AWS and NVIDIA to expand GPU supply with plan for 2 million additional accelerators and new AI infrastructure

Amazon Web Services and NVIDIA announced a broad expansion of their joint AI infrastructure, saying they will deliver 2 million additional graphics processing units, or GPUs, along with next-generation cloud capabilities aimed at supporting agentic and physical AI workloads. The update, reported by Yahoo Finance on Aug. 26, comes as customers push beyond traditional machine learning deployments toward systems that can plan, act, and coordinate tasks across software and the physical world.

In the announcement, AWS and NVIDIA positioned the buildout as part of a deeper integration across multiple layers of the AI stack, rather than a single hardware upgrade. The companies said the effort includes NVIDIA Vera CPUs, advanced networking, and NVIDIA-related open model components, along with technologies designed to support “physical AI,” a term generally used for AI systems connected to sensors, robotics, edge devices, or other real-world inputs and outputs.

The planned GPU capacity increase is the centerpiece of the news. By adding 2 million more GPUs, the companies are effectively targeting the compute bottleneck that often limits large-scale training and inference for frontier models and emerging agentic applications. The announcement did not spell out the delivery schedule, geographies, or whether the full figure is allocated to specific AWS regions, but it frames the incremental supply as a response to “accelerated” customer demand for AI infrastructure.

NVIDIA also referenced its model ecosystem in the integration. According to the report, the infrastructure effort includes “Nemotron Open Models,” which refers to NVIDIA’s open model work built to run on its platform and in partner environments. Open models are typically released with more transparent licensing and tooling so customers and developers can fine-tune, deploy, and adapt models more readily across different workflows, which can reduce friction for enterprises that want control over their AI pipelines.

On the networking side, the companies said the expanded infrastructure will include next-generation connectivity. For large AI systems, networking performance matters because modern training and inference can require many servers working together, with data and gradients moving frequently. While the announcement highlights “advanced networking,” it did not provide specific bandwidth, latency, or interconnect details in the report.

NVIDIA’s Vera CPUs were also cited as part of the platform approach. CPUs are the general-purpose processors that handle orchestration and control-plane tasks, while GPUs handle the heavy parallel compute for AI workloads. By pairing Vera CPUs with GPU acceleration and networking, the companies are indicating a more “balanced” system design intended to improve end-to-end throughput for complex AI services rather than optimizing only one component.

The companies did not disclose financial terms, contract sizes, or the specific configuration of the “next-generation infrastructure” beyond naming key elements such as GPUs, Vera CPUs, networking, and model-related components. They also did not provide a breakdown of how much capacity is intended for training versus inference, or which agentic or physical AI customer categories are expected to consume the new resources first. Any assessment of impact on near-term revenue or market share therefore remains speculative based on the public framing alone.