AWS to add 2 million more NVIDIA GPUs as it expands next-generation AI infrastructure

Amazon’s cloud unit, AWS, and NVIDIA are expanding their partnership aimed at building next-generation infrastructure for AI workloads. In an update published by Amazon, AWS said it will deploy 2 million additional NVIDIA GPUs across its global infrastructure during 2027 and 2028, paired with deeper collaboration across CPUs, networking, and robotics.

The companies also described new compute options for AI applications. AWS said NVIDIA’s Vera CPUs are coming to AWS, adding another form of processing alongside GPU-based acceleration. For developers, that means certain workloads can be mapped to different processor types depending on what is most efficient for training and running AI models, including what Amazon calls “agentic AI,” software that can take actions toward a goal.

In addition to raw capacity and compute options, Amazon emphasized the underlying cloud platform stack that connects AI systems at scale. It said NVIDIA GPU-based and Trainium-based EC2 instances, including those using NVLink Fusion, are built on AWS’s Nitro System and connected through Elastic Fabric Adapter, or EFA. Amazon framed that as supporting security, reliability, and network performance for customers running AI training and inference at high throughput.

AWS also tied the expanded hardware and platform roadmap to government-focused deployments. The companies said they will build “AI factories” for the U.S. government, including 100,000 GPUs on secure AWS infrastructure. Amazon did not provide additional details in the announcement about which agencies or procurement vehicles are involved, but it positioned the effort as an engineered supply of compute for large-scale AI development and use.

For model availability, AWS said NVIDIA’s Nemotron family of open models will be offered on AWS as fully managed, serverless services through Amazon Bedrock. It also said customers can access the same Nemotron models via Amazon SageMaker for teams that want to run and fine-tune models on their own infrastructure. In practical terms, Bedrock is meant to reduce the operational burden of deploying and scaling models, while SageMaker is a broader toolkit for building, training, and adjusting models more directly.

Amazon also highlighted ways it intends to speed up data and analytics processes that sit around AI workloads. It said GPU-accelerated data processing in Amazon EMR using NVIDIA’s cuDF can deliver up to 3.7 times faster processing speeds and 30% better price-performance for Apache Spark workloads versus CPU-only configurations. Separately, AWS said GPU-accelerated vector indexing on Amazon OpenSearch Service can deliver up to 9 times faster indexing at a quarter of the cost, relative to CPU-based approaches, aiming at better performance for applications that rely on vector search, such as semantic retrieval and question answering systems.

On the robotics front, Amazon said its robotics business is integrating NVIDIA’s physical AI platform into Amazon Robotics. The company described the stack as including Jetson, Omniverse, and Isaac, and said the integration is intended to accelerate next-generation warehouse automation by using simulation, synthetic data generation, and real-world validation. The operational goal is to help robots and warehouse systems learn from high-quality simulated scenarios before testing in live environments, which can reduce time and cost during deployment.

Beyond the AI-specific items, the same Amazon release referenced additional infrastructure and security initiatives within its broader portfolio, including “Ring introduces TAKE,” described as a new industry standard for default encryption and control. While not central to the NVIDIA-GPU announcement, it reinforces that AWS and Amazon are positioning platform security and controlled access as part of their larger infrastructure messaging.

Amazon did not disclose financial terms, customer names, or the expected timetable for the U.S. government “AI factories” beyond stating the general buildout for the GPU deployments. It also did not specify whether the additional 2 million GPUs are tied to particular EC2 instance families or how capacity will be allocated between training and inference workloads. As a result, the magnitude of the near-term impact on AWS revenue, margins, or customer migration timelines is not determinable from the announcement alone.