Developing an AI Factory Digital Twin with an Agent-Assisted Workflow
Building a foundation for an AI factory digital twin An brings site plans, equipment assets, and engineering requirements into a consistent model that teams can inspect, revise,...
By Software Development Team
Building a foundation for an AI factory digital twin
An AI factory digital twin brings site plans, equipment assets, and engineering requirements into a consistent model that teams can inspect, revise, and prepare for simulation. Connecting these inputs through an editable workflow allows teams to explore design alternatives and incorporate engineering feedback as the design develops.
The NVIDIA Omniverse DSX Blueprint, part of NVIDIA DSX Sim, provides reference digital twin assets, sample scripts, and workflows for developing AI factory digital twins. Built on OpenUSD and NVIDIA Omniverse libraries, the blueprint provides a foundation for assembling facility and equipment models with SimReady assets and connecting them to engineering simulation tools.
Frontier AI models can work with this foundation through agents and connected applications. Developers can guide agents to combine design inputs, assemble and revise OpenUSD scenes, and inspect the results. The blueprint supplies reusable assets and workflows, while agents help apply and adapt them to specific design requirements.
This workflow turns reference assets, site data, and design requirements into an editable OpenUSD campus model. It can support layout exploration, equipment placement revisions, and preparation for engineering review and downstream simulation. Developers connect tools and guide changes, while engineers review the results and use their findings to inform the next iteration.
Providing agents with a design foundation
A practical question for an AI factory is how to configure a hall when available power is fixed. Changing the compute configuration can affect equipment quantity, placement, electrical distribution, and cooling. Evaluating these alternatives requires the relevant disciplines to work from consistent requirements and design data.
An agent needs context before it can produce a useful design proposal. Photos or renderings alone may result in a model that looks plausible but has inaccurate dimensions or proportions. Reference assets and project documents provide additional grounding, but the resulting model still requires dimensional checks and engineering review.
A grounded workflow begins with the NVIDIA Omniverse DSX Blueprint and its reference design, then incorporates project requirements, drawings, equipment information, and site constraints. Each source serves a different purpose:
- Existing geometry supplies reusable buildings and equipment.
- Architectural references inform appearance.
- Aerial imagery and elevation data provide geographic context.
- Project drawings and equipment information establish dimensions and placement.
- Requirements define the conditions that the design must satisfy.
For the campus example, the NVIDIA Santa Clara Headquarters building, Voyager, served as architectural inspiration for an AI factory site based on the DSX reference design. The inputs included:
- DSX 3D models: Data halls, central utility buildings, substations, generators, cooling equipment, vehicles, and vegetation.
- Architectural references: Aerial and ground-level images of Voyager showing the roof, facade, solar canopies, and landscaping.
- Site plans: PDFs describing the campus arrangement and a more detailed first phase, including buildings, roads, utility areas, and access areas.
- Requirements documents: An indexed collection of architectural drawings, form-factor studies, Basis of Design documents, room layouts, and other records.
- Geographic data: USGS aerial imagery and elevation data for the selected site, used to place the campus in its surrounding terrain.
Together, these sources provide a starting point for an agent-assisted workflow and a basis for reviewing its output. They also help teams trace design choices to their sources, such as linking building placement to a site plan and facade styling to an architectural reference. Dimensional accuracy still depends on checking the model against applicable drawings, specifications, and site constraints.
Building an OpenUSD campus model with an AI agent
To implement this workflow, a frontier AI model runs in an agent harness, software that manages context, instructions, and tool calls. Skills encode repeatable tasks and AI factory domain knowledge. Developers connect modeling and analysis operations through tool interfaces, including the Model Context Protocol (MCP) where supported, and define which changes the agent may propose or apply.
In the campus example, DSX reference buildings and equipment from the DSX Blueprint provided the starting geometry. Site plans guided placement and first-phase details, while photographs and architectural references from the NVIDIA Voyager campus guided the architectural additions.
For the solar canopy, the frontier AI agent used USD Content Agents to help generate geometry and materials. The design was then refined using Voyager references, producing an editable architectural element that could be reviewed alongside the existing buildings and equipment. The canopy remained conceptual and required engineering checks for dimensions, structural properties, and suitability for simulation.
The reused assets and new architectural geometry were combined in additive USD layers, keeping changes separate from the source assets. In the Omniverse RTX viewport, the team could inspect the canopy in the context of the campus, review its appearance and placement, and revise the model as design decisions changed. USGS aerial data supported the landscape and elevation plan.
Extending the model for a fixed power budget
The campus model could support a capacity-planning study focused on configuring an AI factory hall within a fixed power budget. The proposed extension would connect the modeled layout with compute, electrical, and cooling requirements to evaluate alternative configurations.
One possible starting point is an illustrative NVIDIA Vera Rubin NVL72 inference comparison. It contrasts a baseline using MaxP, the highest static GPU-level power setting, with DSX MaxLPS, an AI factory design and operating framework that coordinates facility design, dynamic power allocation, and performance-per-watt optimization. Applying this comparison to a hall would require evaluating support power, cooling, and networking infrastructure.
Normalized to a 1 MW power basis and assuming a power usage effectiveness (PUE) of 1.1, the illustrative scenario shows DSX MaxLPS capacity of 400 GPUs and aggregate token throughput of 1.35x the MaxP baseline. These figures are reference assumptions for the proposed study, not results generated by the campus model.
In the proposed workflow, developers would connect the agent to project data and engineering tools so it could help identify equipment and design assumptions that require review. The reference comparison would inform capacity targets. Engineers would evaluate feasible configurations using complete racks, including their placement, electrical distribution, cooling connections, and network resources.
Using Systems Modeling Language (SysML), teams can represent requirements and link them to system elements. If the modeling tool exposes those relationships through an API, MCP, or another supported interface, an agent could query them and flag dependencies potentially affected by a proposed change. Drawings and CAD would provide dimensions and placement references, while equipment catalogs would supply parameters for engineering analysis.
The handoff would consist of a proposed design revision accompanied by the relevant requirements, equipment information, and follow-up engineering studies. After engineering review, the agent could help update the OpenUSD scene to reflect the selected rack configuration and layout. Preparing the scene for a specific simulation would require additional checks and configuration involving geometry, equipment properties, system connections, and boundary conditions.
Assembling the digital twin with OpenUSD and SimReady
After a proposed configuration and its affected requirements have been reviewed, the design can be prepared for analysis. This requires more than geometry. Engineers also need 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, helping keep changes traceable across applications.
The DSX system architecture connects asset preparation, data storage, simulation services, and the digital twin application. For a fixed-power scenario, this provides a framework for combining the revised layout with the equipment information needed to evaluate power and cooling implications.
SimReady Foundation describes the asset data required for particular simulation uses. OpenUSD provides the shared scene, while SimReady specifications help teams determine whether assets contain the data required for the intended analysis. Teams can select appropriate requirements and validators, then check the geometry, properties, and connection data they cover. An agent can help identify missing information and associate it with the affected equipment for review.
Equipment assets are also available through the SimReady Central Library on Hugging Face, a free, ungated library containing more than 2,000 simulation-ready OpenUSD assets. For AI factory digital twins, teams can discover and reuse OEM-provided racks, power equipment, and cooling equipment, reducing asset preparation work.
The AI factory pipeline samples provide examples for CAD conversion, optimization, metadata, and validation. Through connected tools, an agent can help invoke these preparation steps and collect their reports.
The resulting handoff is an assembled scene tied to the proposed design revision, with supporting data, completed checks, and remaining gaps visible. This gives the simulation team the context needed to prepare a study and evaluate how the configuration behaves within the fixed power budget.
Using simulations to guide design revisions
Once the revised scene is assembled and data gaps have been identified, the required studies can be prepared. In the fixed-power scenario, electrical loading and cooling analysis help determine how the proposed equipment arrangement behaves under specified operating conditions. The application uses geometry, equipment properties, and system connections, with additional solver preparation where necessary.
Where supported solver connections have been implemented, an agent can submit studies, retrieve results, and associate findings with affected equipment and requirements. Engineers review the assumptions and results to determine whether the design supports the proposed configuration or requires further revision. Each finding should identify the tested design revision and distinguish calculated results from stored results or AI approximations.
The capacity question then becomes a basis for design decisions: which equipment configuration to pursue, what supporting infrastructure it requires, and where additional analysis is needed. These decisions feed back into the editable design, allowing the team to refine the proposal and repeat affected studies.
Selecting compute for development, visualization, and simulation
Hardware for a workflow like the campus example should be selected according to the task, such as developing agent tool connections, editing the scene, or running engineering studies.
- NVIDIA DGX Spark for personal agent development: Supports agent prototyping, compatible local AI models, and engineering-tool connections. An RTX system can be paired with it for DSX visualization and simulation.
- NVIDIA RTX PRO Blackwell Workstation Edition GPU for interactive design and simulation: Supports OpenUSD asset preparation, digital twin review, and supported engineering simulations.
- NVIDIA DGX Station for larger local AI workloads: Supports development and local execution of larger AI models and agents. An RTX PRO 6000 Blackwell Workstation Edition GPU can be added for visualization and supported simulation applications.
- NVIDIA RTX PRO Server for shared development and simulation: The NVIDIA 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. Deployment size depends on inference, visualization, and supported simulation workloads, as well as scene complexity, solver requirements, and concurrent users.
Connecting applications to engineering decisions
Extending the campus model into an engineering study requires applications that manage requirements, provide equipment data, and calculate physical behavior. Several application providers address different parts of this workflow:
- Dassault Systèmes: CATIA Magic supports system requirements and design decisions through SysML tools. Engineers can model system elements and interfaces, check model consistency, and evaluate designs against the architecture. In an agent-assisted workflow, simulation findings could inform proposed updates to the SysML model, subject to engineering review. Dassault Systèmes also supports the Agent2Agent (A2A) protocol for agent-to-agent integration, complementing MCP-based access to tools and data.
- PTC: Windchill and NVIDIA Omniverse libraries connect digital twin studies with the product data they represent. Windchill manages version-controlled Creo and other CAD data. The announced integration is intended to bring that data into a shared 3D environment for engineering review and simulation, addressing the transition from managed equipment design to an AI factory digital twin.
- Cadence: The Reality Digital Twin Platform addresses questions about GPU capacity and cooling. It couples equipment and power models with computational fluid dynamics to simulate MaxP, MaxQ, and DSX MaxLPS. Built on NVIDIA Omniverse libraries, SimReady assets, and OpenUSD, the platform supports collaborative and agent-assisted optimization of AI factory performance.
- Siemens: Simcenter STAR-CCM+ supports evaluation of fluid flow and heat transfer under specified conditions. Siemens has also developed DSX-aligned reference architectures that combine power distribution, automation, and simulation-ready assets, representing supporting infrastructure alongside compute equipment.
Starting with the NVIDIA Omniverse DSX Blueprint
A practical starting point is to choose one engineering question, connect the required data and tools, assemble the digital twin with OpenUSD, and use SimReady requirements to prepare assets for the intended analysis. The result should be reviewed before the workflow is expanded.
Relevant resources include: