Nvidia taps Jacobs for a three-year data center digital twin at an R&D facility

Nvidia has tapped Jacobs to deploy a data center digital twin at one of its large-scale U.S. research and development facilities under a three-year software-as-a-service agreement, Construction Dive reported from a Jacobs release that frames the platform as a way to improve operational planning and facility performance.

The functions Jacobs lists are specific—dynamic power load balancing, energy forecasting, liquid coolant leak detection, predictive maintenance and operator training—and Amer Battikhi, an executive vice president at Jacobs, said in the release that simulated scenarios give operators greater visibility into infrastructure management. The list reads as a catalogue of what goes wrong in a dense hall: a load swinging faster than the utility can answer, a power bill that moves with the weather, a leak that does not announce itself, equipment degrading between service visits, and staff learning a system without taking it offline. Leak detection is the item that describes the building itself, since it implies liquid cooling in the halls being watched, though the coverage does not say what the facility's rack design is, where it sits, or what it cost to build.

Power and cooling lead the list, which tracks the hardware, and Siemens Grid Software's guidance that AI racks are reaching 230 kilowatts, and that grid planners should work from scenarios rather than a single load forecast, points to the same density that gives a data center operator a reason to model the hall, where a load swing plays out in seconds. Load balancing is the operator's half of the problem; grid access and load-class rules decide returns before a data center ever signs a lease, and none of that sits inside the twin's reach.

An R&D facility is a particular kind of first customer, since research workloads shift as designs shift, while a production hall runs a steadier profile and has less to rehearse. Nvidia is buying the software for a facility the coverage describes only as large-scale and in the U.S., and the agreement takes the shape of a service rather than a completed installation: three years of a subscription, sold on capability rather than a finished build. Operator training rounds out the list as the one function no sensor supplies.

Three years is the only commercial term the release provides, and it puts no price on the agreement; the blank price column has become infrastructure's default language, and a facilities software contract is no exception. Whether the model stays inside the fence as a tool for load balancing and leak detection, or gets shared with the utility on the other side of the meter, is the question worth carrying forward: there, the interconnection queue still sets the schedule.