Valar Atomics says it has used power from its Ward 250 reactor in Utah to run an Nvidia Blackwell machine. The demo was small, but the question behind it is not: whether AI data centers can get power directly from reactors instead of waiting on a strained grid.
The demonstration was modest by design. Valar connected an Nvidia RTX Spark desktop machine built on Blackwell architecture to electricity drawn from Ward 250, its helium-cooled reactor in Orangeville, Utah, and used it to briefly host a live website. No GPU hall. No 30-megawatt campus. Just a reactor, a desktop machine and a public claim that the circuit worked.
You should treat the scale honestly. A website loading from a desktop PC is not a data center. But it is still a useful signal, because Nvidia's name now sits beside a nuclear startup that has moved from paper promises to a working test reactor at a time when AI power demand is running ahead of the grid.
Valar's Ward 250 reached criticality in June, according to recent reporting from Barron's on the federal reactor push, alongside advanced reactor tests from Antares Nuclear and Deployable Energy. Criticality does not mean commercial power. It means the reactor has achieved a self-sustaining nuclear chain reaction, closer to turning an engine over than driving the car down the highway. That distinction matters, and too much nuclear hype tends to skip it.
The Ward 250 design uses helium cooling, graphite moderation and TRISO fuel, the ceramic-coated uranium fuel type often discussed for high-temperature reactors. Earlier reporting from The Wall Street Journal said the Ward 250 system was flown by the US military from California to Utah in February, unfueled, using three C-17 aircraft. The same report said testing would begin at 250 kilowatts and that Valar ultimately wanted a 5-megawatt system.
That is the useful context for the Nvidia demo. Valar says the reactor was producing around 100 kilowatts of thermal energy for the July 1 test, converted to electricity through a thermoelectric generator. That is tiny beside the power draw of AI infrastructure. A single serious AI data center can use tens or hundreds of megawatts. Google's electricity use jumped 37% in 2025, according to the company's latest sustainability report, and the company said AI demand is making its clean-power goals harder to meet.
Frankly, that is why this small test is getting attention. Nvidia does not need a 100-kilowatt reactor to run a desktop machine. It needs evidence that reactors can be built close to compute, fast enough to matter and without waiting years for new transmission lines. If you run chips for a living, electrons delivered eventually are not the same as power available on site.
Valar and Nvidia also announced a feasibility study for a 30-megawatt data center in Utah that would run on Ward 250 power and use little water. That part is still a study, and you should read it that way. Valar has not said when construction would begin, who would finance the facility, or what permitting path a commercial version would need. The jump from a 100-kilowatt thermal test to a 30-megawatt data center is enormous.
The water point is real enough to watch. Ward 250's helium cooling loop does not rely on the same water-heavy cooling model used across much of the data center industry, and Utah's high desert is not a place where water can be treated as an afterthought. If Valar can pair reactor heat management with Nvidia's data center cooling work, the result could matter beyond one site in Emery County. If it cannot, the demo stays a good photo with a narrow technical footnote.
Big technology companies are already moving toward nuclear power, but mostly through contracts on the existing grid. Microsoft has a 20-year agreement with Constellation tied to restarting Three Mile Island Unit 1 in Pennsylvania. Amazon has backed nuclear-linked power at its Susquehanna data center campus. Google signed a deal with Kairos Power for up to 500 megawatts of small modular reactor capacity starting in the 2030s. Those are large commitments, but they are still grid arrangements.
Valar is trying to show something different: put the reactor near the chips and avoid the grid bottleneck altogether.
The pressure is not theoretical. PJM Interconnection, the grid operator covering much of the Mid-Atlantic and parts of the Midwest, has warned that data centers could add 30 gigawatts of demand by 2030. Reuters and other outlets have reported growing concern from grid operators and regulators over how fast large loads are showing up. When electricity demand arrives faster than transmission, generation and permitting can move, even wealthy tech companies run into a hard limit.
That is the real story in Utah. The first Nvidia chip powered by Ward 250 is not proof that nuclear-powered AI campuses are around the corner. It is proof that the argument has moved from slide decks into hardware, however small the first step looks. Now Valar has to do the harder work: scale the reactor, prove the economics, satisfy regulators and show that a 30-megawatt campus can run on more than a clever demonstration.
Also read: Anthropic Opens Talks With Samsung to Build Its First Custom AI Chip • Alibaba and Tencent Back Kling AI Instead of Racing to Beat It • Jersey Mike's Mentioned AI 22 Times In Its Billion Dollar IPO Filing
