Bechtel plans to take NVIDIA’s Omniverse DSX gigawatt-scale “AI factory” reference design and convert it into modular, repeatable building blocks. The plan is aimed squarely at compressing schedules and de-risking delivery for the current wave of hyperscale AI build-outs. The company will showcase its approach at NVIDIA GTC 2025 in Washington, D.C., positioning modularization as the lever that gets customers from groundbreaking to the “first revenue token” milestone—the moment a site processes its first production workload—more quickly and predictably.
The pitch blends two Bechtel strengths: industrial modularization and tightly integrated program execution. Rather than the traditional, multi-contractor hand-off model common in data centers, Bechtel is standing up a single, end-to-end engineering, procurement, construction, and commissioning (EPCC) model. In practice, that means standardizing design packages and fabricating major subsystems off-site—power, cooling, mechanical and electrical rooms, network blocks—so they arrive as pre-validated units ready for rapid setting and hook-up.
Underpinning the effort is NVIDIA’s Omniverse blueprint, which Bechtel is using as the canonical design pattern to align physical scope, interfaces, and performance targets across multi-generation deployments. By turning that pattern into a kit of parts, Bechtel says it can replicate proven layouts across regions while adapting to local utility, land, and permitting realities. The company also notes that the “AI factory” concept stretches beyond white space: scope includes on-site power integration, water treatment, and the support infrastructure needed to operate at gigawatt scale.
Catherine Hunt Ryan, president of Bechtel Manufacturing & Technology, frames the partnership as a reliability play as much as a speed play: combining NVIDIA’s hardware optimization with Bechtel’s megaproject controls to yield facilities that “are faster to build, more reliable to operate, and ready to scale globally.” In the capital-intensive AI cycle, the sooner a site generates first compute, the sooner operators can prove out utilization, tune power and cooling, and unlock revenue.
For owners racing from pilot clusters to multi-billion-dollar campuses, a modularized, reference-aligned approach offers fewer unknowns at each interface and an easier path to replicate. If Bechtel’s blueprint-to-modules workflow lands as advertised, it could become a default template for the next wave of AI infrastructure programs.
