AI Needs Atoms Now. New Nuclear Arrives Next Decade.

The federal government's flagship nuclear loan this year is $4.2 billion. What it buys is 433 megawatts, and not a single new plant: three existing Vistra reactors get coaxed to run harder, their licenses stretched another 20 years ^[1]. Set that against the ask in one state alone. Texas's grid operator has logged 466,497 megawatts of data-center interconnection requests, more than five times the most power that grid has ever actually delivered ^[2]. Nationwide, the reliability watchdog warned last winter that demand was running 20 gigawatts above the year before ^[3]. The shortage is now. The reactors are not. The dates do the arguing. Bank of America projects a $10 trillion nuclear market over 30 years, then tucks the caveat into its own analysis: meaningful adoption of small modular reactors, the factory-built designs meant to sidestep the decade-long megaplants of the past, begins in 2030 to 2035 ^[4]. X-energy calls nuclear "truly the only real answer" to the steady, round-the-clock power the big AI firms need; its contract with Amazon covers 64 reactors and is dated 2039 ^[5]. Nano Nuclear, a company with no revenue and no operating license, signed a non-binding framework promising 6 gigawatts for AI industrial zones, by 2040 ^[6][7]. Even the fast path, restarting a dead plant, runs on paperwork: Constellation does not expect its Three Mile Island reviews to be finished until the end of 2027 ^[8]. By the industry's own paperwork, the answer sits a decade out while the shortage was booked for last winter. What actually arrives this decade is resurrection. The Nuclear Regulatory Commission has approved the restart of Palisades in Michigan. It is the first time a shuttered US commercial reactor has been cleared to come back, and it comes with a $1.52 billion loan guarantee ^[8]. The Vistra loan is the same move by another name: an uprate, squeezing more megawatts out of a reactor that already exists rather than building one. Meta has already locked 20-year power contracts onto those very plants ^[1]. What gets built for AI today is gas, increasingly behind the meter: on the data center's own property, past the utility and the grid entirely. Meta is installing 800 mini-turbines in El Paso; OpenAI is buying 29 jet-derived turbines from Boom Supersonic; 58 gigawatts of gas power are in development in Texas alone ^[9]. The fuel the industry says it wants arrives in the 2030s. The fuel it can install arrives in months. Here is where the familiar story breaks. The wall everyone assumes is holding nuclear back, the regulator, is dissolving ahead of schedule. In September the NRC issued the first-ever US construction permit for a small modular reactor, to the Tennessee Valley Authority, finishing its safety review in 14 months, four months ahead of its own timetable ^[10]. A House committee has unanimously advanced six bills to streamline licensing further ^[11]. Licensing is no longer the binding constraint. The real ceilings start with a number most of the industry would rather not repeat: Vogtle, the last large reactor completed in the United States, ran more than $30 billion over budget ^[12]. That single figure is why TVA, holding America's first small modular reactor permit, has announced no construction date, and is still negotiating who pays so that its local ratepayers do not shoulder the cost of a first-of-a-kind technology ^[10]. Private capital watched Vogtle and will not price that risk. Then there is the line to get plugged into the grid: the Federal Energy Regulatory Commission just rejected Oklo's bid to rejoin the PJM interconnection queue for a 750-megawatt project, a setback Oklo says delays it more than a year ^[13]. And there is the fuel itself. Advanced reactors run on high-assay low-enriched uranium, and exactly one American company, Centrus, is approved to produce it ^[14]. The ceilings that remain are the ones Washington cannot deregulate away: the queue, the construction bill, the fuel line. Which is how a technology pitched as the only real answer stopped being a business and became a government program. Under a binding term sheet for at least $80 billion of Westinghouse reactors, the government arranges financing and permitting, is entitled to 20% of Westinghouse's cash distributions above $17.5 billion, and may take a minority stake if the company goes public by January 2029 ^[15]. The Energy Department has gone further: it planned to directly purchase and own up to ten reactors ^[15]. The fuel chain, meanwhile, is being stood up under the Defense Production Act, the wartime authority used to commandeer supply lines, because mining, conversion, and enrichment gaps leave the country dependent on foreign uranium ^[16]. Any claim of self-sufficiency runs through Canada regardless: Cameco supplied 17% of the world's uranium in 2024 and owns 49% of Westinghouse itself ^[17]. The department's chief of staff put it plainly.

We expect this new partnership to support the global growth opportunities for both Westinghouse's and Cameco's nuclear products, services, and technologies, adding significant long-term value for our stakeholders and enhancing energy, national, and climate security around the world. — Cameco

One deadline in this whole ledger was not set by a market. The government set it for itself: three pilot reactors operating by July 4, 2026. Meeting it has meant rewriting the department's own safety rules. Groundwater protection softened from "must" to "should be avoided." Security requirements were cut from more than 500 pages to 23. Oversight moved from the independent NRC to the department itself ^[18]. A department attorney was blunt about the pace.

That is the speed at which we are expecting you to move. And our job is to meet you. — Seth Cohen

That date, July 4, was set by fiat. Every other reactor in the ledger still waits on a queue, a negotiation, or a fuel line.