The U.S. Nuclear Regulatory Commission has cleared a major safety review for an advanced nuclear reactor in Wyoming, removing one of the last regulatory barriers before a construction permit can be issued. The decision lands as Amazon, Google, and Microsoft pour billions into nuclear energy to feed data centers that already consume a fast-growing share of the national grid. Together, these moves signal that the age of nuclear-powered supercomputing is no longer theoretical, even if the hardest engineering and timeline questions remain unanswered, and the race to power AI-scale computing is accelerating.
NRC Greenlights TerraPower’s Natrium Reactor
The NRC staff issued its final safety evaluation for TerraPower’s Kemmerer Unit 1 construction permit application and concluded that no safety aspects would preclude issuing the permit. That finding followed the publication of the final Environmental Impact Statement, designated in NUREG-2268, which evaluated alternatives and environmental consequences for the sodium-cooled Natrium reactor planned for a site near Kemmerer, Wyoming. Taken together, the two documents represent the most advanced licensing progress any next-generation reactor design has achieved in the United States in decades.
What makes this relevant to the computing world is the type of customer these reactors are designed to serve. A sodium-cooled fast reactor can deliver steady, carbon-free baseload power at a scale that matches the round-the-clock demand profile of a large data center or supercomputing cluster. Traditional renewables like wind and solar fluctuate with weather and time of day, forcing operators to rely on battery storage or fossil-fuel backup. A reactor that runs continuously for years between refueling cycles eliminates that gap. If the Kemmerer construction permit is formally granted, the project would become a proof point for every tech company eyeing nuclear as the answer to AI’s electricity appetite.
Data Centers Are Devouring the Grid
The urgency behind these nuclear investments becomes clear in the numbers. Data centers consumed 176 TWh of electricity in 2023, accounting for approximately 4.4% of total U.S. power consumption. By 2028, the U.S. Department of Energy projects that share will climb to between 6.7% and 12%, with absolute consumption reaching 325 to 580 TWh depending on the growth scenario. That range reflects deep uncertainty about how quickly AI training runs, inference workloads, and new facility construction will scale, but even the low end nearly doubles current demand within five years.
The Lawrence Berkeley analysis produced the underlying projections in a Congress-mandated update that tracks U.S. data center electricity use historically back to 2014 and models scenario ranges to 2028 using server shipment data and updated assumptions. The associated technical dataset gives policymakers and grid operators a detailed baseline for planning, but it also exposes a stark gap, the country’s existing generation and transmission infrastructure was not built for this kind of demand surge. New power plants and high-voltage lines take years to permit and construct, which is precisely why the nuclear option has gained so much corporate attention so quickly.
Microreactors Move from Concept to Concrete Testing
While large reactors like Natrium target utility-scale output, a parallel track is advancing smaller units that could sit adjacent to, or even on the grounds of, individual computing facilities. The Department of Energy selected Westinghouse and Radiant for the initial experiments at Idaho National Laboratory’s DOME microreactor test bed, with the first fueled reactor experiment potentially starting as early as spring 2026. Microreactors generate far less power than a full-scale plant, but their compact footprint and factory-built design could allow them to be deployed where traditional grid connections are weak or nonexistent.
For a hyperscale data center operator, the appeal is straightforward: a microreactor co-located on site would reduce dependence on congested transmission lines and insulate operations from regional grid stress. The spring 2026 target for fueled testing at DOME is aggressive, and hardware demonstrations often slip, especially when they involve first-of-a-kind nuclear technology. Still, the fact that DOE has moved from concept studies to selecting specific vendors and scheduling physical experiments marks a real acceleration in the federal pipeline. If these tests validate the technology and regulators are satisfied with safety and security performance, commercial microreactor deployments could follow within a few years, giving data center builders a new tool for siting facilities in locations that today lack sufficient power infrastructure.
Big Tech’s Billion-Dollar Nuclear Bets
Amazon, Google, and Microsoft are investing billions in nuclear power, ranging from restart agreements at shuttered plants to funding for entirely new reactor designs. The scale of the spending reflects a shared calculation, AI model training and inference will require so much electricity that no single energy source can fill the gap alone, and nuclear is the only zero-carbon option that delivers firm, always-on power at the gigawatt scale. For companies that have pledged steep reductions in greenhouse-gas emissions, buying more fossil power to run AI clusters would undermine their own climate commitments.
Yet the gap between investment announcements and operating reactors is wide. These projects are years away and rely on technology that has not yet been proven at commercial scale, from advanced fuels to novel cooling systems and digital control architectures. U.S. technology companies are pushing into nuclear power with a speed that contrasts sharply with the slow pace of traditional utility planning, but the history of reactor construction in the United States is filled with cost overruns and schedule delays that caution against assuming everything will arrive on time. Even if TerraPower’s Natrium plant in Wyoming and DOE-backed microreactor demonstrations hit their current milestones, the industry will still need to prove that these designs can be replicated, financed, and interconnected to the grid at the pace AI-era data centers demand.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
