Schneider Electric Unveils Modular Design for NVIDIA DSX AI Factories

Why modular construction is gaining ground

Schneider Electric argues that GPU capacity is growing so quickly that traditional construction methods struggle to keep pace. The company says the fastest way to build AI factories is to use prefabricated data center modules.

The timing is tied to NVIDIA's roadmap. According to Schneider Electric, NVIDIA's next-generation Vera Rubin is already in production, with systems beginning to ship in late 2026. The more powerful Vera Rubin Ultra is expected to follow in 2027. Organizations investing in AI want to use this capacity quickly to speed up time to revenue, the company says.

The modular approach allows a data center to be built piece by piece, so customers can realize value from AI investments once the first components are in place.

NVIDIA DSX and the focus on tokens per watt

Schneider Electric describes AI factories as evolving into gigawatt-scale compute environments with thousands of racks and unprecedented power and cooling needs. NVIDIA DSX is designed to optimize these interconnected systems as a whole, from compute and networking to power and cooling.

The guiding principle is to keep resource consumption as low as possible while maximizing tokens per watt (TPW). Achieving that requires an understanding of factors such as the availability of behind-the-meter power, network capacity, and water and energy consumption. The more efficient the operation, the better the TPW, according to the company. Schneider Electric contrasts this with traditional compute deployments, where servers were typically overprovisioned and underutilized.

Reference Design 115 and RapidScale

Schneider Electric says it has made two DSX-related moves. It released Reference Design 115, which modularizes the DSX architecture, and it reintroduced its Power Modules, Power Skids, Data Halls, All-in-One Modules and IT Pods as RapidScale, its industrialized offering for modular AI infrastructure.

The reference design is a pre-engineered blueprint for a fully modular AI factory. It defines a scalable unit and a deployment unit:

• The scalable unit includes the white space and is built around RapidScale IT Pods. The company says these deploy 80% faster than traditional stick-built pods or hot-aisle containment systems.
• The pods are supported by skidded prefabricated cooling plants, power distribution units, remote power panels and fan walls that supply liquid and air cooling and power.
• Four scalable units make up one deployment unit, supported by outdoor power and cooling modules.

The deployment unit can be repeated as AI factory capacity needs grow, which supports parallel construction. Units are engineered, assembled and validated at Schneider Electric's factories and shipped to the site ready to be plugged in. Because all module components are tested at the factory, on-site commissioning is minimal, the company says.

Certifications and what comes next

Schneider Electric says it has delivered prefabricated data center modules for more than two decades and has now optimized them for AI. It reports that its RapidScale Power Skids earned an industry first UL 2140 certification, and that its RapidScale Power Modules meet UL2755.

The company describes the DSX reference design as the first phase of modularized AI factory implementations. As new GPUs are introduced, the design will evolve to accommodate them, with the aim of helping customers shorten time to token and maximize tokens per watt.

For organizations planning AI capacity, the offer is a pre-engineered path that can be expanded in phases as computing needs grow. The performance figures, including the 60% and 80% speed claims, come from Schneider Electric itself.