TerraPower, founded by Bill Gates, plans to announce its first project to supply a data center with power this year, according to Bloomberg. The company has not disclosed the customer’s identity, but the project, whose construction is expected to begin in 2027, will be its second power plant, while its first plant is already under construction in Wyoming.
The move comes as AI data centers seek continuously available power sources, since training and running models require large amounts of energy. Nuclear startups are betting that their ability to generate electricity around the clock can address part of this problem, but operating nuclear reactors alongside fluctuating data center loads presents a technical and economic challenge.
Why Is It Difficult for Reactors to Keep Up with Data Centers?
Nuclear reactors operate most efficiently when they continue running at a high and steady level. According to the National Laboratory of the Rockies, nuclear reactors in the United States produce their maximum output 92.5% of the time, but they generally cannot increase or decrease their output by more than about 5% of their nameplate capacity per minute.
Small modular reactors, which many startups are developing, can respond more quickly, at a rate of roughly 10% of their nameplate capacity per minute. Even so, operating any reactor at low power reduces the amount of electricity sold, which is particularly costly in nuclear power because of high capital expenditures. Even if the industry succeeds in reducing the manufacturing cost of small modular reactors, it could take a decade or more for those benefits to emerge.
Data centers, particularly those relying on behind-the-meter power, need to handle rapid changes in electrical load. Demand can drop or rise quickly depending on GPU workloads, from training AI models to responding to prompts. Facilities typically turn to large batteries to smooth out these fluctuations, adding a new cost, while gas turbines have faced operational pressure because of these changes.
How Does TerraPower’s Design Work?
TerraPower designed a 345-megawatt molten-salt-cooled reactor to address this problem. Instead of changing the nuclear reactor’s output when demand changes, the reactor continues splitting atoms and producing heat, then stores the excess heat in a large tank of molten sodium.
When electricity demand rises, the plant can use the stored heat to produce more steam and drive the turbines. This keeps the high-cost equipment operating even when demand is low, while the heat inventory is used to provide additional capacity when needed.
What Changes in Practice?
This approach combines the high utilization rate characteristic of nuclear power with storage technology that allows the plant to adapt to changing loads. This could be useful when connecting the reactor directly to a data center or to a grid that relies more heavily on intermittent sources such as wind and solar power.
The design does not mean TerraPower has secured the data center power market; the prospective customer has not been announced, and the project is still in the pre-construction phase. But heat storage could give the company an advantage in competing for data center deals because it addresses the gap between the reactor’s steady output and fluctuating AI electricity demand. TerraPower announced in January that Meta had agreed to purchase eight Natrium plants, without confirming that this deal is the anticipated data center project.