seeing from startups, enterprises, and government customers on its platform. If you run inference workloads on Bedrock or SageMaker, this points toward more capacity coming online in 2027-2028, though Nvidia’s own guidance that supply is its binding constraint means “more capacity” is relative, not a guarantee of cheaper GPU-hours anytime soon, the same lesson underlying Nvidia’s broader shift into financing AI infrastructure directly.