Layer 1: The Reportable Facts

AWS and Nvidia announced on August 26, 2026 that they plan to deploy 2 million additional Nvidia GPUs across AWS global infrastructure in 2027 and 2028. Amazon’s release names Nvidia Blackwell Ultra, Rubin, and Rubin Ultra GPUs and says the capacity will support workloads including agentic AI, scientific discovery, enterprise automation, and physical AI. Bloomberg separately reported that Amazon will add the 2 million Nvidia GPUs over the next two years despite Amazon’s own chipmaking work, while TechCrunch reported that the announcement followed an earlier plan for more than 1 million Nvidia GPUs across AWS infrastructure starting in 2026.

The verified package reaches beyond accelerators. AWS and Nvidia said they are working to bring Vera CPU-based infrastructure to AWS; extend NVLink Fusion with Nvidia custom high-bandwidth memory for next-generation Trainium; integrate Nvidia GPU-based and Trainium-based EC2 instances with the AWS Nitro System and Elastic Fabric Adapter; continue supporting Nvidia Nemotron open models on Amazon Bedrock and Amazon SageMaker; and accelerate data processing and vector indexing through Nvidia CUDA-X libraries on Amazon EMR and Amazon OpenSearch.

The federal component is explicit. 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. Amazon’s release says the collaboration is meant to support workloads classified at Impact Level 6 and above, which places part of this capacity inside the national-security compute stack rather than the ordinary commercial cloud story.

The physical-AI component is also explicit. Amazon Robotics is working with Nvidia’s physical-AI platform, including Jetson, Omniverse libraries, and Isaac, across simulation, synthetic data, robot training, route optimization, functional safety, and real-to-sim validation. In other words, the same AWS-Nvidia infrastructure deal touches both cloud AI and the robot training loop.

Layer 2: The System Read

Inference: this is less a capacity announcement than a supply-chain positioning move. AWS is converting Nvidia allocation into a 2027-2028 control layer for customers that need to know whether their future models, agents, data systems, or robots will have access to top-tier compute. In a market where GPUs are the bottleneck, a multi-year allocation becomes a product feature, a sales weapon, and a strategic moat.

Inference: AWS is not choosing between Nvidia and Trainium; it is designing a mixed compute stack. The Nvidia GPUs preserve access to the dominant AI accelerator ecosystem, while Trainium gives Amazon leverage on cost, differentiation, and bargaining power. The NVLink Fusion and NVHBM work matters because it points toward heterogeneous racks where AWS custom silicon and Nvidia components coexist inside the same AI-factory architecture.

Inference: the U.S. government piece changes the competitive meaning of the deal. Once secure AI factories become part of federal and national-security workloads, cloud infrastructure becomes sovereign-adjacent industrial infrastructure. That favors providers able to combine chips, networking, security boundaries, compliance, logistics, and long-term procurement into one package.

Inference: the robotics language is not decorative. AWS and Nvidia are linking cloud-scale simulation, synthetic data, model training, and warehouse robotics into one physical-AI flywheel. If successful, the same infrastructure used to train agents in software will increasingly be used to train machines that operate in physical environments.

Layer 3: What To Watch Next

Watch whether AWS turns the 2027-2028 GPU plan into named instance families, committed customer capacity, or dedicated AI-factory regions. The practical question is not only how many GPUs AWS receives, but how that supply is partitioned among frontier labs, enterprise AI customers, robotics workloads, and public-sector systems.

Watch the Trainium-Nvidia integration path. If NVLink Fusion and NVHBM make Trainium easier to deploy inside rack-scale architectures alongside Nvidia GPUs, AWS could reduce customer fear around custom silicon by presenting Trainium as part of a familiar Nvidia-adjacent fabric rather than a separate island.

Watch federal procurement signals. The 100,000-GPU U.S. government plan could pull other hyperscalers into a race to package secure AI factories for defense, intelligence, and civilian agencies. The next signal will be whether these systems become named programs, classified-cloud expansions, or broader government AI platforms.

Watch power, siting, and supply-chain constraints. A 2 million-GPU expansion implies pressure on data-center campuses, power availability, networking gear, cooling systems, and memory supply. The AI-factory race will be measured not just in chips ordered, but in megawatts, interconnect, secure facilities, and delivery schedules.

Conclusion

The headline number is 2 million additional Nvidia GPUs, but the more important move is architectural. AWS is using a 2027-2028 Nvidia commitment to pre-build the rails for the next phase of AI demand: agents that need persistent inference, governments that need secure sovereign-grade compute, enterprises that need production data pipelines, and robots that need simulated and real-world training loops. If the plan lands, AWS will have turned GPU scarcity into a multi-market distribution system.