TerraPower Reactor Stores Heat to Power AI Data Centers
Get the Energy newsletter
Daily energy & climate — solar, EVs, oil, the policy fights and tech bets shaping the transition. Free.
- TerraPower plans to announce its first data center project this year, per Bloomberg, with the facility expected to break ground in 2027 as the startup's second power plant after one already under construction in Wyoming
- Meta agreed in January to buy eight of TerraPower's Natrium power plants, though TerraPower has not identified its data center customer
- TerraPower's 345-megawatt molten salt-cooled reactor stores excess heat in a vat of molten sodium, letting the reactor run at full capacity while tapping stored heat to spin turbines during demand spikes
- Nuclear reactors operate at a 92.5% capacity factor in the U.S., the highest of any power source, but existing designs can only ramp output up or down about 5% per minute, per the National Laboratory of the Rockies — poorly suited to AI's swinging loads
- Small modular reactors from competitors ramp about 10% per minute but lose money at reduced capacity, and nuclear's high capital expenditures make running at peak output essential until mass manufacturing brings costs down
- TerraPower originally designed its reactor to complement intermittent renewables like wind and solar; that same ramp-flexibility translates directly to absorbing the variable loads from AI training and inference workloads
Why it matters: TerraPower's molten salt storage turns nuclear's inability to ramp output quickly into an asset for AI data centers, where GPU-driven load swings have already broken natural gas turbines and forced costly battery bank additions. Because the reactor stays at full power and stashes excess heat in molten sodium, TerraPower can amortize its high nuclear capex over more operational hours than SMR competitors that bleed money running below peak.
Ask SkimNews
