Amazon will add 2 million Nvidia GPUs to Amazon Web Services (AWS) data centers. Amazon, which is expanding its in-house AI chip business, is sharply increasing cooperation with Nvidia and speeding up infrastructure expansion to meet soaring AI computing demand.

On Aug. 26 (local time), IT outlet TechCrunch reported that Amazon and Nvidia plan to deploy Nvidia's Blackwell Ultra, Rubin and Rubin Ultra GPUs across AWS infrastructure in 2027 and 2028.

The deal goes well beyond the current scale of cooperation. Amazon decided about 5 months ago to deploy more than 1 million Nvidia GPUs across AWS, and will add another 2 million. Nvidia said demand was far stronger than expected, citing rising computing needs from startups, companies, AI research labs and the government sector.

The specific contract amount was not disclosed. Some expect it could reach tens of billions of dollars given the large-scale GPU supply.

The cooperation goes beyond GPU supply. Nvidia will integrate networking equipment that links thousands of GPUs into a single system into AWS infrastructure, along with open models, CPUs, data processing software and robotics platforms. Nvidia's Nemotron family of open models is also set to be offered through Amazon Bedrock, AWS' managed foundation model platform, and the cloud service SageMaker.

Server CPU supply will also expand. Nvidia Chief Financial Officer Colette Kress (콜렛 크레스) said Nvidia's Vera CPUs will be supplied to AWS along with the additional 2 million GPUs starting in the third quarter. Some will be integrated with Rubin GPUs, and some will be provided as standalone products.

A notable point is that Amazon expanded cooperation with Nvidia even as it strengthens its in-house chip strategy. Amazon is expanding its in-house AI accelerator Trainium and Graviton CPUs to reduce reliance on Nvidia. It is also discussing a plan to sell Trainium to other companies' data centers. Amazon said annualised revenue from its in-house chip business has surpassed $25 billion. It also said it secured commitments worth a total of $225 billion from AI research labs including Anthropic and OpenAI.

Even so, Nvidia's influence remains strong in large-scale AI infrastructure expansion. Nvidia said it will apply its physical AI platform to Amazon's logistics robots. This includes Omniverse, Cosmos, Isaac and Jetson.

Nvidia's results also show AI infrastructure demand remains strong. Nvidia posted second-quarter revenue of $96.2 billion, and data center revenue rose 117 percent from a year earlier to $89.0 billion. It forecast third-quarter revenue of $108.0 billion. Production shipments of next-generation Rubin GPUs also began this quarter.

It is also expanding supply chain investment in line with rising AI demand. Nvidia said it has secured supply and production capacity worth a total of $279 billion for current and future data center projects. That is a sharp increase from $119 billion in the previous quarter. Of that, $92 billion is scheduled to be spent in the remainder of the current fiscal year, and $87 billion is planned for fiscal 2028.

Nvidia CEO Jensen Huang (젠슨 황) said at the earnings release that AI is now performing productive and useful tasks. He stressed that securing more computing resources will enable more profitable AI work and expand service revenue.

Amazon's additional purchase of 2 million GPUs shows that AI competition is expanding beyond model development into large-scale infrastructure investment spanning data centers, semiconductors, networks and software. It is viewed as another confirmation of how fast AI computing demand is growing, as even Amazon, which is building up its own AI chips, continues large-scale investment in Nvidia GPUs.