AWS Orders Two Million More Nvidia GPUs for 2027–28
Amazon will deploy two million additional Blackwell Ultra and Rubin-generation GPUs while extending Nvidia technology across AWS and robotics.
A hyperscale commitment
Amazon Web Services and Nvidia announced that AWS plans to deploy two million additional Nvidia GPUs across its global infrastructure during 2027 and 2028. The order covers Blackwell Ultra, Rubin and Rubin Ultra accelerators and follows an earlier commitment, announced five months ago, to deploy more than one million Nvidia GPUs. Financial terms were not disclosed, but an order of this scale implies spending worth tens of billions of dollars.
The agreement extends well beyond accelerator purchases. AWS plans to incorporate Nvidia’s Vera CPUs and networking technology into future systems, while offering Nvidia’s Nemotron open models through Amazon Bedrock and SageMaker. Amazon will also use Nvidia’s Omniverse simulation platform, Cosmos world models, Isaac robotics software and Jetson edge hardware across warehouse and physical-AI work, including robot training, synthetic-data generation and route optimization.
Cooperation alongside competition
The expansion is striking because Amazon has invested heavily in reducing its reliance on Nvidia. Its Trainium accelerators compete for AI training and inference workloads, while its Graviton CPU business has reached a reported $25 billion annualized revenue run rate. AWS also has large infrastructure commitments from customers including Anthropic and OpenAI, giving it a strong incentive to develop lower-cost in-house silicon.
Yet customer demand still requires Nvidia compatibility and supply at enormous scale. The new order indicates that custom chips are expanding AWS’s capacity rather than displacing Nvidia in the near term. It also secures AWS access to Rubin-generation systems before those products become the industry’s standard deployment base.
Why it matters
Two million additional GPUs materially expand the amount of frontier-model training and inference capacity scheduled to come online in 2027–28. The agreement also shows Nvidia converting its accelerator lead into a broader position spanning CPUs, networking, models and robotics. For Amazon, the unresolved question is whether demand and utilization will grow fast enough to justify the capital committed before the hardware arrives.