TerraPower's nuclear reactor has a hidden edge for powering AI data centers

Nuclear power startups have long pitched themselves as a fix for AI data centers' need for constant, reliable electricity. TerraPower, founded by Bill Gates, is among the latest, with Bloomberg reporting the company plans to announce its first data center project this year. The customer hasn't been disclosed, though TerraPower said in January that Meta had agreed to purchase eight of its Natrium power plants.

The new project, expected to break ground in 2027, would be TerraPower's second power plant; its first is already under construction in Wyoming.

The storage advantage

Nuclear reactors typically run best at full output — in the U.S., they generate at maximum capacity 92.5% of the time. However, traditional reactors adjust output slowly, and even newer small modular reactors (SMRs) being developed by various startups can only ramp somewhat faster, still not fast enough to match the rapid swings in power demand from AI workloads as GPUs process tasks. These swings have proven so severe they've caused natural gas turbines to fail, pushing operators toward costly battery banks to smooth demand.

TerraPower designed its 345-megawatt, molten salt-cooled reactor with a different goal in mind: complementing intermittent renewable sources like wind and solar. Instead of adjusting the reactor's power output to match demand, the plant keeps splitting atoms at a steady rate and stores the excess heat in a large tank of molten sodium. When demand spikes, that stored heat generates additional steam to drive the turbines.

This design lets the plant's expensive equipment keep operating even during low-demand periods, spreading the investment cost over more operating hours. By combining nuclear power's high capacity factor with flexible energy storage, TerraPower aims to serve both renewable-heavy grids and data centers — a combination that could give it an edge as companies race to secure power for AI.