Amazon triples Nvidia chip order to three million GPUs

Amazon Web Services (AWS) has expanded its partnership with Nvidia by ordering an additional 2 million GPUs, tripling its previous commitment of 1 million chips made five months ago. Scheduled for delivery in 2027 and 2028, this massive order includes Nvidia Blackwell Ultra, Rubin, and Rubin Ultra GPUs. While financial terms were not disclosed, the deal is estimated to be worth tens of billions of dollars. AWS will also integrate Nvidia's Vera CPUs, networking hardware, and the Nemotron family of open models, which will be hosted on Amazon Bedrock and SageMaker.

This aggressive acquisition occurs even as Amazon scales its own custom silicon. The company's chip business recently surpassed a $25 billion annualized revenue run rate, supported by $225 billion in commitments from major AI labs like Anthropic and OpenAI. Amazon continues to promote its Trainium deep learning chips as alternatives to Nvidia's H100 and Blackwell processors, alongside its Arm-based Graviton CPUs. However, the massive Nvidia order underscores that proprietary hardware cannot yet satisfy the surging demand from enterprise customers, startups, and governments.

The deal highlights Nvidia's dominant market position. The chipmaker reported Q2 sales of $96.2 billion, with data center revenue accounting for $89 billion—a 117% year-over-year increase. Nvidia expects Q3 revenue to reach $108 billion as Rubin shipments begin. To secure future manufacturing capacity, Nvidia increased its supply commitments to $279 billion, up from $119 billion last quarter. This includes $92 billion for the remainder of the fiscal year and $87 billion in fiscal year 2028. Beyond cloud infrastructure, Amazon is adopting Nvidia's physical AI stack—including Omniverse, Cosmos, Isaac, and Jetson—to power its warehouse robotics fleet.

For AI practitioners and developers, this massive infrastructure expansion ensures that AWS will remain a highly capable environment for training and deploying next-generation models. Access to cutting-edge Rubin and Blackwell Ultra architectures, alongside Nvidia's software ecosystem, means developers can expect sustained compute availability. However, it also means practitioners must navigate a hybrid cloud environment where they must choose between Nvidia's industry-standard stack and Amazon's increasingly cost-effective, proprietary Trainium alternatives.