O que aconteceu
AWS and NVIDIA announced an expanded strategic collaboration centered on increasing AI computing capacity and integrating more of NVIDIA’s hardware and software with AWS services. The companies say AWS plans to deploy 2 million additional NVIDIA Blackwell Ultra, Rubin and Rubin Ultra GPUs across its global infrastructure in 2027 and 2028, including AI factories and capacity for agentic AI, scientific discovery, enterprise automation and physical AI workloads.
AWS and NVIDIA said they are expanding a 16-year collaboration to meet what they describe as accelerating demand for AI infrastructure. The central commitment is a plan for AWS to deploy 2 million additional NVIDIA GPUs across its global infrastructure during 2027 and 2028. The planned systems include NVIDIA Blackwell Ultra, Rubin and Rubin Ultra GPUs, with capacity intended for workloads involving AI agents, scientific discovery, enterprise automation and physical AI. Because the announcement uses future-tense language, it describes a planned deployment rather than a completed installation or currently available inventory.
The companies also said AWS will expand its NVIDIA Blackwell capacity, including RTX PRO 4500 Blackwell Server Edition GPUs for Amazon EC2 G7 instances. NVIDIA and AWS state that G7 instances deliver 4.6 times the AI inference performance and 2.1 times the graphics performance of previous-generation G6 instances. Those are claims in the companies’ release; the source does not identify the test conditions, workloads, pricing or independent validation behind the comparisons. AWS is described as the first major cloud provider to offer instances accelerated by the RTX PRO 4500, according to the release.
The broader collaboration covers several layers of the AI stack. AWS and NVIDIA said they are working to bring NVIDIA Vera CPU-based infrastructure to AWS, extend NVLink Fusion with NVIDIA’s custom high-bandwidth memory technology in cooperation with memory suppliers, and integrate NVIDIA’s platform with the AWS Nitro System and Elastic Fabric Adapter. They also plan to continue offering NVIDIA Nemotron open models through Amazon Bedrock and Amazon SageMaker, accelerate data processing and vector indexing through NVIDIA CUDA-X libraries on Amazon EMR and Amazon OpenSearch, and advance robotics workloads through Amazon Robotics’ use of NVIDIA’s physical AI platform.
Leia a fonte primária: nvidianews.nvidia.com ↗
Por que isso importa
The announcement would expand the cloud infrastructure available for large-scale AI development and deployment, but it is primarily a forward-looking plan rather than evidence that the capacity has already been delivered. It also links AI infrastructure to government workloads, enterprise data processing and robotics, extending the partnership beyond GPU instances alone.
If delivered, the additional capacity could affect how quickly organizations move AI projects from experimentation into production. The release specifically frames the expansion around agentic systems, which may require repeated model calls and supporting data infrastructure, and physical AI, which can require simulation, training and validation workloads. It also points to scientific and enterprise uses, suggesting that the intended audience extends beyond a small number of frontier laboratories. The source does not establish how much of the planned capacity will be available to ordinary cloud customers or how access will be allocated.
The announcement is significant because it combines hardware supply, cloud deployment and software integration in one commercial arrangement. AWS customers would, in principle, be able to use NVIDIA GPUs alongside AWS custom silicon and NVIDIA Vera CPUs, while NVLink Fusion could connect NVIDIA and Trainium components within a common rack-scale architecture. The companies describe this as a way to support heterogeneous AI infrastructure. That may give developers more options for matching different workloads to different processors, but the source does not provide independent benchmarks comparing mixed configurations with alternatives or explain the engineering limits of moving workloads between them.
The government component raises the public-interest stakes. AWS and NVIDIA said they plan to build AI factories for the U.S. government, including 100,000 GPUs on secure AWS infrastructure for federal and national-security workloads at Impact Level 6 and above. The release does not name agencies, programs, deployment sites, budgets, timelines or the specific tasks these systems would perform. It therefore establishes a stated infrastructure commitment, not evidence of a particular national-security deployment or an assessment of its consequences. The planned use also makes security, procurement oversight and the handling of sensitive workloads important areas for further reporting.
O que assistir a seguir
The key questions are whether the planned GPU capacity arrives on schedule, where it is deployed, what customers can access and at what cost, and whether the companies’ performance and efficiency claims hold in independent production use. The release does not provide a detailed implementation timetable, customer commitments, power requirements or specific federal programs.
First, watch execution against the 2027–2028 target. The source does not break the 2 million GPUs into annual milestones, regions, data-center sites or specific product quantities beyond naming the Blackwell Ultra, Rubin and Rubin Ultra families. It also does not say how much capacity has been ordered, financed, installed or reserved for customers. Manufacturing, packaging, networking, power availability and regulatory or export-control changes could affect the final scale, but the release offers no project-level evidence on those issues.
Second, examine whether the claimed efficiency and performance translate into real workloads. AWS and NVIDIA report up to 3.7 times faster data processing and 30% better price performance for certain GPU-accelerated Amazon EMR configurations, as well as up to 9 times faster vector indexing at one-quarter of the cost on Amazon OpenSearch. The source does not specify the datasets, baselines, workload shapes, utilization rates or total operating costs. Independent testing would be needed to determine how representative those figures are for customers running retrieval-augmented generation, semantic search, large-scale analytics or other production applications.
Finally, watch the practical rollout of the model, security and robotics pieces. The release says Nemotron models will be available through Bedrock and SageMaker, but it does not give new model names, release dates, pricing or regional availability. It says Amazon Robotics and NVIDIA will work across simulation, synthetic data generation, robot training, route optimization, functional safety and real-to-sim validation, but does not announce a specific robot, facility, deployment result or safety evaluation. The most important unknown is whether these integrations produce measurable customer benefits and verifiable safeguards at the scale the companies describe.
