Amazon just tripled its Nvidia order five months after the last one. It's also quietly building the chips meant to compete with it.
The companies announced an expanded partnership Wednesday. Amazon is adding 2 million Nvidia GPU chips to its AWS data centers.
A Deal That Grew Fast
The new chips include Nvidia's Blackwell Ultra, Rubin and Rubin Ultra models. They're scheduled to arrive across 2027 and 2028.
This follows Amazon's prior commitment to more than 1 million Nvidia GPUs just five months ago. Nvidia said plainly that "demand has exceeded those expectations."
Neither company disclosed financial terms. Based on typical GPU unit costs, the deal is worth tens of billions of dollars.
That figure is a rough estimate, not a confirmed number. Neither Amazon nor Nvidia has said what either side is actually paying.
The partnership goes beyond chip sales specifically. Nvidia's networking hardware, open models, CPUs and data processing software will also integrate across AWS infrastructure.
Amazon Is Also Building Its Own Competition
Amazon has spent years developing its own chips to reduce Nvidia dependence. Its Trainium chips compete directly with Nvidia's hardware for deep learning workloads.
That business is growing fast.
Amazon's custom chip revenue crossed a $25 billion annualized run rate. That's backed by $225 billion in total commitments from AI labs including Anthropic and OpenAI.
Amazon's AI chief, Peter DeSantis, said AWS is now in talks to sell Trainium chips to other companies entirely. That would put Amazon in direct competition with Nvidia beyond its own data centers.
Despite that competition, Nvidia remains central to Amazon's plans. CFO Colette Kress said Nvidia will also ship an unspecified number of its new Vera CPUs to AWS.
Lead partners for that chip include Oracle and SpaceXAI.
The Robotics Angle
The partnership extends into Amazon's warehouse robot fleet too. Kress said Amazon plans to adopt Nvidia's full physical AI stack across that operation.
That stack includes Omniverse for simulation, Cosmos for world models, Isaac for robotics development, and Jetson for edge AI hardware. MAIN covered Nvidia's newest Jetson chip specifically last week.
On the enterprise side, AWS will also serve Nvidia's Nemotron open models through Amazon Bedrock and SageMaker.
Nvidia CEO Jensen Huang framed the broader moment on the earnings call.
"AI is generating profitable tokens," he said. "If we had more compute, we could generate more profitable tokens, which results in more profit for all of the services."
What This Means for Miami
This is the clearest sign yet of how central Nvidia remains to AI infrastructure.
That's true even as its biggest customers build competing chips of their own. Amazon can pursue both strategies simultaneously without either one canceling out the other.
For Miami investors weighing exposure to either company, that dual-track approach is worth understanding directly. Betting against Nvidia because Amazon builds Trainium misses that Amazon is expanding its Nvidia orders at the same time.
Anthropic Signs $45 Billion Compute Deal With Nscale