The NVIDIA Omniverse DSX Blueprint and OpenUSD-based libraries provide a reference foundation for building editable digital twins of AI factory campuses. The approach brings site plans, equipment assets and engineering requirements into a consistent OpenUSD scene that teams can inspect, revise and prepare for simulation. Frontier AI models, deployed as agents, can orchestrate inputs, propose design changes and invoke asset-preparation and simulation workflows while engineers verify dimensions and system behavior.
Purpose and significance
Creating a digital twin for an AI factory requires coordinated input from architecture, power, cooling and compute disciplines. A shared, editable model enables exploration of layout alternatives, incorporation of engineering feedback and preparation for downstream analyses that inform design decisions.
DSX Blueprint and input sources
The NVIDIA Omniverse DSX Blueprint (part of NVIDIA DSX Sim) supplies reference digital twin assets, sample scripts and workflows to help developers assemble facility and equipment models using SimReady assets and integrate them with engineering tools. In the example workflow, the project combined these input types:
- DSX 3D models: data hall, central utility buildings, substations, generators, cooling equipment, vehicles and vegetation.
- Architectural references: images of NVIDIA’s Santa Clara headquarters, Voyager, to guide roof, facade, solar canopies and landscaping design choices.
- Site plans (DSX Reference Design): PDFs describing campus arrangement and a first-phase layout of buildings, roads and utility areas.
- Requirements documents: an indexed collection of architectural drawings, Basis of Design documents, room layouts and related records.
- Geographic data: USGS aerial imagery and elevation data to place the campus in the surrounding terrain.
These sources provide the traceability needed to link design choices back to their origins—for example, matching building placement to a site plan or facade styling to an architectural reference—though dimensional accuracy still requires checking against applicable drawings and specifications.
Agent-assisted OpenUSD campus model construction
In the demonstrated workflow, a frontier AI model runs inside an agent harness that manages context, instructions and tool calls. Reusable skills encode repeatable tasks and domain knowledge. Developers connect modeling and analysis operations through tool interfaces (including the Model Context Protocol where supported) and set boundaries for what the agent may propose or apply.
For the campus example, DSX Blueprint buildings and equipment provided starting geometry, while site plans directed placement and first-phase details. Photographs and Voyager references guided architectural additions. For a solar canopy, the frontier agent used USD Content Agents to generate geometry and materials and then refined the result against the Voyager imagery, producing an editable architectural element for review. The canopy remained a conceptual element requiring engineering checks for dimensions, structure and simulation suitability.
The team combined reused assets and new geometry in additive USD layers so changes stayed separate from source assets. In the Omniverse RTX viewport they inspected the canopy in campus context, reviewed appearance and placement, and iterated as design decisions evolved. USGS aerial data provided landscape and elevation context.
Extending the model for a fixed power budget study
The campus model can support a capacity-planning study: how to configure an AI factory hall within a fixed power budget. This requires connecting layout with compute, electrical and cooling requirements to evaluate alternative configurations.
An illustrative comparison from NVIDIA Vera Rubin NVL72 contrasts a MaxP baseline (highest static GPU power setting) with DSX MaxLPS (a DSX design and operating framework coordinating facility design and dynamic power allocation). Normalized to a 1 MW power basis with an assumed PUE of 1.1, the illustrative scenario shows a DSX MaxLPS capacity of 400 GPUs and aggregate token throughput about 1.35× the MaxP baseline. These are reference assumptions for the study concept, not results produced by the campus model.
In the workflow, the agent would identify equipment and assumptions needing review and use the reference comparison to inform capacity targets. Engineers would evaluate feasible configurations with complete racks—considering placement, electrical distribution, cooling connections and networking resources—using drawings, CAD dimensions and equipment catalogs for parameter inputs.
Requirements modeling and traceability
Teams can represent requirements in Systems Modeling Language (SysML) and link them to system elements. If the modeling tool exposes those relationships via an API, MCP or another interface, the agent can query dependencies and flag affected elements for review. The handoff from the agent is a proposed design revision accompanied by requirements, equipment data and follow-up engineering studies; after review, the agent can update the OpenUSD scene to reflect the chosen rack configuration.
Preparing the scene for a specific simulation requires additional checks and configuration (geometry, equipment properties, system connections and boundary conditions).
Assembling the digital twin with OpenUSD and SimReady
Simulation needs go beyond geometry: engineers require equipment properties and connection data to evaluate power and cooling. OpenUSD brings building geometry, equipment assets and associated data into a shared scene; USD attributes and custom schemas can preserve links to engineering records and requirements so changes remain traceable across applications.
The DSX system architecture links asset preparation, data storage, simulation services and the digital twin runtime. SimReady Foundation describes the asset metadata required for particular simulation uses: teams can assess whether assets contain geometry, properties and connection data needed for intended analyses and select validators appropriate to the equipment and analysis.
Developers can source equipment from the SimReady Central Library on Hugging Face, a free, ungated library of more than 2,000 simulation-ready OpenUSD assets, which can reduce asset-preparation effort when OEM-provided racks or power and cooling equipment are available.
AI factory pipeline samples provide examples for CAD conversion, optimization, metadata and validation; through connected tools, an agent can invoke these steps and collect reports. The output is an assembled scene tied to the proposed design revision, with completed checks and visible gaps for the simulation team to address.
Running power and cooling simulations to guide revisions
With an assembled scene and identified data gaps, the team prepares studies to evaluate electrical loading and cooling behavior under specified operating conditions. Where solver connections are supported, an agent can submit studies, retrieve results and associate findings with affected equipment and requirements. Engineers review assumptions and results to decide whether the design supports proposed configurations or needs modification. Findings should clearly identify the tested revision and distinguish calculated solver outputs from stored data or AI approximations.
Decisions about capacity then inform which equipment configurations to pursue, required supporting infrastructure and where further analysis is necessary; the editable design is updated and affected studies are repeated.
Recommended compute for development, visualization and simulation
Hardware selection depends on tasks:
- NVIDIA DGX Spark: personal agent development, local AI model prototyping; pair with an RTX system for DSX visualization.
- NVIDIA RTX PRO Blackwell Workstation Edition GPU: interactive design, OpenUSD editing and supported engineering simulation.
- NVIDIA DGX Station: larger local AI workloads and model development; add an RTX PRO 6000 Blackwell Workstation Edition GPU for visualization and supported simulation.
- NVIDIA RTX PRO Server: shared development and simulation following the NVIDIA RTX PRO AI Factory reference architecture (example sizing calls for eight RTX PRO 6000 Blackwell Server Edition GPUs per server and grouping four servers into a scalable unit).
Sizing should reflect scene complexity, solver requirements and concurrent users.
Integrations that connect design to engineering decisions
Several vendors are integrating their tools with Omniverse and the DSX workflow to maintain linkage between managed product data and digital twin studies:
- Dassault Systèmes: CATIA Magic and SysML tools help keep system requirements and design decisions consistent; they support Agent2Agent (A2A) and MCP-based access.
- PTC: Windchill integration with Omniverse libraries aims to bring version-controlled Creo and other CAD data into a shared 3D environment for engineering review and simulation.
- Cadence: Reality Digital Twin Platform couples equipment and power models with computational fluid dynamics to analyze MaxP, MaxQ and DSX MaxLPS scenarios and visualize impacts on power, cooling and token throughput.
- Siemens: Simcenter STAR-CCM+ enables detailed fluid flow and heat transfer evaluation; Siemens also provides DSX-aligned reference architectures combining power distribution, automation and simulation-ready assets.
These integrations aim to preserve the chain from managed equipment design through digital twin analysis and back into engineering decisions.
How to get started
Choose a single engineering question, connect required data and tools, assemble the digital twin in OpenUSD and use SimReady requirements to prepare assets for the intended analyses. Review the initial outcome before expanding the scope.
Practical next steps
- Build on the NVIDIA Omniverse DSX Blueprint.
- Assemble your digital twin with OpenUSD using reference assets and site data.
- Prepare and validate assets using DSX guidance and SimReady Foundation.
- Use agent-assisted pipelines to automate preparation steps and surface missing data.
- Engage engineers for verification and iterate until the design and simulations converge.
This agent-assisted OpenUSD approach supports traceable, repeatable development of AI factory digital twins from initial design through engineering analysis and back into revised designs.
