Amazon Web Services has agreed to deploy an additional 2 million Nvidia GPUs across its data centers, the companies announced Wednesday, five months after AWS committed to more than 1 million Nvidia chips for delivery this year. The new order covers Blackwell Ultra, Rubin, and Rubin Ultra GPUs, arriving in 2027 and 2028 rather than immediately. Nvidia said in a statement that demand from AWS customers has “exceeded those expectations” set back in the spring.
Neither company disclosed financial terms. Given publicly known GPU unit costs, an order of that scale runs into the tens of billions of dollars, though the precise figure is not confirmed by either side.
The timing is the real story. AWS is not reacting to a current capacity shortage; it is pre-buying chips that will not land in its data centers for one to two years. That is a wager that enterprise, startup, and government demand for AI compute keeps climbing through 2028, not a response to demand it is straining to meet today. If that bet is wrong, AWS carries the depreciation risk on hardware ordered well ahead of the workloads that would justify it.
The order also sits awkwardly next to Amazon’s own chip ambitions. AWS has spent years building Trainium, a direct alternative to Nvidia’s GPUs for training and running AI models, and Graviton, its Arm-based CPU family pitched against the conventional server silicon sold by Intel and AMD. Peter DeSantis, AWS’s infrastructure chief, has said the company is in talks to sell Trainium to outside data center operators, positioning it as a competitor to Nvidia rather than just an internal cost hedge. Amazon has also said its custom chip business passed an annualised revenue pace of $25 billion, underpinned by $225 billion in commitments booked from AI labs, Anthropic and OpenAI among them.
Tripling a rival’s order five months after the last one is a signal about who actually controls the supply AWS’s customers need right now. A company spending billions to reduce its dependence on Nvidia is, at the same time, agreeing to buy millions more of Nvidia’s chips because its own silicon cannot yet cover demand at the scale customers are asking for. That is not a contradiction so much as a hedge: build toward independence over years, but do not let a supply gap show up in this quarter’s or next year’s capacity.
The partnership extends past GPUs. Nvidia will also send AWS its new Vera CPU, some units integrated with Rubin GPUs and others sold standalone, according to Nvidia CFO Colette Kress. Nvidia’s Omniverse, Cosmos, Isaac, and Jetson robotics platforms will support Amazon’s warehouse robot fleet, and AWS will host Nvidia’s Nemotron open models on Bedrock and SageMaker. A portion of the new GPU capacity is also earmarked for AI systems used by the US government, though neither company specified how much.
For AWS customers, the practical read is that 2026 and 2027 capacity constraints are unlikely to ease from this deal; the chips it covers do not ship until 2027 and 2028. Enterprises planning multi-year AI infrastructure contracts should treat this as evidence that Nvidia supply, not Amazon’s own silicon, remains the binding constraint on AWS capacity through at least the next two years.
Reported by TechCrunch on 26 August 2026.
