At a glance
- The future of the AI factory is modular
- As GPU capacity grows at dizzying speed, traditional construction methods struggle to keep up
- The fastest way to build AI factories is to leverage prefabricated data center modules
- NVIDIA plans to introduce higher-capacity GPUs soon
- NVIDIA’s next-generation..
Modular AI Factory Infrastructure Offers a Faster Path to First Token
As GPU capacity scales at an accelerating rate, traditional data center construction methods are struggling to keep pace with AI demand. Prefabricated modular infrastructure is emerging as the fastest route to deploying AI factories, with implementation timelines reduced by up to 60 percent compared with conventional stick-built approaches.
The urgency is driven by NVIDIA’s product roadmap. The company’s next-generation Vera Rubin systems are already in production and expected to begin shipping in late 2026, with the more powerful Vera Rubin Ultra anticipated in 2027. Organizations investing billions in AI compute want to leverage this capacity as quickly as possible to accelerate time to revenue. For these operators, a modular AI factory enables value realization as soon as the first components are in place, rather than waiting for an entire facility to be completed.
DSX Architecture Optimizes Tokens per Watt
AI factories are evolving into gigawatt-scale compute environments comprising thousands of racks, creating unprecedented power and cooling requirements. Supporting that density requires extensive networking, electrical, and thermal infrastructure. NVIDIA’s DSX architecture is designed to optimize these interconnected systems as a whole, spanning compute, networking, power, and cooling.
The guiding principle behind DSX is to minimize resource consumption while maximizing tokens per watt. Achieving this requires a deep understanding of behind-the-meter power availability, network capacity, and water and energy consumption. This represents a fundamental shift from traditional compute implementations, where servers were typically overprovisioned and underutilized. With AI investments reaching billions of dollars, designing architecture to maximize throughput has become a critical priority.
Reference Design 115 Modularizes DSX
Schneider Electric has partnered with NVIDIA to support customers building AI factories, releasing Reference Design 115 to modularize the DSX architecture and accelerate deployment. The reference design provides a pre-engineered blueprint for a fully modular AI factory built on DSX principles of output maximization balanced with optimized resource consumption.
The design is organized around scalable units and deployment units. Each scalable unit includes white space built around Schneider’s RapidScale IT Pods, which are optimized for high-density AI workloads and deploy 80 percent faster than traditional stick-built pods or hot-aisle containment systems. These compute and support pods are backed 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 comprise a deployment unit, supported by outdoor power and cooling modules. The deployment unit is designed to scale repeatedly based on AI factory capacity needs, supporting a parallel construction approach in which units are added as compute requirements increase.
RapidScale Industrializes Modular Deployment
RapidScale units are pre-engineered, assembled, and validated at the factory before being shipped to site with all power and cooling infrastructure needed to support white space. Because module components are factory-tested, on-site commissioning is minimal. This plug-in delivery model accelerates time to first token by reducing the complexity of field integration.
Schneider Electric has reintroduced its Power Modules, Power Skids, Data Halls, All-in-One Modules, and IT Pods under the RapidScale brand—an industrialized solution for designing, delivering, and supporting modular AI infrastructure at scale. The company has delivered prefabricated data center modules for more than two decades and has now optimized them specifically for AI deployments.
Certifications Simplify Commissioning
Third-party certifications further streamline deployment. Schneider Electric’s RapidScale Power Skids have earned an industry-first UL 2140 certification, while RapidScale Power Modules meet UL 2755 standards. These certifications adhere to strict safety, performance, and integration requirements, simplifying commissioning and shortening time to first token.
As new GPUs are introduced, the DSX reference design will evolve to accommodate changes, continuing to help customers accelerate time to token and maximize tokens per watt. For operators facing gigawatt-scale AI ambitions, modular infrastructure offers a deployment path that matches the speed of silicon innovation.
MegaWatt Intelligence Desk
Hyperscale & Grid Analyst. `More stories by $MegaWatt Intelligence Desk →`
