NVIDIA published a developer walkthrough of an agent-assisted workflow for building an AI factory digital twin. A frontier model runs inside an agent harness that carries instructions, skills and tool calls, including Model Context Protocol interfaces where supported, and the developer defines which changes the agent may propose or apply. Starting from the Omniverse DSX Blueprint reference design, the team supplied site plans, equipment data, architectural references from NVIDIA's Voyager building and USGS aerial imagery for terrain placement. The agent used USD Content Agents to generate the geometry and materials for a solar canopy, and edits were kept in additive USD layers so source assets stayed untouched. NVIDIA states that dimensional accuracy still depends on checking the model against drawings and specifications plus engineering review. The capacity figures in the post — a 1 MW basis, PUE of 1.1, 400 GPUs and 1.35x aggregate token throughput against a static maximum-power baseline — are described as reference assumptions for a proposed study, not outputs of the campus model.
Our takeThe honest part is the boundary: the agent turns grounded inputs into geometry and placement proposals, and every dimensional claim still routes through a drawing check and an engineer. That makes it useful for exploring alternatives rather than signing off a design. The 1.35x figure is a stated assumption, not a measurement, so it belongs to the study design and says nothing about how well the agent modeled anything.