AWS and NVIDIA are making their AI infrastructure partnership much bigger.
Amazon Web Services plans to deploy 2 million additional NVIDIA GPUs across its global infrastructure. The rollout is scheduled for 2027 and 2028. It will cover AI factories and other AWS infrastructure around the world.
The scale alone grabs attention. But this isn’t simply a giant GPU order.
AWS and NVIDIA are also expanding their work across CPUs, networking, AI models, data processing, and robotics. NVIDIA Vera CPUs are coming to AWS as part of the deal. The companies also plan secure AI infrastructure for the U.S. government.
AWS Is Adding Another 2 Million NVIDIA GPUs
AWS already had an enormous GPU expansion underway. At NVIDIA GTC 2026, Amazon announced plans to add more than 1 million NVIDIA GPUs starting this year. Demand has since moved beyond those expectations. Now AWS plans another 2 million GPUs for 2027 and 2028. The new capacity will spread across its global infrastructure. It will support agentic AI, scientific research, enterprise automation, and physical AI.
Blackwell Ultra and Rubin Are Part of the Expansion
This isn’t just about increasing the number of chips. AWS plans to deploy newer generations of NVIDIA hardware. That includes Blackwell Ultra, Rubin, and Rubin Ultra GPUs. AWS will also expand its existing Blackwell capacity. Amazon EC2 G7 instances will use NVIDIA RTX PRO 4500 Blackwell Server Edition GPUs. AWS says G7 delivers 4.6 times higher AI inference performance than the previous G6 generation.
NVIDIA Vera CPUs Are Coming to AWS
GPUs get most of the attention in AI infrastructure. CPUs still have plenty of work to do. AWS plans to introduce infrastructure based on NVIDIA’s Vera CPU. Vera targets workloads behind agentic AI and reinforcement learning. Those tasks include code execution, tool use, analytics, sandboxing, and orchestration. In short, agents need more than GPUs when they start performing complex actions. Vera is designed to handle more of that supporting work.
AWS Isn’t Walking Away From Its Own AI Chips
There’s an interesting wrinkle here. Amazon is simultaneously pushing its own Trainium AI chips. So the NVIDIA expansion doesn’t mean AWS is choosing one side. Instead, Amazon wants customers to mix different compute options. AWS says customers can use NVIDIA GPUs, Trainium chips, or both. That strategy could become important as AI infrastructure grows more specialized. One chip architecture may not make sense for every workload.
Trainium Is Getting Closer to NVIDIA’s Technology
AWS and NVIDIA are also connecting their chip ecosystems more tightly. Next-generation Trainium chips will support NVIDIA NVLink Fusion. Amazon’s Annapurna Labs will also work with NVIDIA’s custom high-bandwidth memory technology. The combination could improve memory performance and efficiency. More importantly, it creates a path for Trainium and NVIDIA GPUs to operate within common rack-scale systems. AWS gets to keep developing its own silicon without building an entirely separate infrastructure world around it.
The U.S. Government Is Getting a 100,000-GPU AI Push
Part of the partnership goes straight into government AI. AWS and NVIDIA plan AI factories for U.S. federal agencies. The companies expect those systems to include 100,000 NVIDIA GPUs. They will run on secure AWS infrastructure. The environment will support federal and national-security workloads at Impact Level 6 and above. That brings the AWS-NVIDIA partnership into some of the most sensitive areas of government computing.
AWS Wants Faster AI Data Processing Too
Training models isn’t the only place GPUs matter. AWS and NVIDIA are accelerating the data pipelines around AI applications. Amazon EMR can use NVIDIA cuDF for GPU-accelerated processing. AWS says this delivers up to 3.7 times faster processing for Apache Spark workloads. It also claims 30% better price-performance than CPU-based configurations. Amazon OpenSearch is getting GPU acceleration as well. AWS says vector indexing can run up to nine times faster at one-quarter of the cost.
NVIDIA Nemotron Models Stay Inside the AWS AI Stack
The partnership extends into models, not just hardware. NVIDIA’s Nemotron open models are available through Amazon Bedrock. Customers can access them as managed serverless models. They can also use Amazon SageMaker when they want more control over deployment and fine-tuning. This gives AWS another model family alongside the growing selection available through Bedrock. NVIDIA, meanwhile, gets another route into enterprise AI workloads.
Amazon Robotics Is Getting More NVIDIA Physical AI
Robotics is another big piece of the agreement. Amazon Robotics is integrating NVIDIA’s physical AI platform into its development work. That includes Jetson, Omniverse, and the Isaac robotics platform. The companies will work across simulation, synthetic data, training, and route optimization. Real-world validation will also play a role. The obvious testing ground is Amazon’s huge warehouse network. That’s where physical AI can move from demos into daily operations.
Two Million GPUs Says Plenty About Where AI Is Heading
The number is difficult to ignore.
AWS announced more than 1 million NVIDIA GPUs earlier in 2026. Months later, it added plans for another 2 million. Amazon says customer demand exceeded its earlier expectations.
That gives a pretty clear signal about the next phase of AI. The industry is moving beyond chatbot experiments. Companies are building agents, automated workflows, scientific systems, and robots. Those applications need enormous amounts of compute.
AWS also isn’t betting everything on NVIDIA. Trainium remains central to Amazon’s strategy. Its custom chip business has already become a major operation.
What we’re seeing instead is a more complicated AI infrastructure market. NVIDIA GPUs, Amazon silicon, specialized CPUs, high-speed networking, and new memory systems are starting to blend together.
The AI race isn’t only about who has the smartest model anymore.
Increasingly, it’s about who has enough infrastructure to actually run them.
Sources
- Amazon — AWS and NVIDIA expand partnership for next-gen AI infrastructure
https://www.aboutamazon.com/news/aws/aws-nvidia-2-million-gpus-ai - Amazon Press Center — AWS and NVIDIA to Deliver 2 Million Additional GPUs and Next-Generation Infrastructure for Agentic and Physical AI
https://press.aboutamazon.com/aws/2026/8/aws-and-nvidia-to-deliver-2-million-additional-gpus-and-next-generation-infrastructure-for-agentic-and-physical-ai
