• Reuters reported on 19 March 2026 that NVIDIA would sell one million GPUs to Amazon Web Services (AWS), with sales beginning in 2026 and extending through 2027. The transaction also includes Spectrum networking and Groq products, which sit outside the GPU count.
  • AWS and NVIDIA had announced three days earlier that deployment of more than one million NVIDIA GPUs would begin across AWS Regions in 2026. A sales window and a deployment plan do not establish shipments, installed capacity, service availability or utilization; those stages need separate evidence.

Two disclosures describe different parts of the same programme

AWS announced on 16 March that it would add more than one million NVIDIA GPUs, including Blackwell and Rubin architectures, across its global cloud regions starting in 2026. NVIDIA described the same infrastructure expansion as covering more than one million GPUs, Groq 3 language processing units and Spectrum networking.

On 19 March, NVIDIA vice-president Ian Buck gave Reuters a commercial timeline: sales of one million GPUs to AWS would begin in 2026 and run through 2027. The later disclosure adds a sales horizon; it does not convert the earlier deployment announcement into a completed delivery count.

The package is broader than one million identical “AI chips”

The one-million unit refers to GPUs. The announced GPU mix spans Blackwell and Rubin, while AWS also plans RTX PRO 4500-based EC2 instances. Buck said the transaction includes Spectrum networking and Groq chips; NVIDIA’s announcement identifies Groq 3 LPUs for low-latency inference. These components have different functions and should not be added together as one chip total.

The wider collaboration also covers NIXL data movement over AWS Elastic Fabric Adapter, Nitro-based instances, data-processing integrations and Nemotron models on Bedrock. Those technical announcements help explain the operating stack, but the reviewed sources do not show that every integration is a line item in the reported purchase.

Sale, delivery, installation and availability are not interchangeable

A purchase commitment can precede shipment; shipment can precede arrival, installation and cluster integration; installed hardware can precede a generally available EC2 service; available capacity can remain reserved, constrained or unused. The disclosures provide programme scale and timing, not a dated conversion through each stage.

Amazon’s first-quarter release illustrates the distinction. It said more than 2.1 million AI chips had landed over the previous 12 months, more than half of them Trainium, and separately said one million-plus NVIDIA GPUs had been announced for deployment starting in 2026. That 2.1 million figure covers Amazon’s broader silicon intake and is not a delivered count for this NVIDIA programme.

Control is split between the supplier and the cloud operator

NVIDIA controls product roadmaps, production and the GPU, LPU, networking and software stack it supplies. AWS controls procurement, region selection, data-centre readiness, power, cooling, network construction, Nitro and EFA integration, instance design, pricing, quotas and the date customers can actually obtain capacity.

Rubin adds a dated dependency. NVIDIA says Rubin is in production and partner products will be available in the second half of 2026, naming AWS among the first cloud providers expected to deploy Rubin instances. That schedule supports a future path, not a finding that Rubin capacity is already live on AWS.

What the evidence proves—and what it does not

Reuters provides a named-executive account of the sales window and mixed transaction. AWS and NVIDIA provide primary evidence for the announced architecture, integrations and deployment start. Amazon’s later results preserve the distinction between chips already landed and the separately announced NVIDIA deployment.

The reviewed sources do not disclose contract value, generation-by-generation quantities, shipment or installation totals, regional allocation, programme-wide customer availability, utilization, customer reservations, performance measured independently or revenue recognition. Claims that the deal by itself excludes startups, resolves GPU scarcity or proves a market forecast are unsupported and have been removed.

What to watch next

  • Dated shipment, installation and EC2 general-availability disclosures by AWS Region.
  • A reconciled count of ordered, delivered, installed, live and customer-available GPUs.
  • Rubin partner availability in the second half of 2026 and any schedule change.
  • The mix of Blackwell, Rubin, RTX PRO, Groq 3 and Spectrum products, with units kept separate.
  • Contract value, capital spending, NVIDIA revenue recognition, AWS pricing, quotas and utilization.
  • How NVIDIA capacity complements rather than replaces AWS Trainium and Inferentia.

Sources