AWS-Nvidia Deal: 2 Million GPUs and Vera CPUs

Alongside Nvidia’s earnings, the two companies that sit at the center of the AI build-out disclosed how deep their alliance runs. On August 26, 2026, Amazon Web Services and Nvidia announced an expansion of their partnership under which AWS will deploy 2 million additional Nvidia GPUs across its global infrastructure over the next two years, bring Vera CPUs into its fleet, and jointly build secure AI data centers for the U.S. government. It is one of the largest single infrastructure commitments the cloud industry has disclosed, and it lands as concrete evidence behind the demand Nvidia described in its quarterly results.

Two million GPUs, three generations

The centerpiece is volume. AWS plans to deploy roughly 2 million additional Nvidia GPUs across its data centers and AI factories through 2027 and 2028 — the companies framed the runway as extending into AWS’s fiscal 2029. The chips span three generations of Nvidia’s roadmap: Blackwell Ultra, Rubin, and Rubin Ultra. That mix matters. Rather than a one-time buy of current silicon, it is a rolling commitment that carries AWS’s capacity build across Nvidia’s next two architecture transitions, smoothing the awkward demand cliffs that usually accompany a generational handoff.

The capacity is aimed at workloads AWS says are moving “from pilot to production”: agentic AI, scientific discovery, enterprise automation, and the emerging category of physical AI — models that perceive and act in the real world rather than only generating text. Those workloads are exactly the ones that consume the most compute per query, which is why hyperscalers keep signing for more GPUs than they can seemingly use.

Vera CPUs join the stack

The deal is not GPUs alone. AWS will also bring Nvidia’s Vera CPU into its infrastructure, giving customers a high-performance central-processor option to sit alongside accelerated compute for agentic workloads. Nvidia said the cloud provider could purchase “millions of CPUs” over the life of the arrangement, on top of the 2 million GPUs.

That is a notable strategic wrinkle. AWS designs its own Graviton Arm-based server CPUs and has spent years steering customers toward them for general-purpose compute. Adopting Vera signals that for the specific demands of agentic AI — where CPU and GPU are tightly coupled in a single system — AWS is willing to run Nvidia’s processors rather than insist on its own. It is a pragmatic concession to how Nvidia’s rack-scale systems are engineered, where the CPU and GPU are co-designed to feed each other at high bandwidth.

A federal AI build-out

The most consequential piece for the long term may be the government angle. AWS and Nvidia said they will build AI factories for the U.S. government, with plans to deliver 100,000 GPUs on AWS’s secure infrastructure for federal and national-security workloads. The systems are designed to support data classified at Impact Level 6 (IL6) and above — the tier reserved for classified national-security information.

That places the two companies at the center of federal AI procurement at a moment when Washington is racing to field AI capabilities across defense and intelligence. A 100,000-GPU secure enclave is a serious sovereign-scale system, and running it inside AWS’s accredited government regions — rather than a bespoke on-premises facility — is a bet that the cloud operating model can meet the security bar for the most sensitive workloads. For AWS, federal AI is a high-margin, sticky market; for Nvidia, it is another large buyer insulated from the commercial cycle.

Open models and robots

Two further pieces round out the announcement. AWS will offer Nvidia’s Nemotron family of open models through Amazon Bedrock and SageMaker, widening the menu of models customers can run on managed AWS services and reinforcing the industry’s drift toward open weights as a first-class option. And Amazon will adopt Nvidia’s Omniverse, Cosmos, Isaac, and Jetson stack for its warehouse robotics fleet — the simulation, world-model, and edge-compute tools that underpin physical AI.

The robotics tie-in connects the deal to Amazon’s own operational ambitions. The company runs one of the world’s largest fleets of warehouse robots, and building it on Nvidia’s physical-AI stack folds fulfillment automation into the same partnership that powers its cloud. It also puts AWS and Nvidia on the same side of the fast-moving race to commercialize robotics, where simulation-trained models are increasingly the differentiator.

What it means

This announcement is best understood as the supply-side confirmation of everything Nvidia told investors the same afternoon. When Nvidia guides to $108 billion next quarter and flags $279 billion in supply commitments, deals like this are what those numbers are made of: named customers, multi-year, spanning current and future silicon. AWS is effectively pre-committing a slice of its capital budget to Nvidia’s roadmap through 2028.

The winners are the two signatories, in complementary ways. Nvidia locks in one of its largest customers across three GPU generations plus a new CPU line, deepening the switching costs that make its position so durable. AWS secures priority access to scarce accelerators — the single hardest input to obtain in the current cycle — and gains a differentiated federal and physical-AI story to sell against Microsoft and Google. Both extend their reach into government workloads that competitors will now have to answer.

The tension worth watching is what the deal says about AWS’s own silicon. Amazon has invested heavily in Trainium and Inferentia accelerators and Graviton CPUs precisely to reduce its dependence on Nvidia and improve unit economics. Committing to 2 million more Nvidia GPUs and adopting Vera CPUs suggests that, for the frontier workloads customers actually demand today, in-house chips are a complement rather than a replacement. That is a rational choice while demand outruns supply — but it also means the cloud providers’ long-sought leverage over Nvidia remains, for now, mostly aspirational. The question for the next few quarters is whether any hyperscaler can shift enough volume to its own silicon to change that balance, or whether Nvidia’s roadmap keeps setting the pace the whole industry builds around.