Amazon doubles down on Nvidia with 2M more GPUs despite building rival chips

Amazon and Nvidia have dramatically expanded their relationship, with AWS agreeing to deploy another 2 million Nvidia GPUs across its data centers in 2027 and 2028. The chips — spanning Nvidia’s Blackwell Ultra, Rubin, and Rubin Ultra lines — come just five months after Amazon committed to more than 1 million GPUs starting this year, a pace Nvidia says demand has already outstripped. No financial terms were disclosed, but given current GPU unit costs, the deal is comfortably in the tens of billions of dollars.

The timing is what makes this interesting. Amazon has spent years building its own silicon — Trainium chips positioned as a direct alternative to Nvidia’s H100 and Blackwell for deep learning workloads, plus its Graviton CPUs challenging Intel and AMD. The custom chip business crossed a $25 billion annualized revenue run rate last quarter, fueled by $225 billion in commitments from AI labs like Anthropic and OpenAI. Amazon’s AI chief Peter DeSantis has even floated selling Trainium to outside data center operators. And yet, when it comes to meeting surging demand from startups, enterprises, and governments, Nvidia remains the default choice.

The partnership extends well beyond raw GPU purchases. Nvidia’s networking hardware, open models, CPUs, data processing software, and robotics platform will all be integrated across AWS. On the enterprise side, AWS will serve Nvidia’s Nemotron family of open models on Bedrock and SageMaker. Amazon also plans to adopt Nvidia’s full physical AI stack — Omniverse, Cosmos, Isaac, and Jetson — to power its warehouse robot fleet.

Nvidia’s numbers put the demand in context: $96.2 billion in Q2 revenue, with data center sales up 117% year over year to $89 billion. The company expects $108 billion next quarter and has committed $279 billion to secure manufacturing capacity — up from $119 billion just last quarter. As Jensen Huang put it on the earnings call, “AI is generating profitable tokens… If we had more compute, we could generate more profitable tokens.” Amazon appears to agree, even if it’s simultaneously trying to build an escape hatch.