TerraPower Uses Molten Salt Storage to Power AI Data Centers
Bill Gates-backed nuclear startup leverages thermal energy storage to smooth volatile GPU workloads without throttling reactors
Nuclear reactor startup TerraPower is pitching its 345-megawatt Natrium design as a flexible solution for AI data center power demands. By storing excess nuclear heat in a integrated vat of molten salt, the facility can rapidly absorb and dispatch energy to smooth volatile GPU training spikes without throttling atomic output or relying on costly battery arrays.
Key details
Traditional nuclear reactors operate most efficiently at constant maximum output, achieving a high 92.5% capacity factor in the U.S. However, standard commercial reactors ramp power up or down slowly at only 5% of rated output per minute, while advanced small modular reactors (SMRs) manage approximately 10% per minute.
For behind-the-meter AI data centers, rapidly changing compute loads during heavy training runs or inference queries create severe power fluctuations. These swings have caused natural gas turbines to suffer mechanical failure, forcing operators to install capital-intensive battery systems to stabilize grid voltage.
TerraPower's 345-megawatt molten salt-cooled reactor avoids these operational constraints by continuously splitting atoms at peak capacity. When data center energy demand drops, excess thermal energy is diverted into a molten salt storage system. When compute loads suddenly surge, the plant draws from this thermal reservoir to rapidly generate additional steam, ramping electrical output without adjusting the nuclear core or idling expensive capital equipment.
Why this matters
The rapid expansion of AI infrastructure is exposing fundamental limits in conventional generation assets. Large-scale GPU clusters subject local microgrids to extreme, instantaneous load shifts that conventional baseload plants cannot track.
By integrating thermal energy storage directly into the generation architecture, TerraPower addresses both the high capital expenditure of nuclear power and the load volatility of AI workloads. Storing thermal energy in molten salt enables high-capacity factor operation while providing load-following capability, eliminating the need for standalone battery storage farms and mitigating grid integration bottlenecks.
Context
Nuclear energy developers are positioning small modular reactors as a vital, carbon-free power source for hyperscale data center operations. TerraPower, founded by Bill Gates, is currently constructing its first commercial nuclear plant in Wyoming.
In January, Meta reached an agreement to purchase power from eight TerraPower Natrium plants. Reports indicate that TerraPower plans to unveil its first dedicated AI data center customer project later this year, with construction slated to begin in 2027.
What happens next
TerraPower is preparing to announce its inaugural AI data center partnership before the end of the year, detailing site selection and grid interconnection plans. Industry analysts will monitor whether molten salt thermal buffering can deliver lower total cost of ownership compared to gas turbine and battery microgrids as ground breaking approaches in 2027.
Source: TechCrunch Published on AI Usage Global, author: AUG Bot
