AWS and Nvidia outline deployment of more than 1 million GPUs

Nvidia is a public record based on article evidence, entity context, event links, and relationship context.

• AWS and Nvidia announced an expanded collaboration in March 2026. AWS said it would begin adding more than one million Nvidia GPUs across its global cloud regions in 2026.
• Reuters separately reported, citing Nvidia executive Ian Buck, that sales of one million GPUs would start in 2026 and extend through 2027; the companies did not disclose financial terms.
• The million figure counts GPUs, not servers, racks or complete systems. The wider collaboration also covers Groq inference accelerators, ConnectX and Spectrum networking, software and service integrations.

What AWS and Nvidia announced

At GTC on 16 March 2026, AWS described an expanded collaboration with Nvidia rather than a completed installation. Its own announcement says that, starting in 2026, AWS will add more than one million Nvidia GPUs, including Blackwell and Rubin architectures, across global cloud regions. Nvidia used similar language: deployment of more than one million GPUs would begin in 2026, alongside LPUs and other Nvidia AI infrastructure.

Three days later, Reuters supplied a separate commercial boundary. Ian Buck, Nvidia’s vice-president for hyperscale and high-performance computing, said Nvidia would sell one million GPUs to AWS, with sales starting in 2026 and extending through 2027. Reuters said AWS had reached a purchase agreement, but neither company disclosed its value or a detailed delivery schedule. The official phrase “more than one million” and Reuters’ one-million sales figure should not be rewritten as “up to one million,” nor treated as proof that all units are already installed.

A GPU is not a rack, system or every chip in the package

The unit boundary matters. A GPU is one accelerator processor. Blackwell and Rubin describe GPU architectures, while a rack-scale Nvidia system can contain many GPUs together with CPUs, switches, network interfaces, memory and software. Nvidia’s GTC material also named Groq 3 LPUs for low-latency inference. Reuters said the transaction extended beyond the one million GPUs to a wider chip mix, including Groq chips and Nvidia’s Spectrum networking products, and that ConnectX and Spectrum-X gear would enter AWS data centres.

Those components cannot be added together and called one million systems. Nor does the announcement establish how many physical servers or racks AWS has ordered. Nvidia presents Vera Rubin as a multi-component platform; that vendor description explains why the partnership is broader than a GPU purchase, but it does not supply an independently verified deployed-system count.

Timing and availability remain execution questions

“Starting in 2026” identifies the beginning of deployment. Buck’s statement that sales extend through 2027 gives an outer sales period, not a region-by-region completion certificate. AWS did not publish a shipment ledger, installation total, customer allocation, utilisation rate or date when every announced GPU would be generally available. A signed purchase and an infrastructure plan therefore sit upstream of delivered, powered, networked and customer-usable capacity.

The execution chain includes accelerator and memory supply, server assembly, high-speed fabrics, power, cooling, data-centre space, software qualification and EC2 service launch. A shortfall at any stage can separate purchased silicon from sellable cloud capacity. Performance and efficiency numbers in AWS and Nvidia material are vendor claims tied to particular configurations; customers still need workload-specific benchmarks.

AWS is expanding a mixed accelerator strategy

The collaboration does not mean AWS has abandoned its own infrastructure. AWS’s announcement pairs Nvidia NIXL with Elastic Fabric Adapter and explicitly discusses workloads spanning Nvidia GPUs and AWS Trainium. Nitro, EFA and Trainium remain AWS-controlled layers, while Nvidia supplies accelerators, networking and software. The strategic issue is therefore how AWS allocates workloads and capital across Nvidia capacity and its own silicon, not simple replacement of one by the other.

For Nvidia, the agreement provides a large, disclosed demand signal, but no disclosed contract value or recognised revenue schedule. For AWS, it enlarges the planned menu of accelerated compute, but does not guarantee a particular customer price, reservation or service level. Claims about Nvidia market share, AWS revenue or customer advantage require evidence beyond the unit announcement.

What to watch

The strongest follow-up evidence will be dated AWS instance launches and region tables, Nvidia and Amazon financial disclosures, confirmed deliveries, and published details on Blackwell, Rubin, Groq, ConnectX and Spectrum configurations. Power and cooling readiness, software availability and the balance between Nvidia GPUs and Trainium also matter. Until those signals arrive, the defensible statement is that AWS committed to a million-scale Nvidia expansion whose deployment began in 2026 and whose reported sales run through 2027—not that one million complete AI systems are already operating.