Amazon will add another two million Nvidia GPUs to its data-center plans, sharply expanding a partnership announced only five months after the companies agreed to deploy more than one million Nvidia processors across AWS infrastructure.
The additional chips are scheduled for 2027 and 2028 and include Blackwell Ultra, Rubin and Rubin Ultra GPUs. Nvidia said demand had exceeded the expectations behind the earlier agreement, while the companies attributed the expansion to interest from startups, enterprises, AI laboratories and governments.
Neither side disclosed financial terms. TechCrunch estimates that the order is worth tens of billions of dollars based on GPU costs, but the precise value and Nvidia’s expected return remain unknown.
The agreement reaches beyond a larger GPU order
Nvidia’s networking equipment, open models, CPUs, data-processing software and robotics tools will also be integrated across Amazon’s operations. Vera CPUs are included in unspecified quantities, sometimes paired with Rubin systems and sometimes delivered independently.
The arrangement places Nvidia technology in both the cloud and Amazon’s physical operations. Amazon plans to adopt Nvidia’s physical AI stack for warehouse robots, including the Omniverse simulation platform, Cosmos world models, Isaac robotics development tools and Jetson edge-computing hardware.
AWS will also offer Nvidia’s Nemotron open models through Amazon Bedrock and SageMaker. The partnership therefore spans the systems used to train and run models, the managed services through which customers access them and the robotics stack used inside warehouses.
That breadth makes the deal more strategic than a hardware purchase. It connects Nvidia’s chips and software to several layers of Amazon’s AI business, potentially increasing the cost and effort required to replace individual components later.
Amazon is buying Nvidia while building rival chips
The expansion arrives as Amazon invests heavily in its own processors. AWS has developed Trainium as an alternative to Nvidia accelerators for deep-learning workloads and has discussed selling those chips to other companies for deployment in their data centers. Its Arm-based Graviton CPU also competes with traditional server processors.
Amazon says its custom-chip business has passed a $25 billion annualized revenue run rate. The company attributes that growth to $225 billion in total commitments from AI laboratories including Anthropic and OpenAI.
Buying two million additional Nvidia GPUs does not end that internal effort. Instead, it shows that Amazon expects demand to require both proprietary silicon and external supply. Customers may want different economics, model compatibility or performance characteristics, and AWS can use multiple processor families to serve those needs.
The balance also carries strategic tension. Nvidia becomes more deeply embedded across AWS even as Amazon tries to reduce dependence on it. TechCrunch reports the expansion as evidence that Nvidia remains dominant in AI chips, but the long-term competitive outcome is not established by this order alone.
Nvidia’s commitments grow with its revenue expectations
Nvidia reported second-quarter sales of $96.2 billion, above analyst estimates cited by TechCrunch. Data-center revenue accounted for $89 billion and rose 117 percent from the prior year. The company expects third-quarter revenue of $108 billion as initial Rubin sales begin.
To support current and future projects, Nvidia has committed $279 billion to manufacturing and supply capacity, up from $119 billion in the previous quarter. That includes projected spending of $92 billion for the rest of the fiscal year and $87 billion in fiscal 2028.
CEO Jensen Huang argues that AI is producing useful, profitable work and that more compute can generate more profitable tokens. That is Nvidia’s commercial thesis, not proof that every customer adding capacity will achieve equivalent returns.
The Amazon agreement strengthens the demand case, but it also raises the bar for utilization. Millions of processors must be installed, networked, powered and sold through services that customers value. Investors will be watching whether infrastructure spending converts into durable profit as smoothly as the suppliers predict.
