NVIDIA Outlines AI-Agent Workflow for AI Factory Digital Twins

NVIDIA outlined an agent-assisted process for assembling AI factory digital twins, giving engineering teams a way to organize facility, equipment and design data before review and simulation.

The Omniverse DSX Blueprint supplies reference assets and sample workflows for a scene built with OpenUSD, which lets teams combine 3D elements and related information. The campus example draws on DSX geometry, Voyager architectural references, site plans, indexed engineering records, and aerial and elevation data for the location. An agent sketched a conceptual solar canopy from visual references; the added geometry sits in USD layers separate from the original files.

Developers can connect an agent to modeling and analysis tools and choose which edits it may propose or apply. Engineers still need to confirm dimensions and equipment specifications, check system connections, and set simulation conditions before treating the scene as an engineering model.

NVIDIA also makes a free SimReady Central Library available through Hugging Face, with more than 2,000 OpenUSD assets covering items such as server racks and power and cooling equipment. The DSX Blueprint describes a 50-acre reference site and sample thermal and electrical analyses, and requires customization before production use.

The digital twin serves as a shared design model, not proof of a completed engineering study. By bringing facility geometry, equipment data and technical records into one scene, teams can prepare analyses of power and cooling changes. Those studies still depend on accurate dimensions, verified specifications and suitable simulation settings.

An accompanying Vera Rubin NVL72 inference comparison sets reference conditions for a proposed study: 400 GPUs, a 1-megawatt power basis and a power usage effectiveness (PUE) of 1.1. For aggregate token throughput, DSX MaxLPS is assigned a rate 1.35 times the MaxP reference. These are assumptions for the proposed study, not results from the campus model, and Vera Rubin testing remains in progress.

The example sets up inputs for analysis but does not report a completed, validated simulation or customer deployment. The example focuses on organizing design inputs for engineering review as part of the broader AI-factory digital twin and agent work.