TerraPower’s reactor fuels AI data centers with a secret weapon
▼ Summary
– TerraPower plans to announce its first data center project this year, with construction expected to begin in 2027; the customer is undisclosed, though Meta agreed in January to buy eight of its Natrium plants.
– Nuclear reactors have the highest capacity factor (92.5% in the U.S.) but are slow to ramp, adjusting only about 5% of rated output per minute, while small modular reactors can react faster at 10%.
– Running nuclear plants at reduced capacity is financially challenging due to high capital expenditures, and mass manufacturing cost reductions for SMRs remain unproven and could take over a decade.
– Data center power loads swing rapidly due to AI training and prompts, stressing natural gas turbines and requiring large battery banks, which add costs.
– TerraPower’s molten salt-cooled reactor stores excess heat in molten sodium, allowing it to tap the reservoir to boost steam generation during demand spikes while keeping the reactor running at full capacity, combining high capacity factor with flexible energy storage.
Nuclear startups are increasingly positioning themselves as the answer to one of the tech industry’s most pressing headaches: the relentless demand for always-on electricity. TerraPower, the company co-founded by Bill Gates, is the latest to enter that race, with Bloomberg reporting that it will unveil its first data center project later this year.
The identity of the customer remains under wraps, though the company did announce in January that Meta has agreed to purchase eight of its Natrium power plants. The upcoming data center venture, slated to break ground in 2027, will mark TerraPower’s second plant, following the one already under construction in Wyoming.
Not every reactor design is built for the unique demands of a data center, but TerraPower holds a significant advantage: thermal energy storage. That edge, interestingly, was born out of a focus on renewables like wind and solar, not AI.
Nuclear reactors perform at their best when running at full capacity. They boast the highest capacity factor of any power source in the U. S., generating at maximum output about 92.5% of the time. But they also have a weakness: they are slow to adjust. Existing reactors can only ramp up or down by roughly 5% of their rated output per minute, according to the National Laboratory of the Rockies (NLR).
The newer small modular reactors (SMRs) that many startups are chasing can respond quicker, around 10% per minute, per NLR. But throttling down is not a winning business strategy. Reduced output means reduced revenue, and that is a problem for a technology with the highest upfront capital costs in the energy sector. While mass manufacturing of SMRs is expected to lower those costs eventually, that promise remains unproven, and even if it pans out, the benefits may take a decade or more to materialize. Every company in this space concedes that its first plants will be pricey, making it essential to run them hard and often.
For data centers, especially those relying on behind-the-meter power, the challenge is acute. Loads can spike or plummet in seconds as GPUs churn through AI training or respond to user prompts. These swings are so violent that even natural gas turbines have been known to fail under the strain. Smoothing that curve typically requires massive battery banks, which only adds to the cost.
TerraPower’s design, a 345-megawatt reactor cooled by molten salt, was built to sidestep these issues. The original goal was to create a plant that could complement the intermittency of wind and solar, ramping up and down with ease. Data centers present a similar pattern, albeit on the demand side rather than the supply side.
The trick is that TerraPower doesn’t throttle its reactor. Instead, it keeps splitting atoms at a steady clip, and the excess heat is diverted into a large tank of molten sodium. When power demand surges, the plant taps that reservoir to generate extra steam and spin the turbines faster. The expensive equipment never sits idle, even when demand is low, which means the company can spread its capital costs over far more operating hours.
This strategy combines the best of nuclear power, a high capacity factor, with a storage solution that makes the plant flexible enough for a renewable-heavy grid or a hyperscale data center. It is a nimble approach that could give TerraPower a real edge in the race to power the AI boom.
(Source: TechCrunch)
