Amazon and Nvidia just deepened one of the biggest partnerships in AI infrastructure. The two companies announced Wednesday, during Nvidia’s quarterly earnings call, a deal to add another 2 million Nvidia GPUs to Amazon’s data centers, according to TechCrunch AI. Neither side shared financial terms, but based on GPU unit costs, TechCrunch AI reports the agreement is worth tens of billions of dollars.
What stands out here is the speed. Just five months ago, Amazon agreed to deploy more than 1 million Nvidia GPUs across AWS. Now that number has more than tripled. Nvidia’s explanation was blunt: since the first deal, “demand has exceeded those expectations.”
⚡ What’s in the deal
This goes well beyond Amazon buying more chips. The expanded partnership pulls Nvidia’s full stack into AWS:
- 2 million Blackwell Ultra, Rubin, and Rubin Ultra GPUs, landing in AWS data centers in 2027 and 2028.
- Nvidia’s networking hardware that links thousands of GPUs into a single system.
- Nvidia Vera CPUs, some paired with Rubin, some standalone.
- Nemotron open models served on Amazon Bedrock and SageMaker.
- Nvidia’s physical AI stack (Omniverse, Cosmos, Isaac, Jetson) to power Amazon’s warehouse robots.
The companies said “surging demand” from startups, enterprises, AI labs, and even governments pushed them to work more closely.
🤔 Why this matters
Here’s the twist. Amazon is buying more Nvidia hardware even as it builds chips to compete with Nvidia. Its Trainium chips are a direct alternative to Nvidia’s H100 and Blackwell for deep learning, and AWS is reportedly in talks to sell Trainium to other companies. Its Arm-based Graviton CPU challenges Intel and AMD.
Amazon’s custom silicon business isn’t small either. On its last earnings call, the company said the unit crossed a $25 billion annualized revenue run rate, backed by $225 billion in total commitments from labs like Anthropic and OpenAI.
So Amazon is both Nvidia’s biggest customer and one of its rivals. That it’s still tripling down on Nvidia tells you who’s winning the AI chip race right now. Nvidia remains the default.
📊 The numbers behind it
Nvidia’s quarter backs that up. Per TechCrunch AI:
- $96.2 billion in Q2 sales, beating estimates.
- $89 billion of that from data centers, up 117% year over year.
- $108 billion revenue guidance for Q3, some of it from next-gen Rubin GPUs now in production.
- $279 billion committed to secure supply and manufacturing, up from $119 billion last quarter.
That supply commitment is the tell. Nvidia is locking down memory and manufacturing capacity years out because it expects demand to hold.
CEO Jensen Huang framed the whole thing around economics. “AI is generating profitable tokens,” he said on the call. “If we had more compute, we could generate more profitable tokens, which results in more profit for all of the services.”
🔭 What comes next
The first of the 2 million chips start arriving in Q3. Vera CPUs are already shipping to lead partners including Oracle and xAI, and Nvidia expects every major hyperscaler and AI lab to adopt them.
The open question is whether all this compute actually pays off. Huang’s argument is that more GPUs equal more profit. Investors will be watching Rubin’s early sales and the returns AI companies see after pouring hundreds of billions into infrastructure. If that math holds, expect more mega-deals like this one. If it doesn’t, this is the moment people will point back to.
For practitioners, the near-term signal is simple: Nvidia capacity on AWS is expanding fast, and Rubin-class hardware is coming online sooner than many expected. Full details are available at the original source, TechCrunch AI.