Developing an AI Factory Digital Twin with an Agent-Assisted Workflow

NVIDIA's developer blog published a walkthrough of an agent-assisted workflow for building AI factory digital twins on its Omniverse DSX Blueprint, part of NVIDIA DSX Sim. The blueprint ships reference digital twin assets, sample scripts and workflows built on OpenUSD and NVIDIA Omniverse libraries, giving developers a starting point for assembling facility and equipment models with SimReady assets and linking them to engineering simulation tools. The post describes a worked example: a campus site model based on the DSX reference design and inspired architecturally by NVIDIA's Santa Clara headquarters building, Voyager.
The inputs in that example fall into distinct roles. DSX 3D models supplied data halls, central utility buildings, substations, generators, cooling equipment, vehicles and vegetation. Photography of Voyager from aerial and ground views informed roof, facade, solar canopies and landscaping. PDF site plans described the campus arrangement and a more detailed first phase covering buildings, roads, utility areas and access. An indexed requirements collection included architectural drawings, form-factor studies, Basis of Design documents and room layouts. USGS aerial imagery and elevation data placed the campus in surrounding terrain.
The agent side is a frontier AI model running in an agent harness, the software layer that manages context, instructions and tool calls. Skills encode repeatable tasks and AI factory domain knowledge, and developers expose modeling and analysis operations as tools, including through the Model Context Protocol (MCP) where supported, while defining which changes the agent may propose or apply. In the example, the agent used USD Content Agents to generate geometry and materials for a solar canopy, then refined it against the Voyager references.
Changes were kept separate from source assets using additive USD layers. This is a core OpenUSD pattern: layers stack like transparent sheets, so new geometry and edits sit on top of the referenced originals rather than overwriting them, which lets teams inspect, revise or discard changes without damaging the base content. The campus was reviewed in the Omniverse RTX viewport, with USGS terrain data used for the landscape and elevation plan. NVIDIA is explicit that the canopy stayed conceptual and would need engineering checks on dimensions, structural properties and simulation suitability.
A second, proposed extension turns the campus model toward capacity planning: how to configure an AI factory hall inside a fixed power budget. The reference point is an illustrative NVIDIA Vera Rubin NVL72 inference comparison pitting MaxP, the highest static GPU-level power setting, against DSX MaxLPS, a framework that coordinates facility design, dynamic power allocation and performance-per-watt optimization. Normalized to 1 MW with a power usage effectiveness of 1.1, the scenario shows DSX MaxLPS capacity of 400 GPUs and aggregate token throughput at 1.35x the MaxP baseline. NVIDIA stresses these are reference assumptions for a proposed study, not outputs of the campus model, and that Vera Rubin testing is still in progress.
In that proposed workflow, engineers would evaluate feasible configurations using complete racks, including placement, electrical distribution, cooling connections and network resources. Systems Modeling Language (SysML) models could hold requirements linked to system elements; if a modeling tool exposes those relationships through an API, MCP or similar interface, the agent could query them to flag dependencies a proposed change might touch. The handoff would be a design revision bundled with requirements, equipment data and follow-up studies, with engineers reviewing before the agent updates the OpenUSD scene to match the chosen rack layout.
Simulation readiness is handled through SimReady Foundation specifications, which describe the asset data particular analyses require, and which let teams pick appropriate validators and check geometry, properties and connection data. The SimReady Central Library on Hugging Face is described as free and ungated with more than 2,000 simulation-ready OpenUSD assets, including OEM racks and power and cooling equipment, which could cut asset preparation work. AI factory pipeline samples cover CAD conversion, optimization, metadata and validation, and an agent can invoke these steps and collect their reports. The resulting handoff is an assembled scene tied to a specific revision with checks completed and gaps visible.
Once a scene exists, electrical loading and cooling studies use its geometry, equipment properties and system connections, with extra solver preparation where needed. Where supported solver connections exist, an agent can submit studies and retrieve results, associating findings with affected equipment and requirements. NVIDIA advises that each finding identify the tested revision and distinguish computed results from stored results or AI approximations, a sensible discipline: a twin that blurs simulation output with model guesses becomes hard to defend in design review.
NVIDIA also sketches hardware selections by task. DGX Spark is positioned for personal agent development and testing local models, paired with an RTX system for visualization and simulation. An RTX PRO Blackwell Workstation Edition GPU targets interactive design, asset preparation and supported simulations. DGX Station covers larger local AI workloads, optionally with an RTX PRO 6000 Blackwell Workstation Edition GPU for visualization. For shared development, the RTX PRO AI Factory reference architecture specifies eight RTX PRO 6000 Blackwell Server Edition GPUs per server and groups four servers into a scalable unit, sized by scene complexity, solver requirements and concurrent users.
The ecosystem layer spans several vendors. Dassault Systèmes' CATIA Magic uses SysML to model system elements and interfaces, check consistency and evaluate designs against architecture, and supports the Agent2Agent (A2A) protocol alongside MCP. PTC is working on an integration between Windchill, which manages version-controlled Creo and other CAD data, and Omniverse libraries, aimed at the handoff from managed equipment design to the twin. Cadence's Reality Digital Twin Platform couples equipment and power models with computational fluid dynamics to simulate MaxP, MaxQ and DSX MaxLPS, built on Omniverse libraries, SimReady assets and OpenUSD. Siemens' Simcenter STAR-CCM+ handles fluid flow and heat transfer analysis, and Siemens has developed DSX-aligned reference architectures.
Read together, the blueprint's practical value is the pipeline shape rather than any single asset: documents and imagery ground an agent, additive layers keep edits reversible, SimReady checks gate whether assets can be simulated, and solver integrations close the loop back into requirements. The limits are equally clear. Dimensional and structural correctness still depends on human checking against drawings and site constraints, the 1.35x throughput and 400-GPU figures are illustrative assumptions, and traceability only works if requirements tools expose their relationships programmatically. This suggests adoption will hinge less on agent capability than on whether a team's CAD, PLM and simulation stack can be wired up through MCP, A2A or vendor APIs.
Why it matters: Teams planning AI data centers now have a documented, OpenUSD-based route from site documents to an inspectable twin with simulation hooks, which lowers the integration work for developers building facility design tooling. The agent-generated geometry is explicitly conceptual, so engineering review and validated asset data remain the gate before anyone relies on results (an inference, not a claim from NVIDIA). Product teams in CAD, PLM and simulation should note the MCP and A2A interfaces named here, since those appear to be how agents will reach their tools.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
Building an AI factory digital twin requires developers to bring site plans, equipment assets, and engineering requirements into a consistent model that teams...