TerraPower, the nuclear startup founded by Bill Gates, plans to announce its first data center project this year, according to TechCrunch AI, which cites Bloomberg reporting. The company hasn’t named the customer yet. But the project is expected to break ground in 2027 and would be TerraPower’s second power plant, after the one already under construction in Wyoming.
What stands out here isn’t just another nuclear-for-AI pitch. It’s the reason TerraPower thinks it can win: energy storage. And oddly enough, the design comes from thinking about wind and solar, not GPUs.
The problem nuclear has with data centers
Nuclear plants are built to run flat out. In the U.S., reactors generate at maximum power 92.5% of the time, the highest capacity factor of any power source, TechCrunch AI reports. That’s by design. Existing reactors are slow to change speed, moving only about 5% of their rated output per minute. The newer small modular reactors (SMRs) that most startups are chasing can shift roughly 10% per minute.
Here’s why that matters. Data center loads don’t sit still. When they’re training models or answering prompts, demand can crash and spike fast as GPUs fire up and wind down. Those swings are brutal. TechCrunch AI notes that natural gas turbines have been breaking under the stress.
Running a reactor at half capacity to match those dips is a money-loser. Nuclear carries the highest capital costs of any generating technology, so every hour it’s not producing electricity hurts. Startups are betting that mass-producing SMRs drives costs down, but that’s unproven, and even if it works, the payoff could be a decade out.
TerraPower’s workaround: a giant vat of molten salt
TerraPower’s 345-megawatt reactor is molten salt-cooled, and it handles demand swings in a clever way. Instead of ramping the reactor up and down, it keeps splitting atoms at a steady pace. The extra heat gets stored in a huge reservoir of molten sodium.
When demand spikes, the plant pulls from that stored heat to make more steam and spin the turbines harder. When demand drops, the reactor keeps running and banking heat.
The result:
- The reactor stays at peak output, protecting that high capacity factor.
- The turbines flex to match sudden load swings without gas or massive battery banks.
- The expensive gear keeps earning even when demand is low, spreading the investment across more operating hours.
TerraPower actually designed this for a renewable-heavy grid, where wind and solar come and go. Turns out data centers are intermittent too, just on the demand side instead of the supply side. Same problem, mirror image. The storage tank solves both.
Why this matters
The AI power crunch is real, and nuclear startups have positioned themselves as the always-on answer. But most of them are still fighting the same fundamental tension: reactors want to run steady, data centers don’t behave steady. TerraPower’s molten salt storage sidesteps that fight instead of engineering around it with batteries that pile on cost.
There’s also a notable name in the background. Back in January, TerraPower said Meta agreed to buy eight of its Natrium plants, per TechCrunch AI. The company won’t say whether Meta is the customer for this new project, so read into that what you want.
A few things worth watching:
- Who the customer turns out to be, and whether it’s a hyperscaler like Meta.
- Whether the 2027 groundbreaking holds, since nuclear timelines slip often.
- Whether TerraPower’s storage edge holds up once SMR competitors respond.
The bigger picture: the winner in AI power might not be whoever builds the cheapest reactor. It could be whoever solves the flexibility problem. TerraPower is betting a tank of hot salt is the answer. For the full breakdown, check the original report at TechCrunch AI.