The pitch from Silicon Valley sounds straightforward: build small, advanced nuclear reactors next to the data centers that train AI models, and the electricity problem solves itself. Three projects in particular have attracted billions in backing from tech giants eager to lock in carbon-free baseload power. But federal licensing records reviewed through early June 2026 tell a different story. Not one of these reactors is on track to produce a single watt of grid electricity before the year is out, and the gap between corporate announcements and regulatory reality could leave AI infrastructure plans short of clean power well into the next decade.
Where each reactor actually stands
The project closest to breaking ground is TerraPower’s Natrium reactor, a sodium-cooled fast reactor planned for Kemmerer, Wyoming, backed by Bill Gates and supported by a $2 billion cost-share award from the Department of Energy. On December 1, 2025, the Nuclear Regulatory Commission completed its final safety evaluation and issued construction permit CPAR-1 to US SFR Owner, LLC, a TerraPower subsidiary. That permit authorizes building the physical plant. It does not authorize loading fuel, achieving first criticality, or selling electricity. Each of those steps requires additional NRC review, and for a first-of-a-kind reactor design, the construction-to-operation timeline typically stretches several years. No NRC docket entry or TerraPower filing specifies a firm date for fuel loading or grid synchronization.
Oklo’s Aurora microreactor sits further back in the queue. The company’s SEC quarterly filing projects a target of late 2027 or early 2028 for its first powerhouse deployment. That projection is Oklo’s own, not an NRC-endorsed schedule. The NRC’s status page lists Aurora’s proposed output at 75 MWe and confirms the project remains in pre-application activities, a defined regulatory status meaning Oklo has not yet submitted the formal application that triggers the NRC’s multi-year construction permit review. Notably, the NRC denied Oklo’s earlier Aurora design application in January 2022, citing insufficient safety information. The company restarted the process under a revised approach, but pre-application engagement can itself last years, and completion depends on fuel qualification results and NRC staff availability, neither of which has a publicly confirmed date.
Kairos Power’s Hermes reactor in Oak Ridge, Tennessee, received a construction permit on December 14, 2023, making it the first new non-light-water test reactor permitted in the United States in decades. But Hermes is a 35 MWth test reactor designed to validate fluoride-salt-cooled technology. It generates no electricity and is not licensed to export power to any grid. Counting Hermes as part of an AI energy supply chain overstates what the hardware can deliver. Its contribution will be engineering data and licensing experience, not megawatts.
A fourth project frequently discussed alongside these three is X-energy’s Xe-100 high-temperature gas reactor at the Long Mott Generating Station in Texas, developed in partnership with Dow. X-energy submitted its construction permit application on March 31, 2025, proposing four modules totaling 320 MWe. The NRC’s environmental and safety reviews are in their earliest stages, with consultations under the Endangered Species Act and Section 106 of the National Historic Preservation Act still underway. Until the NRC publishes a detailed review schedule, any projection of an in-service date remains speculative.
The demand side is not waiting
Tech companies are not standing still while regulators work. Meta has announced nuclear supply agreements with TerraPower, Oklo, and energy company Vistra, and stated publicly that it is assembling up to 6.6 GW of new and existing clean energy capacity by 2035. Microsoft has signed a 20-year power purchase agreement to restart a unit at Three Mile Island. Amazon has invested in X-energy and purchased a nuclear-powered data center campus in Pennsylvania.
But the specific share of Meta’s 6.6 GW target that depends on reactors still in pre-application or early construction review has not been disclosed. Whether any of these tech buyers have contracted interim power sources to bridge the years between deal announcements and actual reactor output is not addressed in available filings. The 2035 horizon provides more runway than a year-end deadline, yet the near-term electricity gap is real: AI training clusters already consume hundreds of megawatts per campus, and demand projections from grid operators continue to climb.
Why the timelines keep slipping
Several structural factors explain the distance between ambition and execution.
First, the NRC’s own capacity is a bottleneck. Multiple commissioners have publicly acknowledged that reviewing novel reactor designs requires specialized staff the agency is still hiring and training. Every first-of-a-kind application demands new safety frameworks, and the commission cannot shortcut that work without undermining its statutory mandate.
Second, fuel supply remains unresolved for several of these designs. Both Natrium and Xe-100 require high-assay low-enriched uranium (HALEU), a fuel type that has no established commercial supply chain in the United States. The Department of Energy has funded HALEU production through Centrus Energy’s demonstration cascade in Piketon, Ohio, but output remains limited. Until fuel fabrication scales up, reactor construction timelines are constrained by a supply chain that does not yet fully exist.
Third, the recent history of small modular reactors in the U.S. offers a cautionary data point. NuScale Power’s Carbon Free Power Project with the Utah Associated Municipal Power Systems, once the most advanced SMR effort in the country, was canceled in November 2023 after cost estimates nearly doubled and subscriber utilities pulled out. That collapse did not doom the broader SMR sector, but it demonstrated how quickly first-of-a-kind economics can unravel, even with substantial federal support.
What the regulatory record actually tells us
The strongest evidence in this story comes from primary federal records: NRC permit issuance pages, safety evaluation announcements, and SEC filings. These documents carry legal weight. When the NRC classifies Oklo as being in pre-application activities, that is a defined regulatory status with specific procedural implications. When Oklo’s 10-Q filing names late 2027 as a target, the company’s officers have certified that projection to investors under securities law.
Corporate press releases and partnership announcements describe purchasing intent, not regulatory reality. Meta’s 6.6 GW goal and its named reactor partnerships signal strategy, but they do not accelerate the pace at which the NRC reviews safety cases or issues permits. Readers should treat procurement announcements as demand signals and NRC docket entries as the authoritative record of what can actually be built and when.
It also matters to distinguish between project types. A non-power test reactor like Hermes can move through licensing faster because it does not export electricity and operates at lower power densities than a commercial plant. Demonstration units like Natrium and Xe-100 occupy a middle ground: intended to prove new technologies while eventually selling power, which complicates both financing and regulatory scrutiny. Microreactors like Aurora promise modularity and faster deployment, but their novel fuel types and compact designs raise questions regulators must resolve case by case.
Clean baseload power for AI is coming, but not this year
None of this means the nuclear deals are empty gestures. Long-term power purchase agreements and equity investments help developers secure financing and demonstrate the market demand that justifies federal support and regulatory resources. The documented timelines, however, point to a clear sequencing: first come years of licensing and construction, then come electrons. The earliest realistic window for commercial power from any of these advanced reactors falls closer to the late 2020s or early 2030s than to December 2026.
In the meantime, tech firms will likely rely on a patchwork of grid purchases, utility-scale solar and wind, battery storage, and life extensions at existing conventional nuclear plants to keep pace with surging AI workloads. Those interim choices will shape the sector’s emissions trajectory long before the first sodium-cooled or gas-cooled reactor synchronizes with the grid. For now, the race to power AI with advanced nuclear is real, but the finish line is further away than the announcements suggest.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
