AI Factory Digital Twins Link Facility Design to Tokens per Watt

Dana Tilley frames the problem as one of industrial speed. The technology behind the AI boom, he says, is “moving faster than anything we've ever seen in the industrialized world.” For AI infrastructure, that speed collides with a second constraint: the facilities themselves are becoming harder to design, build, and operate through conventional workflows.

Jim Gao describes AI factories as “so much larger, more complex, and more dynamic” than past data centers. Catherine Kniker makes the same point from the lifecycle side: because the world needs to move quickly through that complexity, digital twins have become more important.

The design and operating model for AI factories needs a shared representation that multiple disciplines can use. Tilley says that before digital twins, “design, engineering, construction, planning, and all were done in relative silos.” In his account, NVIDIA Omniverse and the DSX blueprint provide an AI architecture that lets multiple partners collaborate across the ecosystem instead of passing work through disconnected silos.

The on-screen examples reinforce that point without reducing the twin to a rendering. NVIDIA Omniverse DSX views show both an AI data center interior — server racks with overhead piping — and an exterior with large cooling units. Jacobs’ interface shows a 3D rack view with “NVIDIA GB200 NVL72,” an electrical simulation schematic, analytics for “Token Efficiency” and “Power Usage Effectiveness (PUE),” and three alternative data center configurations with location and power-source details.

The practical implication, as described by Tilley and illustrated by the Jacobs screens, is that the digital twin is meant to serve as a communication layer across design and planning concerns that are otherwise easy to separate. Tilley calls the DSX blueprint “a great starting point and frame of reference.” When the digital twin is layered on top of that process, he says, it becomes “your communication tool and continuity through the life cycle of the asset.”

That continuity matters because the facility is not treated as a static construction project. Tilley says the objective is to move “from concept to tokenization” faster. Jacobs’ screens include token efficiency, while Gao frames the operating target as tokens per watt.

Catherine Kniker argues that the collaboration problem cannot be solved by a visual model alone. Because so many participants are involved in designing, building, and operating an AI factory, she says, a PLM or product lifecycle backbone becomes “really important.”

In the DSX Omniverse blueprint, Kniker identifies PTC’s Windchill product as a core part of that backbone. The workflow she describes is direct: teams can publish from Windchill into Omniverse in USD format, then simulate the result and test different placements. If the simulation shows that changes are needed, the model can be adjusted before physical construction begins.

PTC’s on-screen examples show Windchill and CAD interfaces with a server rack in a data center room, a detailed server chassis, and a fully populated rack with liquid-cooling pipes. Their role is to show that the digital twin is not only a presentation of the building envelope. It includes equipment and cooling details that have to be coordinated with the rest of the facility design.

Kniker states the value proposition plainly: “The design is proven to work efficiently before ground is ever broken.” The emphasis is on moving design decisions into a digital environment where engineering participants can test placement and system interactions before construction commits the project to physical constraints.

Jim Gao pushes the argument beyond design collaboration into live operations. An AI factory, he says, “behaves as one single computer.” That creates “a higher degree of volatility” than operators have historically had to manage in data centers.

The operational example is cooling. A diagnostic interface appears with three server racks and colored heat mapping, indicating temperature conditions. Gao then describes Phaidra’s approach: train AI agents first inside a digital twin simulation of the AI factory. One of those agents is a liquid-cooling agent, which he says can eliminate thermal spikes by predicting them before they occur.

The logic is specific. If operators can eliminate thermal spikes, Gao says, they can safely increase their TCS temperatures. He does not unpack the acronym, but he ties the result directly to energy use: higher safe temperatures mean less energy required for cooling complex systems. That saved power can then be shifted into “revenue-generating IT load.”

Phaidra’s screenshots present that operating claim through anomaly and insight interfaces: a time-series graph for “cdu-01 Primary Supply Temperature Too High,” an “All CDU Insights” list, a “Home feed” with “Top insights,” a PUE anomaly view, a “DH1 CRAH 11 SAT Analysis” line graph, and workspace cards related to cooling distribution unit temperatures and setpoint deviations.

Taken together, Gao’s agent claim and the Phaidra screens draw a line between seeing facility behavior and optimizing against it. The interfaces surface anomalies, cooling conditions, PUE signals, and setpoint deviations. Gao’s claim goes further: agents trained in the digital twin can anticipate thermal behavior, choose better operating actions, reduce cooling energy, and keep optimizing for tokens per watt as conditions change.

That also explains why the facility is described as one computer. If the goal is maximum AI output per unit of power, racks, cooling loops, power systems, and control policies are not independent operating domains. They are coupled parts of the same production system.

Dana Tilley returns to communication because the AI factory lifecycle contains too many interdependent decisions for a linear handoff model. He says adopting the Omniverse DSX blueprint gives teams “a communication tool that cuts across the entire program and entire project.” Catherine Kniker extends that point across “all phases of the life cycle,” arguing that a digital representation enhances how people work together.

The collaboration claim is specific: design, engineering, construction, planning, lifecycle management, simulation, and operations all need to refer to the same evolving representation. Jacobs’ examples show configuration, power, PUE, and token-efficiency views. PTC’s examples show lifecycle-managed equipment models flowing into Omniverse for simulation. Phaidra’s examples show anomaly, insight, PUE, and cooling-analysis interfaces.

Across the lifecycle, the digital twin functions as both a planning environment and a simulation layer. Before construction, it supports layout, configuration, and simulation. During operation, it supports predictive agents that aim to reduce thermal volatility and improve energy allocation.