AI factory simulation: validate power and cooling before you build

Simulation reduces the boundary between design and operation by letting teams run operational scenarios before construction starts. A useful AI factory model does more than render a 3D room; it connects geometry, engineering attributes, live or historical operating data, so teams can test placement, power behaviour, airflow, cooling capacity, structural constraints, and workload impact before physical changes become expensive.

This matters because AI factories operate close to infrastructure limits, where small design assumptions can reduce usable token output, strand power or cooling capacity, or force conservative operating margins. The practical value comes from testing "operate-like" scenarios whilst still in design: rack moves, next-generation accelerator swaps, higher power density, different CDU selections, primary-loop changes, adiabatic cooling tower assumptions, and AI workload profiles.

Separate tools work for point analysis, but they break down when multiple personas need a shared view of the same facility, assets, and operating state. AI factory design requires CFD, power simulation, engineering documentation, asset models, IT telemetry, OT telemetry, and operational history to line up against the same physical equipment.

Without a common integration layer, each tool creates its own viewport, naming convention, and data model, which forces teams to reconcile assets manually. OpenUSD provides the standardisation layer for connecting 3D geometry to multiple data sources and solvers. Schneider Electric, AVEVA, and ETAP have joined the Alliance for OpenUSD to support software interoperability, industrial simulation, collaborative design, and AI infrastructure systems.

Token output depends on what happens inside the compute stack, so the model needs visibility below the rack level. CPU utilisation, memory utilisation, GPU utilisation, network behaviour, disk activity, and workload state help connect IT behaviour to OT consequences such as power draw, heat rejection, flow requirements, and thermal risk.

A rack-level abstraction misses whether a specific workload, node, or component behaviour is driving the infrastructure condition. This is why AI factory simulation must combine IT and OT data: the facility does not produce value by keeping rooms cold; the facility produces value by enabling expensive compute racks to run at the highest safe utilisation. Schneider Electric and NVIDIA reference design work explicitly ties AI factory design to operating points such as MaxP and MaxQ, with MaxQ enabling more tokens per watt when power constraints require optimised compute performance.

CFD simulation can test airflow, temperature, velocity, pressure, and cooling behaviour against both baseline and what-if scenarios. The practical use case is straightforward: start with a baseline room condition, apply a heavier AI workload or move thermal load from one rack location to another, and visualise whether cold airflow, cooling capacity, and containment strategy still support the load. For air-cooled or hybrid environments, a what-if move does not always require complete recabling or repiping; the model can shift the kilowatt dissipation profile and test the resulting thermal behaviour.

Power simulation can quantify how a failure or maintenance condition affects usable IT capacity. A practical scenario is a failed remote power panel combined with a second element entering maintenance mode. The simulation can show whether the facility still sustains the required redundancy posture and how much IT power must be capped to preserve N+1 behaviour.

That output can then inform workload scheduling rules, power caps, or frequency reductions before operations begin ETAP's electrical digital twin capability supports advanced electrical system design and simulation, dynamic what-if scenario analysis, real-time electrical infrastructure performance tracking, energy efficiency optimisation, predictive maintenance, and infrastructure planning based on power usage.

Simulation can evaluate the impact of replacing one GPU generation with another by testing power, cooling, and structural consequences before procurement or construction. A practical example is selecting a rack in the model and testing a future higher-density, heavier platform to understand whether the floor, pipe sizing, cooling loop, and electrical infrastructure can support the change.

This matters because AI factories refresh faster than traditional facilities, and next-generation rack platforms may alter weight, heat rejection, pipe sizing, flow requirements, and power density.

Simulation is becoming a practical requirement for AI factories because the cost of conservative assumptions or design error is too high. If a high-value rack can only be operated at 90% because the facility team lacks confidence in cooling, power, or redundancy margins, the operator gives up token capacity on the most valuable asset in the building.

AI factories require simulation because facility design, workload behaviour, and utilisation economics are tightly coupled. ETAP power simulation, CFD, real-time telemetry, and visual what-if scenarios help operators test the conditions that determine whether the AI factory can run close to the edge safely. OpenUSD-based digital twins further address the problem of siloed data and tools by enabling system-level reasoning across design, construction, and operation.