Asked whether nuclear is the answer for AI data centers, energy investor Ramez Naam gave two answers at once. “AI data centers might be the best thing that's ever happened to the nuclear industry,” he said on Moonshots with Peter Diamandis. Yet large cloud companies talking about small modular reactors as their power supply are looking beyond the five-year investment window he believes they can see clearly.
That gap—between money arriving now and electricity arriving later—runs through his assessment of nuclear fission, the process of splitting atoms to release heat that a power plant converts into electricity. AI customers could help finance enough reactors to rebuild an industry's skills and supply chains. That does not mean a new reactor will be ready when a data center needs it.
The power you can get first
Naam's first moves involve plants that already exist: stop shutting them down, extend their operating lives where possible, upgrade them to produce more power, and restart suitable closed plants. He points to Three Mile Island as a restart opportunity. An upgrade that increases a plant's permitted output is called an uprate; it adds capacity without building an entirely new reactor.
His own accounting of those measures is modest: a few gigawatts, with each gigawatt equal to 1,000 megawatts of generating capacity. For a much larger expansion, he turns to the cost of building new plants.
Outside China and perhaps South Korea, Naam describes nuclear construction as prohibitively expensive. His explanation centers on repetition. An industry that builds occasionally does not develop the same practiced crews, dependable suppliers and refined designs as one that keeps building.
An order book, not a one-off plant
One route is to keep constructing large reactors such as Westinghouse's AP1000, a roughly gigawatt-scale design. Naam argues that financing a fleet makes more sense than treating every plant as an isolated project.
The first reactor of a new model is likely to run over budget and schedule, he says, with problems its builders did not anticipate. Later units give crews a chance to apply what they learned, engineers time to resolve design issues, and suppliers a reason to invest in producing parts. A multi-unit order book can spread early costs and risks across a larger program rather than leaving the first customer to absorb them alone.
France is his illustration in both directions. It built a large nuclear fleet by repeating closely related designs. But Naam argues that its interrupted construction cadence helped erode the expertise and industrial capacity needed for later projects. He points to delays and overruns with the European pressurized reactor as the unhappy side of that history.
Once construction stops, he says, an industry loses not just experience but the suppliers and facilities that make its parts. Restarting then becomes harder and more expensive. In his account, the benefit of repetition depends on maintaining it.
This is where AI customers could matter: their demand could support orders for multiple plants, giving builders work beyond the difficult first unit. Naam's proposed arrangement is either one buyer willing to order enough reactors to pursue lower costs, or several buyers agreeing to share costs and risks.
Smaller is not automatically cheaper
The other route is small modular reactors, or SMRs: smaller plants designed to shift more work from construction sites into repeatable factory production.
Naam was enthusiastic about them 15 years ago, cooled on the idea, and is now cautiously hopeful again. The appeal is manufacturing. A factory can repeat the same operations, refine its tooling and carry lessons directly from one unit to the next. In the most ambitious version, it produces a reactor that can be shipped to its destination by truck or barge.
But a plant assembled from factory-made components is not the same thing as a fully factory-built reactor. Many modular designs still require substantial work at the site.
There is also a genuine advantage to building big. As Naam explains it, large plants use less steel and cement per unit of power output. Smaller reactors give up some of that material efficiency in pursuit of savings from manufacturing more units.
He worries that medium-sized designs, including TerraPower's Natrium, could occupy an awkward middle: too large to finish entirely in a factory, yet too small to capture all the economies of a large plant. He does not rule them out. They might, he says, turn out to be the best compromise.
Natrium has another way to change the economics
Natrium's design is not simply a smaller version of a conventional plant. TerraPower's technical overview describes a sodium-cooled reactor coupled to molten-salt heat storage. The reactor supplies 840 megawatts of heat, and the plant has a baseline electrical output of 345 megawatts. Drawing on stored heat is intended to let electrical output rise to around 500 megawatts for more than five hours while the reactor's heat production remains steady.
The design separates the nuclear island—the reactor side—from the energy island, which stores heat and generates electricity. TerraPower intends that separation to reduce how much equipment needs nuclear-grade construction and allow electricity production to respond to changing demand. Those are additional proposed sources of value, beyond the question of reactor size; they do not eliminate the need to demonstrate construction costs and delivery schedules.
At the smaller end, microreactor developers such as Radiant seek to move reactor assembly fully into the factory. Naam also describes military interest in compact reactors that could reduce the need for repeated fuel deliveries to bases. The economic wager is that repeatable production can outweigh the material disadvantages of going small.
Safe shutdown is not continued power
The discussion also touches on passive safety: using physical processes rather than electrically powered equipment to perform safety functions. That should not be confused with keeping a data center running through a blackout.
Westinghouse's AP1000 explanation describes automatic shutdown followed by cooling through gravity, convection and condensation. The design provides 72 hours of core cooling without operator intervention or AC power; afterward, operators replenish water to maintain cooling. The protection concerns reaching and sustaining a safe shutdown, not continuing to generate electricity.
A shipment of fuel is a step, not a power supply
Naam describes developers' optimistic schedules for the smaller reactors under discussion as around 2030 to the early 2030s, and expects slippage. Smaller units may slip less than larger ones, he suggests, but their first versions will still be expensive. His timetable is an assessment, not a shared industry deadline.
Radiant, for example, has stated an earlier ambition. On July 1, 2026, the company announced that its first shipment of TRISO nuclear fuel, manufactured by Standard Nuclear, had arrived at Idaho National Laboratory's DOME facility. It outlined a five-phase test program for its Kaleidos reactor, progressing from initial criticality—the start of a sustained chain reaction—through increasing power and temperature. The eventual target was at least 150 hours at full power without operator intervention.
Radiant said the resulting reactor and fuel-performance data would support commercial licensing and customer delivery by 2028. That target is earlier than Naam's expectation, but the announcement establishes fuel receipt and planned testing, not completed commercial deployment. Before a customer can count those megawatts toward a data center's power supply, the reactor still has to move through testing, licensing and delivery.