Overview
- The companies announced Thursday that AWS will deploy an additional two million Nvidia GPUs across 2027 and 2028, raising its planned Nvidia capacity to more than three million units when combined with a prior March commitment.
- The order covers Nvidia’s high‑end Blackwell Ultra, Rubin, and Rubin Ultra GPU platforms that are used for large model training and inference workloads.
- The expanded agreement formalizes broader integration of Nvidia technology into AWS, including Vera CPU infrastructure, NVLink Fusion interconnects, memory and networking gear, and Nvidia’s physical AI tools for Amazon Robotics.
- AWS will reserve a secure tranche of 100,000 Nvidia GPUs for U.S. federal and national‑security workloads, while Nvidia says it is supply‑constrained and facing higher memory and component costs that pressure margins.
- AWS already has customer reservations stretching into 2027 and 2028, which together with long data‑center lead times means the deal will require more power, cooling, and construction and could leave capacity short of demand despite the multi‑year purchase.