The Challenge of Nuclear Flexibility

Nuclear power plants are designed to operate at a constant, high capacity—typically maintaining a 92.5% capacity factor in the U.S. This is economically necessary due to high capital expenditures, but it creates a technical mismatch for AI data centers. AI workloads cause power demand to fluctuate rapidly as GPU clusters scale tasks up and down. Traditional reactors are notoriously slow to adjust output, often limited to a 5% change per minute. Even newer small modular reactors (SMRs) are limited to roughly 10% per minute. Forcing a reactor to ramp up and down to match data center load is inefficient and undermines the economic viability of the plant.

Decoupling Generation from Demand

TerraPower’s 345-megawatt molten salt-cooled reactor solves this by separating the heat generation process from the electricity generation process. Instead of modulating the nuclear reaction to match power demand, the reactor runs at a steady state. The excess heat produced during low-demand periods is stored in a large reservoir of molten sodium. When the data center experiences a spike in demand, the plant draws from this thermal storage to generate additional steam, driving the turbines to meet the load.

This approach allows the plant to maintain a high capacity factor—maximizing the return on its high capital investment—while providing the flexible, responsive power required by modern AI infrastructure. By using thermal storage as a buffer, TerraPower avoids the need for massive, expensive battery banks typically required to smooth out volatile data center power loads.