TerraPower bets on molten sodium heat storage to match AI data centre load fluctuations

Nuclear power startups are racing to position themselves as a definitive solution to the biggest challenge facing AI data centres, the need for a uninterrupted, round-the-clock power supply. TerraPower, founded by Bill Gates, is preparing to announce its first nuclear plant project dedicated to data centres this year, Bloomberg reported. Although the company declined to name the prospective client, the move follows its announcement in January of an agreement with Meta to purchase eight Natrium power plants. Construction on the data centre project is expected to begin in 2027, making it the company's second plant after its first facility currently under construction in the US state of Wyoming.

Not all nuclear reactors are suited to how data centres operate. Conventional nuclear reactors achieve the highest capacity factor of any generation technology in the United States, producing their maximum output 92.5 percent of the time. However, they are extremely slow to ramp production up or down, adjusting output by no more than roughly 5 percent of rated capacity per minute, according to data from the Rocky Mountain National Laboratory. Even small modular reactors under development by startups, which respond at roughly 10 percent per minute, yield lower returns when operated below full capacity because their capital expenditures are the highest among all power generation technologies, pending proof that volume manufacturing can reduce costs over a decade or more.

The operational nature of data centres imposes sharp fluctuations in power consumption, particularly during AI model training or inference workloads, when GPU loads surge and plunge faster than conventional grids can adapt. These swings have reached levels that triggered natural gas turbine shutdowns under severe stress, while relying on massive battery storage systems imposes heavy financial burdens to balance consumption and secure operational stability.

TerraPower bypasses this bottleneck with its 345-megawatt molten-salt-cooled reactor, taking advantage of energy storage capabilities originally developed to integrate with intermittent renewable sources such as wind and solar. Rather than throttling or accelerating the nuclear reaction to track demand fluctuations, the reactor continues splitting atoms at a steady rate, directing surplus heat into a massive molten sodium reservoir. When data centre power draw spikes suddenly, the facility draws stored heat from the tank to generate the steam needed to drive turbines and ramp generation instantly without compromising reactor stability.

This engineering approach allows the company to amortise its capital investment over longer operating hours, by running expensive components at full capacity even during periods of lower instantaneous demand. By combining the high capacity factor of nuclear power with the flexibility of thermal energy storage, the company secures an operational advantage that enables it to integrate with renewable-heavy power grids or link directly to demanding AI data centres.