Four U.S. startups have fired up their first small nuclear reactors, aiming them straight at power-hungry AI data centers

Four U.S. nuclear-reactor startups notched a milestone the industry had chased for decades this year: each fired up a small, non-commercial reactor and achieved a self-sustaining nuclear reaction, known as criticality, around a federal deadline meant to jump-start the sector. The companies, Antares Nuclear, Valar Atomics, Aalo Atomics and Oklo, are now positioning that early hardware as a first step toward supplying the enormous and fast-growing electricity appetite of AI data centers, a market traditional utilities are struggling to keep pace with.

A federal deadline that four companies beat

The Department of Energy’s Reactor Pilot Program set a goal of reaching criticality in at least three advanced reactor concepts built outside the national laboratories by July 4, 2026, a deadline meant to compress years of regulatory review into months. Antares Nuclear reached criticality on June 4, Valar Atomics followed on June 18, and Aalo Atomics announced its own milestone in early July, with a fourth microreactor developer confirming criticality around the same window, pushing the program past its original target of three.

That fourth confirmation, covered in trade-press coverage of the deadline, rounded out a group of four companies running their own small test reactors rather than a single shared design.

What criticality does and doesn’t prove

Criticality means a reactor’s core has reached a self-sustaining fission chain reaction, the basic physics milestone every nuclear reactor must clear before it can generate meaningful power. It does not mean the reactor is connected to a power grid, producing commercial electricity or licensed for continuous operation. Aalo Atomics, for instance, described its milestone as achieved in a test reactor built specifically to prove the design before scaling up, a distinction that separates this wave of announcements from an actual operating power plant.

Each of the four companies is also pursuing a different reactor technology rather than a shared design, which is part of the point of the federal pilot program: testing several competing approaches in parallel rather than betting early on a single winner. That variety means the four reactors differ in size, fuel type and cooling method, and each will need its own separate licensing path through the Nuclear Regulatory Commission before any of them can be built at a scale large enough to power an actual data center.

AI data centers become the customer everyone wants

The push behind these startups is less about general electricity demand than one specific customer: AI data centers, whose power draw per facility now rivals that of a small city and keeps climbing as computing clusters grow larger. Large technology companies have already signed development agreements with multiple reactor developers to help secure future power supplies, betting that nuclear plants sized and sited near data centers can deliver the steady, round-the-clock output that intermittent solar and wind cannot easily match on their own.

That demand is reshaping how reactor developers pitch themselves to investors and regulators alike, since a data-center operator generally wants a long-term, fixed-price power contract rather than a stake in an unproven technology, and several of the newest reactor designs are explicitly sized to sit on or near a single large campus rather than feed a broad regional grid. The arrangement also suits utilities under pressure to add generation capacity quickly, since a nuclear plant built and financed primarily for one large customer shifts much of the upfront cost and risk away from ordinary ratepayers.

Bigger commercial projects are already under construction

Separate from the four criticality announcements, TerraPower, a more established nuclear developer, began construction on its first commercial-scale Natrium reactor in Kemmerer, Wyoming, on the site of a retiring coal plant, with commercial operation not expected until around 2030. That project, described in reporting on the company’s data-center ambitions, illustrates how far even the industry’s most advanced private projects still are from actually delivering power, despite the recent run of criticality headlines from smaller rivals.

The gap between a test reactor and a power plant

The Department of Energy has cataloged both the appeal and the obstacles of pairing nuclear power with data centers, noting in its own assessment of the approach that reliable, carbon-free baseload power is the draw, while lengthy licensing timelines, high upfront construction costs, water availability and workforce shortages remain real constraints. None of the four startups that reached criticality this summer has announced a reactor actually delivering power to a data center, meaning the gap between this year’s test-reactor headlines and an operating commercial plant is still measured in years, not months.

Historically, building a new nuclear plant in the United States, from site selection through licensing, construction and grid connection, has taken a decade or longer, and cost overruns have derailed more large reactor projects than they have delivered on schedule. The Reactor Pilot Program was designed specifically to compress the earliest, slowest part of that timeline for smaller designs, but the harder, more capital-intensive stages, full-scale construction, commercial licensing and long-term operation, still lie ahead for every company that reached criticality this summer.

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

More from Morning Overview