“The energy produced through nuclear fission can ultimately power the computational systems driving artificial intelligence,” says Mike Luther, founder of Elemental Nuclear. This is precisely what is taking place at the University of Utah a prototype reactor, a compact turbo generator cycle, and a GPU executing a live task.
The demonstration is less about the modest estimated output of two to three kilowatts than it is about the posed engineering challenge. With data centers consuming ever increasing amounts of power in increasingly challenging places, the current approach to power infrastructure design appears to be out of date. In 2021, data centers consumed roughly 4.5 percent of all electricity in the country; some projections suggest a much sharper increase by 2030. To meet the challenge in Utah, it takes neither a giant new plant nor an incremental approach, just a carefully controlled experiment demonstrating the viability of using reactor’s heat to produce useful computing power.
The school’s research reactor, known as TRIGA, was traditionally used as a lab facility, not a power source. As such, it has not been delivering electricity. Instead, its heat used to be dumped to the environment via special cooling systems. For the upcoming summer demonstration, engineers will connect the reactor to a compact reverse Brayton cycle using helium rather than steam as the working fluid. The cycle involves a series of pumps compressing and cooling the gas, heating it with reactor water and sending it to a turbine where it expands and creates electricity. The output then goes through another cooling stage in a specially designed heat exchanger, resulting in an overall process which consumes close to 50 kilowatts and produces roughly 13 kilowatts of electricity, with a fraction making its way to GPU and performing an actual task.
The project is less about the numbers than it is about the principle behind it. According to reactor director, Dr. Ted Goodell, the project represents “the first time any university reactor has produced electricity,” adding that it shows “small, safe reactors could live at data centers, rather than in labs.” The statement reveals the changing face of engineering practice, which sees the nuclear power technology shifting from a purely centralized one toward more decentralized applications closer to data infrastructure.
The logic behind the idea is clear. With AI facilities in dire need of reliable and compact power generation systems, engineers have started designing SMRs and microreactors as a potential solution at least in terms of licensing and planning considerations. While there is yet no American data center relying on an SMR or a microreactor, the engineering demands remain considerable, including compliance with higher seismic requirements, provision of separate utility corridors, capability of heat dissipation in the long run, dual interconnections, and far more rigorous geotechnical analysis than in traditional facilities. Still, engineers consider nuclear power plants due to their capacity for compactness and availability 24/7.
In Utah, a campus in Millard County has been evaluated as a prospective AI hub, with the plan considering the possibility of installing 2+GW of nuclear generation capacity along with other energy types. In comparison with the gigantic scale, a reactor powering a GPU may seem trivial indeed. Yet, small-scale experiments are essential in revealing interfaces between power generation and computing hardware that may become critical later. At the moment, Utah’s test project involves participants from twelve different universities. Thus, the modest production of electricity becomes a shared infrastructure experiment in the area of integrating nuclear and computation technologies.