Amazon Doubles Its Nvidia GPU Order to 3 Million, Betting Big on AI Compute Through 2028
Amazon and Nvidia announced an expanded partnership on August 26, 2026, during Nvidia's quarterly earnings call, adding another 2 million Nvidia GPU chips to Amazon Web Services data centers for deployment in 2027 and 2028. The order includes Nvidia Blackwell Ultra, Rubin, and Rubin Ultra GPUs. TechCrunch
The expanded deal comes roughly five months after Amazon agreed to deploy more than 1 million Nvidia GPUs across AWS infrastructure starting in 2026 — a commitment that included Blackwell and Rubin architectures across AWS's global cloud regions. Nvidia said that since that earlier agreement, demand for the chips has exceeded expectations. The two companies cited surging demand from startups, enterprises, AI labs, and governments as influencing the decision to deepen the partnership. TechCrunch
Neither Amazon nor Nvidia disclosed financial terms. TechCrunch estimated the deal to be worth tens of billions of dollars. TechCrunch
The scope extends well beyond GPU procurement. Nvidia's networking hardware for connecting GPUs into systems, open models, CPUs, data processing software, and its robotics platform will all be integrated across AWS. As part of the arrangement, Nvidia plans to send an unspecified number of its Vera CPUs to AWS — some integrated with Rubin, others standalone, according to Nvidia CFO Colette Kress. Kress also said Nvidia expects Vera CPUs to be deployed by every major hyperscaler (large-scale cloud operator), neocloud (newer, AI-focused cloud providers like CoreWeave), AI lab, and system OEM, with shipments already underway to lead partners including Oracle and SpaceX AI. TechCrunch
Amazon also plans to adopt Nvidia's full physical AI stack, including the Omniverse simulation platform, to power its fleet of warehouse robots. AWS and Nvidia have collaborated since 2010, beginning with the world's first GPU cloud instance — a single virtual server in the cloud that offered GPU acceleration, which was novel at the time. TechCrunch AWS
The chip deal arrives alongside Nvidia's notification to customers about AI-related price hikes above 15 percent on systems shipped early next year, including those featuring its flagship Vera Rubin platform. Bloomberg
In parallel with its Nvidia procurement, Amazon is advancing its own silicon strategy. Amazon's AI chief Peter DeSantis has said AWS is in talks to sell its Trainium chips — positioned as an alternative to Nvidia's H100 or Blackwell chips — to other companies for use in their own data centers. Amazon said its custom chip business crossed a $25 billion annualized revenue run rate, driven by $225 billion in total commitments from AI labs including Anthropic and OpenAI. TechCrunch
The broader context here is that Amazon is running two strategies at once, and they are not necessarily in conflict. On one track, AWS is deepening its dependence on Nvidia's roadmap — committing to Blackwell Ultra, Rubin, Rubin Ultra, Vera CPUs, networking, and Omniverse. On the other, it is building out Trainium as a merchant silicon offering for external data center customers. AWS's internal capacity needs for training (the process of building AI models by feeding them vast amounts of data) and inference (running those trained models to produce outputs) at scale almost certainly exceed what its own chips can supply in the near term, making Nvidia GPUs a volume necessity even as Trainium targets a different margin profile and customer segment. The $225 billion in commitments from AI labs suggests the demand underpinning both tracks is not speculative.
The price hikes add another variable worth considering. If Nvidia is raising system prices above 15 percent on Vera Rubin shipments, the economics of large-scale GPU deployment shift for every hyperscaler, not just Amazon. AWS's custom silicon push gains a sharper cost-justification narrative when the primary alternative supplier is raising prices on its flagship platform. Whether Trainium can deliver competitive performance per dollar at the workloads that matter most to AI labs remains an open question, but Amazon's willingness to sell it externally signals confidence that the gap is narrowing.
Nvidia's Vera CPU roadmap also deserves attention. Nvidia's ambition to place Vera across every major hyperscaler, neocloud, and AI lab would extend its footprint beyond GPU acceleration into the general-purpose compute layer that currently belongs to AMD EPYC and Intel Xeon. If Kress's expectation materializes, the competitive dynamics in the data center CPU market shift in a meaningful way.
AWS, Google Cloud, Microsoft, and OCI (Oracle Cloud Infrastructure) are all set to be among the first cloud providers to deploy Nvidia Vera Rubin-based instances in 2026. The competitive pressure to offer the latest Nvidia architectures is uniform across the hyperscalers; Amazon's move to lock in 2 million additional GPUs through 2028 is as much about securing supply allocation as it is about capacity planning.
For the broader market, the signal is that GPU demand at hyperscaler scale continues to outpace even aggressive procurement plans. Amazon ordered 1 million Nvidia GPUs in March 2026. Five months later, it doubled that commitment. Nvidia's statement that demand exceeded expectations since the first deal suggests the gap between projected and actual consumption is widening, not stabilizing — which, if sustained, will keep capital expenditure trajectories steep across the industry for the foreseeable future.
