Nvidia is backing a nearly waterless nuclear data center in Utah

A rural stretch of central Utah has become the setting for one of the more unusual experiments in the race to power artificial intelligence: a data center designed to run on a startup’s own nuclear reactor while drawing almost no water from the surrounding desert. The plan pairs a young reactor company with the world’s dominant maker of AI chips, and it was unveiled at a live demonstration on the reactor site itself in early July 2026.

Water has quietly become one of the AI industry’s touchiest constraints. Conventional data centers evaporate enormous volumes to shed the heat their servers produce, a burden that grows heavier as facilities push into arid states with strained supplies. A design that promises steady nuclear power and near-zero local water use at the same time speaks directly to two of the biggest objections communities raise when a large computing campus comes to town.

The partners and the site

The project brings together Valar Atomics, a reactor developer founded in 2023, and chipmaker Nvidia, whose processors sit at the heart of nearly every large AI system. The two announced their collaboration on July 1 at Valar’s site in Orangeville, Utah, describing plans to explore data centers powered directly by advanced nuclear energy, according to an account in the Deseret News. The companies cast the pairing as a first for the industry, matching a startup-built reactor with the computing hardware it would eventually feed.

The centerpiece of the announcement was a demonstration in which Nvidia’s Blackwell-generation chips were run on electricity supplied by Valar’s reactor, a moment the companies presented as proof that the two technologies could be joined in practice rather than only on paper. From there the partners sketched a larger ambition: a roughly 30-megawatt facility in Utah built to drink almost no local water.

How the water figure works

The waterless claim rests on the cooling method rather than the reactor. Nvidia’s data-center reference design uses closed-loop liquid cooling, which circulates the same coolant in a sealed system instead of evaporating fresh water to carry heat away. Coverage of the announcement put the difference starkly, citing a drop from roughly 2.6 million gallons of water per megawatt each year in a conventional evaporative setup to near zero in the closed-loop approach, as detailed in a report from Utah broadcaster KSL.

Eliminating evaporative cooling does not make a data center entirely water-free, but it removes the single largest draw and reframes the facility’s demand from an ongoing withdrawal into a mostly one-time fill. In a region where agriculture and municipalities already compete for scarce supplies, that shift is the difference between a project that alarms local water managers and one they may be willing to permit.

The reactor at the center

The power source is Valar’s Ward250, a small reactor the company says went critical on June 18, 2026, and now generates about 100 kilowatts of electricity. Reaching criticality means the reactor sustained a nuclear chain reaction on its own, the threshold that separates an assembled machine from an operating one. That output is a tiny fraction of the eventual 30-megawatt target, underscoring that the Utah plan is a demonstration meant to prove a concept before any full-scale plant exists.

Valar did not reach that milestone in isolation. The Ward250 was one of several advanced reactors that hit criticality this summer under a federal effort to accelerate the technology, work the government has documented as part of its push to prove new designs quickly, according to the Department of Energy. Being folded into that program gives a startup reactor a measure of federal oversight and a testing pathway that would otherwise take far longer to assemble.

What the desert test is really probing

Behind the headline pairing sits a practical bet that the AI industry can escape two bottlenecks at once. Grid interconnection queues have stretched into years in many regions, and the water demands of large data centers have sparked local opposition from Arizona to Georgia. A self-contained reactor sited next to the servers would bypass the transmission wait, and a closed-loop cooling loop would blunt the water fight, letting a developer promise a community both a dedicated power supply and a minimal draw on the aquifer.

The distance from a 100-kilowatt test reactor to a 30-megawatt commercial plant remains enormous, and the announcement did not resolve the licensing, financing, and construction questions that separate the two. Selling reactor power commercially to a private data center invites a stricter regulatory review than running a demonstration unit, and a company founded only three years ago carries little track record for regulators and lenders to weigh. The economics of building bespoke reactors for individual computing campuses are likewise unproven against simply buying grid power or gas generation.

Still, the Utah experiment has moved a much-discussed idea onto real ground. A live chip demonstration, a reactor that has reached criticality, a named site, and a chipmaker willing to attach its brand together push the concept of reactor-adjacent, water-light AI past the slide-deck stage. Whether the Orangeville project scales into the promised facility or stalls short of it, the pairing has drawn attention to a version of the AI build-out that tries to answer the power and water objections in the same stroke, and the coming construction phase will test whether that promise survives contact with a real budget and a real permit, as observers watching the rural site have already begun to note in local coverage of the experiment.

This article was produced with the assistance of AI and reviewed by Morning Overview editors prior to publication.

More from Morning Overview