This summer, a 50-year-old research reactor at the University of Utah will do something it was never designed for: generate electricity for an AI workload. The TRIGA reactor, originally built by General Atomics for neutron experiments, will feed a miniature data center with 2 to 3 kilowatts of power—just enough to run a single high-performance GPU node.
Ted Goodell, the reactor’s manager, told Gizmodo this is the first known instance of a university reactor producing electricity for computational use. The setup repurposes the reactor’s thermal output: instead of dissipating heat into cooling water, Elemental Nuclear’s compact Brayton cycle generator captures waste heat and spins it into power.
The engineering trade-off here is clear. A full-scale AI data center demands hundreds of megawatts. This test delivers 0.002% of that. It’s a proof-of-concept for microreactors—factory-built, portable fission units designed for remote sites or direct attachment to computing clusters. No transmission losses, no grid dependency.
Why now? AI’s electricity consumption is straining U.S. grids. Goldman Sachs projects Big Tech will need 85 to 90 gigawatts of new nuclear capacity by 2030. Data center power draw could triple in that window. Hyperscalers are already placing bets: Microsoft revived Three Mile Island Unit 1 for 835 megawatts. Meta locked in 6.6 gigawatts across SMR projects with Oklo and TerraPower. Google partnered with Kairos Power for 500 megawatts of molten salt reactors. Amazon committed over 5 gigawatts from X-energy’s high-temperature gas design.
Startups are flooding in. Last Energy raised $100 million for mini-reactors targeting AI loads. Aalo Atomics broke ground on a data-center pod prototype. Oklo, backed by Sam Altman, went public and has a 14-gigawatt pipeline. China is moving faster—26 reactors including 12 SMRs for 960 megawatts, dedicated to AI training.
But challenges remain. Commercialization lags: fuel certification, regulatory approvals, grid interconnection. Westinghouse’s Vogtle plants ran seven years late and $18 billion over budget. NuScale canceled its first U.S. project on cost. SMRs promise two-year factory builds, but the supply chain for enriched uranium and specialized components is strained.
Nvidia is helping: PhysicsNeMo simulates reactor designs on GPUs. Oklo and Los Alamos are using AI digital twins to test plutonium fuels.
Back in Utah, the test is a spark. If a GPU hums on fission heat this summer, it validates a model: colocated reactors feeding racks directly, with zero transmission overhead and emissions-free baseload. AI’s appetite won’t wait for the grid to catch up.
Source: Webpronews