Reference Designs and the Digital Twin

New building types get standardized in a predictable way. When a design is new and risky, the dominant supplier publishes a reference design: a tested configuration that others can copy. The reference design lowers risk for the buyer and fixes the supplier's product at the center of the build. For AI factories, NVIDIA has been doing exactly this, and the control plane has become part of the reference.

Trace the sequence. In October 2024, Vertiv and NVIDIA published the seven-megawatt reference architecture for the GB200 NVL72 rack-scale system, covering power and cooling. In September 2025, Schneider Electric announced two reference designs with NVIDIA. One connected facility power management and liquid-cooling controls to NVIDIA Mission Control, the operations software NVIDIA had announced that March. The other covered GB300 NVL72 data halls of up to 142 kilowatts per rack, and Schneider said it included its ETAP electrical modeling and EcoStruxure IT Design computational fluid dynamics models, so that power and cooling scenarios could be simulated before construction.

The next step moved from hardware to the whole facility. At its GTC conference in Washington in October 2025, NVIDIA introduced the Omniverse DSX blueprint, a framework for building digital twins of gigawatt-scale AI factories. A digital twin is a live simulation of the facility that mirrors the real building's electrical, thermal and computing state. NVIDIA said the blueprint would be validated at Digital Realty's AI Factory Research Center in Manassas, Virginia.

Then, on March 16, 2026, NVIDIA released the Vera Rubin DSX AI Factory reference design alongside general availability of the Omniverse DSX blueprint. The reference design names three software components, and each maps onto a problem this course has already described. DSX Max-Q aims to maximize computing output per watt within a fixed power budget. That is the utilization problem, stated in electrical terms. DSX Flex connects the AI factory to power-grid services so it can adjust its consumption and coordinate with onsite generation. That is the grid problem, which the final module covers. DSX Exchange is described as integrating compute, network, energy, power and cooling-plant signals between IT, operational technology and operations agents. That is the seam between layers, given a product name.

The partner list is the other signal. NVIDIA named Cadence, Dassault Systèmes, Eaton, Jacobs, Nscale, Phaidra, Procore, PTC, Schneider Electric, Siemens, Switch, Trane Technologies and Vertiv as contributors. That list includes the major electrical and cooling vendors, a construction engineer, simulation software companies and at least one company, Phaidra, that sells AI-based control for industrial plants. At the same March 2026 conference, Schneider Electric and AVEVA, the industrial software company Schneider owns, announced that AVEVA's engineering and operations software was embedded in the Omniverse DSX blueprint and ecosystem.

Read the structure, not the press releases. Three things are happening at once. The facility vendors are exposing their equipment as data, so that chillers, CDUs, which is to say coolant distribution units, and switchgear become software objects. The chip vendor is defining the interface through which that data reaches the computing stack. And simulation is being extended from design into operations, so that the model used to plan the building is the same model used to run it.

The benefit is real. A shared data model is the missing piece identified in the previous lesson. If a scheduler can query the thermal and electrical state of each row through a common interface, it can place work more intelligently. If the facility can see the job queue, it can prepare cooling before a large job starts rather than after.

The dependency is also real. When the reference design, the scheduler, the telemetry and the digital twin framework all come from one supplier, the operator's control plane is shaped to that supplier's hardware roadmap. Facility vendors joined because being absent from the reference is costly. An operator adopting the reference gets integration and gives up some freedom to mix components.

There is also a gap between a reference design and a running building. A reference design is a tested pattern, not an operating record. The announcements describe what the components are intended to do. They do not yet tell us how much throughput a DSX-style control plane recovers in practice, or how it behaves in a fault.

Faraday's read is that the reference design has become the unit of competition in AI infrastructure, and the control plane is now inside it. The durable question for any operator is which layers it controls itself and which it has accepted from a supplier. The durable question for an investor is whether the operator knows the difference.