For 50 years, the nuclear reactor buried beneath the University of Utah campus has done exactly one thing: produce neutrons. Students use it to study reactor physics. Researchers use it for neutron activation analysis. It has never generated a single watt of electricity. That is about to change.
In a partnership with startup Elemental Nuclear, the university plans to route heat from its TRIGA Mark I reactor through a small turbine, convert it into roughly 13 kilowatts of electrical power, and feed that power directly into a GPU node running a live artificial intelligence workload. If the demonstration succeeds, it will mark the first time a U.S. university research reactor has produced electricity specifically to drive AI computing hardware.
The test is modest in raw output. Thirteen kilowatts is roughly what a single American household draws at peak demand. But the point is not scale. The point is proof: that a licensed, campus-based nuclear facility can be coupled to a power-conversion system and deliver stable current to the kind of sensitive, high-performance hardware that AI depends on.
A reactor built for teaching, not for power
The University of Utah’s TRIGA reactor operates under NRC License R-126, originally issued on September 30, 1975, under Docket 05000407. It is rated at a maximum steady-state thermal output of 100 kilowatts. The Nuclear Regulatory Commission classifies it as a nonpower reactor, a designation that reflects its purpose: education, training, and experiments, not electricity generation.
TRIGA reactors (Training, Research, Isotopes, General Atomics) are among the most widely deployed research reactors in the world, with dozens installed at universities and national laboratories across more than 20 countries. Their signature safety feature is a uranium-zirconium hydride fuel that has a strong negative temperature coefficient, meaning the nuclear chain reaction naturally slows as the fuel heats up. That built-in self-regulation is one reason TRIGAs have operated safely on campuses for decades.
An NRC inspection report from 2022 confirmed that the Utah facility was operating within its technical specifications and identified no significant safety issues. Nothing in that report, or in any other publicly available NRC filing, suggests the reactor has ever been configured to send thermal energy into an electrical power-conversion system. University materials consistently describe the facility as a research tool, not a generator.
That history is what makes the planned demonstration notable. After five decades of operation in a single mode, the reactor is being asked to do something fundamentally different.
How the test is supposed to work
According to an announcement from the university’s John and Marcia Price College of Engineering, the proof-of-concept system will draw approximately 50 kilowatts of thermal energy from the TRIGA core and pass it through a small turbine to produce about 13 kilowatts of electrical output. That implies an overall thermal-to-electric efficiency of roughly 25 percent, a figure consistent with small, simple-cycle conversion systems such as organic Rankine cycle units.
The electricity will then power a compact computing setup: a GPU node performing a continuously active AI workload in real time. The university has not specified whether the task will involve inference (answering queries from a trained model), fine-tuning, or full model training. A single GPU operating within a 13-kilowatt envelope could handle a range of tasks, from serving a modest vision model to running a smaller language model under continuous load.
For context, a single modern AI server rack in a commercial data center can draw 40 kilowatts or more. A hyperscale AI facility operated by companies like Microsoft or Google may demand 100 megawatts to over a gigawatt, roughly 10,000 to 100,000 times the output of this test. The Utah experiment is not trying to compete at that scale. It is trying to close the loop, from fission to electrons to a functioning AI application, inside a tightly controlled research environment.
Open questions on regulation and engineering
Several important details remain unclear in public records as of May 2026. The most significant is regulatory. The NRC’s safety evaluation report for the most recent license renewal focuses on the reactor’s established mission and does not address coupling the cooling systems to a turbine-generator. Whether the university sought or required a license amendment to conduct the experiment has not been confirmed in publicly available filings.
The answer matters. Depending on how the NRC and the university’s internal reactor safety committee classify the power-conversion hardware, the project could require anything from routine internal approval to a formal license amendment with public notice and comment. That regulatory pathway, whatever it turns out to be, could set a precedent for other universities considering similar experiments.
On the engineering side, Elemental Nuclear has not publicly released schematics, thermodynamic cycle details, or component specifications for the turbine system. Research reactors of this class typically reject heat through a secondary cooling loop to the environment via heat exchangers. Introducing a turbine into that chain raises practical questions about pressure boundaries, temperature control, isolation valves, and how the system would respond to transients, such as a sudden turbine trip or a fluctuation in computing load, without disrupting reactor operations.
Funding sources are also undisclosed. It is not clear whether the project is backed by federal research grants, state funding, private investment in Elemental Nuclear, or internal university budgets. The distinction matters: a federally funded program would likely emphasize reproducible data and regulatory learning, while a privately backed demonstration might prioritize investor-facing milestones.
Where this fits in the nuclear-AI landscape
The Utah test arrives during a broader push to pair nuclear energy with the explosive power demands of artificial intelligence. In 2023 and 2024, several major deals signaled that the technology industry sees nuclear as a serious option for baseload AI power. Microsoft signed an agreement to help restart Three Mile Island Unit 1 through Constellation Energy. Amazon struck a power purchase deal tied to Talen Energy’s Susquehanna nuclear plant in Pennsylvania. Google announced a partnership with Kairos Power to develop advanced reactors for its data center operations.
Those deals involve large commercial reactors or next-generation designs operating at hundreds of megawatts. The Utah project sits at the opposite end of the spectrum: a microreactor-scale demonstration on a university campus, producing just enough electricity to run a single computing node. But it addresses a question the bigger deals do not. Can a small, modular nuclear heat source be directly integrated with AI computing hardware in a way that is safe, licensable, and operationally reliable?
The university has also discussed a potential upgrade from the reactor’s current 100-kilowatt thermal rating to 1 megawatt. If pursued and approved by the NRC, that increase would expand the envelope for future experiments, potentially enabling tens or even hundreds of kilowatts of electrical generation for more demanding computing loads. As of now, however, no publicly available NRC application or detailed schedule confirms when, or whether, that upgrade will move forward.
What the test will actually prove
The significance of this demonstration will not hinge on raw wattage. Thirteen kilowatts will not power a data center, and no one involved is claiming it will. What the test can prove, if it is conducted transparently and the results are published, is something more foundational: that the engineering, licensing, and operational challenges of coupling a nuclear heat source to AI computing hardware can be identified, managed, and documented in a real-world setting.
That kind of data has value. The microreactor industry is still in its early stages, with companies like Kairos Power, Oklo, and X-energy working toward first deployments. Detailed, publicly available results from a campus-scale integration test could inform how future nuclear-powered microgrids are designed for digital workloads, what regulators will require, and where the engineering bottlenecks actually lie.
If key details remain opaque, the impact may be more limited: a symbolic milestone, a small GPU humming along on power from a reactor that, for half a century, was never meant to generate electricity at all. The difference between a landmark experiment and a publicity exercise will come down to what the university and Elemental Nuclear choose to publish after the turbine spins up.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
