Amazon just tripled its order of Nvidia chips over ‘surging demand’
TechCrunch Rebecca Bellan, Kirsten Korosec ● Covered by 3 sources
Amazon just added 2 million more Nvidia chips to AWS. Nvidia says demand blew past the last deal, even as Amazon builds its own chips.
Based on reporting by TechCrunch, Rebecca Bellan, Kirsten Korosec — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Amazon and Nvidia are deepening a relationship that already looked enormous. On Wednesday, the two companies said Amazon Web Services will add another 2 million Nvidia GPU chips to its data centers, with the hardware arriving in 2027 and 2028. The chips include Blackwell Ultra, Rubin, and Rubin Ultra, all built for the kind of heavy lifting that trains and runs AI models.
This is not a small follow-on order. Just five months ago, Amazon said it would deploy more than 1 million Nvidia GPUs across AWS starting this year. Nvidia now says demand has gone beyond that plan. Neither company gave financial terms, but at GPU prices this is still a deal that runs into tens of billions of dollars.
The partnership also goes well beyond Amazon simply buying more chips. Nvidia said its networking gear, CPUs, data-processing software, robotics platform, and open models will all be folded more tightly into AWS. That matters because Amazon is also pushing hard on its own silicon. It has been building chips to cut its dependence on Nvidia, and Peter DeSantis, who leads Amazon’s AI effort, has said AWS is talking about selling Trainium chips to other companies for use in data centers.
Even so, Nvidia is still the one AWS keeps leaning on when the demand gets serious. Amazon said its custom chip business has reached a $25 billion annualized revenue run rate, helped by $225 billion in total commitments from AI labs including Anthropic and OpenAI. But Wednesday’s announcement shows that the market for AI compute is still big enough to pull both strategies forward at once.
Nvidia is also sending Vera CPUs into the mix, with Colette Kress saying some will be integrated with Rubin and others sold separately. Beyond AWS, she said Nvidia expects Vera to show up across major hyperscalers, neoclouds, AI labs, and system OEMs, with Oracle and SpaceXAI already among the early partners. The company is clearly trying to make its chips the default stack, not just the fastest one.
The rest of the deal reaches into Amazon’s warehouses and its enterprise cloud. Amazon plans to use Nvidia’s physical AI stack for robots, including Omniverse, Cosmos, Isaac, and Jetson. On the software side, AWS will host Nvidia’s Nemotron open models on Bedrock and SageMaker. And all of this lands as Nvidia reports second-quarter sales of $96.2 billion, with data center revenue at $89 billion, up 117% from a year earlier.
My take — AI-written commentary, not fact-checked reporting
This is the clearest proof yet that AI infrastructure has become a two-track race: build your own chips, then keep buying Nvidia anyway. Amazon can talk up Trainium all it wants, but when demand gets this hot, Nvidia still gets the call. The funny part is that everyone is trying to reduce dependence on the same supplier while simultaneously handing it more business.
Read more about this at: TechCrunch