TerraPower is building a nuclear reactor that could power AI data centers.

TerraPower unveiled its first contract to power AI data centers, to be commissioned in 2027. It’s the second Natrium reactor after the first one is being built in Wyoming. What makes it stand out is its thermal battery: the 345 MW reactor operates at full capacity all the time, sending excess heat into a giant pool of molten sodium. In times of high demand from AI data centers, this stored heat can be quickly turned back into extra steam and electricity.

1Why GPU Clusters Break Gas Turbines

AI cluster demand fluctuations are different than normal consumer demands. They can ramp up or down in just a few seconds. It’s physically impossible for gas turbines, which usually take care of these fluctuations, to follow suit, as AI data centers have noted. A typical nuclear reactor increases power by about 5% each minute while new SMRs can go up to 10%. These rates are too slow for GPU clusters. With a thermal battery, the TerraPower reactor can operate at full capacity all the time while the sodium absorbs any sudden fluctuations.

2Meta Has Already Ordered Eight Units

In January 2026, Meta signed an agreement for 8 Natrium units. While TerraPower hasn’t specified who would own the new AI data center, speculation points to the same company. The price of each unit and the overall capacity is unknown.

Now nuclear power can compete with gas on paper, not just in theory, to fuel AI workloads. Nuclear power offers the needed flexibility when paired with a thermal battery like molten sodium. That way, it can supply the fluctuating needs of GPU clusters during training and inference of big models. If this project succeeds in meeting its claims starting in 2027, it will change the economics of building AI capacity for the next decade.