TerraPower expects to announce its first dedicated data center project this year, pairing a nuclear reactor with a thermal storage system designed to absorb the abrupt changes in electricity demand created by AI computing. The customer remains unnamed, and the project has not yet demonstrated its economics or performance in operation.
The company told TechCrunch that construction is expected to begin in 2027. It would be TerraPower’s second Natrium project after a plant already under construction in Wyoming. Meta separately agreed in January to buy the output of eight future Natrium plants, illustrating how closely advanced nuclear developers are now courting hyperscale computing customers.
AI loads challenge steady power plants
Nuclear generation is prized for reliability. US reactors run at an average capacity factor of 92.5%, according to figures cited by TechCrunch, producing electricity far more consistently than weather-dependent sources. That steady output is attractive to data centers expected to operate around the clock.
AI workloads, however, are not perfectly steady. Thousands of processors can change power consumption quickly as jobs begin, end or shift between stages. The resulting swings have reportedly damaged gas turbines, while batteries capable of buffering very large facilities remain expensive.
Conventional nuclear reactors can adjust output, but not instantly. Existing plants generally ramp at about 5% of rated capacity per minute, and small modular reactors may reach roughly 10%, according to the National Laboratory of the Rockies. Those rates may still be too slow for sharp changes in GPU demand.
TerraPower’s proposed answer is to avoid forcing the nuclear core to follow every fluctuation. Its 345-megawatt Natrium reactor can continue producing heat at full output while the rest of the system decides whether that heat should become electricity immediately or be stored.
Molten sodium acts as a thermal buffer
Excess heat is held in a large reservoir of molten sodium. When computing demand rises, the stored energy can generate additional steam for the turbines; when demand falls, more heat accumulates in the reservoir. The reactor itself can therefore operate at a stable level while electrical output changes around it.
TerraPower originally designed the arrangement to complement variable renewable generation. The same architecture could suit a data center for the opposite reason: instead of smoothing an intermittent supply, it would smooth a volatile customer.
The company also argues that keeping the reactor and other expensive equipment fully utilised improves the economics of the plant. Thermal storage may be cheaper than building enough batteries to handle comparable energy shifts, although TechCrunch does not provide a direct cost comparison for the proposed project.
The concept has practical appeal, but it remains a vendor proposition. The unnamed data center plant has not broken ground, construction schedules can change, and neither its final cost nor its performance under real AI workloads is available. Molten-salt and sodium systems also introduce engineering and operating requirements that must be proven over time.
Flexible output does not erase nuclear’s cost problem
Advanced reactors still face nuclear power’s largest obstacle: capital expense. New plants require substantial investment before producing any electricity. Small modular reactor companies expect factory manufacturing to lower those costs, but the source notes that the savings remain unproven and may take more than a decade to materialise.
Data center customers could help by committing to long-term purchases, giving developers predictable demand and financing support. Yet a reliable buyer does not automatically make construction cheap or fast. TerraPower must show that its storage-equipped design can be licensed, built and operated at a competitive price.
If the project succeeds, its significance will extend beyond simply attaching a reactor to a server campus. The thermal reservoir could let a fundamentally steady generator respond to machines whose appetite changes by the second. That would make nuclear power more compatible with AI infrastructure without sacrificing the high utilisation needed to justify its cost. For now, the design offers a credible mechanism and a planned customer project—not proof that the commercial equation has been solved.
