With an additional order of 2 million GPUs, AWS has pushed the global AI infrastructure competition to new heights. On August 26, local time, Amazon (AMZN) Web Services (AWS) and Nvidia (NVDA) announced an expanded strategic collaboration, planning to deploy an additional 2 million of Nvidia’s high-end GPUs across AWS’s global infrastructure between 2027 and 2028. This order doubles the 1-million-unit procurement plan announced in March of this year, officially incorporating the Blackwell Ultra, Rubin, and Rubin Ultra chips into AWS’s AI roadmap.
This partnership not only solidifies Nvidia’s dominance in the AI training market but also reveals that Amazon, despite its alternative strategy of developing its own chips, remains deeply reliant on Nvidia’s flagship products. The strategic alignment between the two parties marks that the AI compute competition has moved from the “hundreds-of-millions” level into the “trillion-dollar era.”
The scale of this transaction is unprecedented in the AI infrastructure sector, with the chip lineup covering Nvidia’s current three most powerful generations: the enhanced version of its flagship AI chip, Blackwell Ultra; the core GPU for the next-generation AI platform, Rubin; and the flagship-tier product, Rubin Ultra. The list price for a single GPU is estimated to be at least tens of thousands of dollars. Based on this estimate, the contract value for the 2-million-unit order is conservatively in the hundreds of billions of dollars. Adding the previously committed 1 million units, AWS’s committed procurement volume for Nvidia GPUs in 2026 alone reaches 3 million units.
This is not merely a procurement deal but also a full-stack technology binding. AWS will deploy Nvidia’s Vera CPU for the first time—a high-performance processor designed specifically for agentic AI workloads. Nvidia stated that some Vera CPUs will be integrated with Rubin, while others will be used as standalone products. At the same time, the two parties have deepened their collaboration on NVLink Fusion high-speed interconnect technology, pairing it with a new custom high-bandwidth memory (NVHBM) to allow Amazon’s self-developed Trainium chips to access faster and more energy-efficient memory. In addition, both parties plan to deliver an AI factory of 100,000 GPUs to the U.S. government on AWS infrastructure that meets Impact Level 6 security ratings.
Most dramatically, Amazon has been steadily increasing its investment in self-developed chips in recent years. Its Trainium accelerators have already attracted customers such as Anthropic, which have committed hundreds of billions of dollars in usage, and Amazon plans to sell them externally. However, the additional 2-million-GPU order demonstrates that self-developed chips and Nvidia GPUs are not an “either-or” choice but rather are expanding in parallel. AWS CEO Matt Garman stated that this expansion provides “more ways for cutting-edge labs, enterprises, and governments to build and deploy AI on AWS.”
