아마존 웹 서비스는 2027~2028년 클라우드 인프라 구축을 위해 엔비디아 GPU를 추가로 200만 대 도입할 계획이라고 밝혔다.

AWS and Nvidia unveiled their August 26 plan to deploy 2 million additional GPUs across AWS’s global infrastructure in 2027 and 2028. This future capacity will utilize Nvidia Blackwell Ultra, Rubin, and Rubin Ultra products; the announcement does not indicate that these accelerators have already been delivered or entered service.

The expansion follows AWS’s earlier commitment to add more than 1 million Nvidia GPUs across its global cloud regions starting in 2026. The two initiatives operate on different deployment timelines, with the newer commitment supplementing rather than replacing the initial pledge.

The earlier plan established a lower bound rather than a fixed order size, citing “more than 1 million” units. With the new addition of 2 million GPUs, the combined published figures yield a lower bound exceeding 3 million units if both programs proceed as outlined. This arithmetic justifies the headline’s “triples” shorthand, though it does not confirm an exact procurement ratio. Because the original commitment may exceed its stated minimum, the precise baseline and exact multiple remain undefined.

TechCrunch’s coverage of the threefold expansion noted the new commitment arrived roughly five months after the first, while highlighting that neither company disclosed financial terms. Referring to the arrangement as an “order” serves as convenient shorthand; formally, both companies state that AWS plans to deploy the capacity.

The two-year timeframe represents a deployment window rather than a single delivery date. AWS has not released a region-by-region rollout schedule, detailed allocations among Blackwell Ultra, Rubin, and Rubin Ultra architectures, or launch dates for customer-facing EC2 instances powered by the expanded fleet. Scaling deployment at this magnitude requires coordinated manufacturing, logistics, data center readiness, cluster integration, and service validation. The public commitment outlines the target volume and general timeline but does not guarantee that all GPUs will arrive at the start of 2027 or go live simultaneously across regions.

Nvidia’s collaboration overview similarly frames the initiative as a planned deployment across AWS’s global footprint, omitting shipment milestones, procurement structures, or ownership details for the installed hardware. Beyond raw compute, the expansion functions as a comprehensive systems-integration program encompassing processors, memory interconnects, and networking. AWS and Nvidia intend to introduce Vera CPU-based infrastructure to the cloud platform, expand NVLink Fusion work alongside Nvidia high-bandwidth memory, and advance Spectrum networking solutions optimized for large-scale GPU clusters.

The underlying architecture bridges Nvidia components with AWS-native technology. Both Nvidia GPU-based and AWS Trainium-based EC2 instances leverage the AWS Nitro System and Elastic Fabric Adapter, while Nvidia and Amazon’s Annapurna Labs are scaling their joint efforts on rack-scale systems. Neither company has disclosed the projected volume of Vera CPUs or the specific configurations and regions slated for deployment.

At the software layer, the partnership encompasses Nemotron open models via Amazon Bedrock and SageMaker, GPU-accelerated data processing on Amazon EMR utilizing cuDF, and vector indexing on Amazon OpenSearch with cuVS. Each integration operates under independent availability and performance parameters; none serve as proof that the newly planned GPU capacity is currently operational.

Physical AI represents another pillar of the agreement. An Associated Press report on the AWS–Nvidia initiative independently documented both the additional compute capacity and the integration of Nvidia technology into Amazon’s warehouse-robotics ecosystem. Amazon Robotics will deploy Nvidia’s Jetson, Omniverse, and Isaac platforms to support simulation, synthetic-data generation, robot training, route optimization, functional safety, and validation workflows. The broader initiative also encompasses planned AI factories for the U.S. government, allocating 100,000 GPUs on secure AWS infrastructure to support federal and national-security workloads.

No purchase price, payment schedule, unit cost, or projected revenue impact has been disclosed for the additional capacity. Estimating deal value by applying retail GPU pricing to the headline figure would be misleading, as enterprise infrastructure agreements typically bundle accelerators, servers, networking, memory, software, and support services under customized commercial terms.

Current disclosures also leave unresolved how the government allocation intersects with the wider deployment, how capacity will be partitioned across GPU generations, or which AWS instance families will ultimately surface the resources to customers. These operational specifics will dictate when the commitment transitions from planned infrastructure to production-ready cloud capacity.

As of August 28, the verified trajectory confirms a multiyear expansion: the initial program commences in 2026, while the supplementary fleet targets 2027–2028. Future shipment milestones, designated regions, instance releases, model distribution plans, and subsequent financial disclosures will serve as the next indicators of execution. Until then, the aggregated figure reflects planned additions rather than AWS’s currently installed inventory.