Amazon Places Additional Order for 2 Million NVIDIA GPUs
Nashnova编辑部
Amazon will buy an additional 2 million high-end Nvidia GPUs over the next two years for AWS data-center expansion, bringing its total commitment — combined with a 1-million-unit order announced in March — to 3 million chips, one of the largest single-vendor procurement pledges in cloud-computing history.
How big is this order?
The new 2 million GPUs will be deployed in AWS data centers in 2027 and 2028, spanning Nvidia's Blackwell Ultra, Rubin, and Rubin Ultra product lines.
Each chip carries a list price of at least tens of thousands of dollars; large customers typically receive discounts. Neither company disclosed the deal's total value.
Combined with the 1 million units announced in March, Amazon's total Nvidia GPU commitment now stands at 3 million. This means → Amazon alone has locked up a significant share of Nvidia's high-end capacity for the next two years.
Isn't Amazon building its own chips? Why buy so many?
Amazon's chip division, Annapurna Labs, has developed proprietary CPUs and AI accelerators — custom chips designed to compete directly with Nvidia GPUs.
Yet in-house and external procurement run in parallel. In plain terms = custom silicon is the long-term cost play; Nvidia GPUs are the foundation that actually runs today's workloads — Amazon is walking on both legs.
Amazon also said its next-generation Trainium AI chip will use Nvidia's networking interconnect technology. This reflects the reality that even on the in-house track, Amazon cannot fully bypass Nvidia's ecosystem.
What does this mean for Nvidia?
All three major cloud giants — Amazon, Microsoft, and Google — are developing custom chips, raising concerns they could erode Nvidia's share over time.
But a 3-million-unit commitment makes one thing clear: for AI training and inference workloads through at least 2027–2028, Nvidia GPUs remain irreplaceable core infrastructure.
This means → custom chips are filling edge cases in the near term, not displacing Nvidia at the center of the compute stack. For Nvidia, its largest customers' "in-house anxiety" has not yet translated into actual order losses.
市场有风险，内容仅供研究参考，不构成投资建议。