TerraPower's Natrium Reactor Has a Secret Weapon for AI Data Centers: A Giant Vat of Molten Salt

The race to power artificial intelligence has produced some strange bedfellows, but few are as consequential as the one now forming between hyperscale data center operators and nuclear startups. This week, that race gained a new front-runner candidate: Bill Gates-founded TerraPower is preparing to announce its first data center project this year, according to Bloomberg reporting highlighted by TechCrunch on August 19. The company did not name the customer — but the timing is hard to miss. In January, Meta agreed to a landmark deal supporting the development of up to eight Natrium power plants in the United States.

The data center project, expected to break ground in 2027, would become TerraPower’s second power plant. The first — a 345-megawatt Natrium unit near a retiring coal plant in Kemmerer, Wyoming — officially began construction in April 2026, marking America’s first utility-scale advanced nuclear plant to reach that milestone in decades.

Why not every reactor can serve a data center

At first glance, nuclear seems like the obvious answer to AI’s power problem. Reactors deliver carbon-free electricity around the clock, and U.S. nuclear plants boast the highest capacity factor of any generating technology — running at maximum output 92.5% of the time. For an industry whose GPU clusters gulp megawatts at all hours, that reliability reads like a match made in heaven.

The problem is the load itself. AI data centers do not draw steady power. Training runs ramp up and wind down; inference traffic spikes when millions of users hit a model at once; clusters of GPUs can swing demand violently in minutes. Those swings are so punishing that natural gas turbines — the conventional fast-ramping workhorse — have reportedly been breaking under the stress. Behind-the-meter installations typically paper over the gaps with massive battery banks, which only inflate already eye-watering capital costs.

Nuclear reactors are, ironically, terrible at this. Traditional plants can ramp only about 5% of their rated output per minute, according to figures cited from the National Laboratory of the Rockies. Newer small modular reactors (SMRs), the category most nuclear startups are chasing, improve that to roughly 10% per minute. But throttling a nuclear plant is economically perverse: nuclear carries the highest capital expenditure of any generating technology, so every hour spent running below full output is money burned. Early plants are the most expensive of all, and the mass-manufacturing cost curves startups promise remain unproven, possibly a decade away.

The molten salt trick

TerraPower’s Natrium design sidesteps the entire dilemma with an approach borrowed from renewable integration: decouple the reactor from the load using energy storage.

The Natrium plant pairs a 345-megawatt sodium-cooled fast reactor with an integrated molten salt energy storage system — a giant vat of salt that holds roughly a gigawatt-hour of heat. The reactor keeps splitting atoms at a steady clip no matter what the grid wants. When demand is low, excess heat flows into storage. When demand spikes, the plant taps that reservoir to boil extra steam and spin the turbines harder, boosting electrical output to as much as 500 megawatts for over five and a half hours.

In other words, the expensive parts — the reactor, the fuel, the licensing — never stop earning. The plant takes the best feature of nuclear power, its relentless capacity factor, and welds on the flexibility of a peaker plant. TerraPower originally designed the system to complement intermittent wind and solar on a renewable-heavy grid, but as TechCrunch notes, a data center’s spiky GPU load is essentially the same problem mirrored: volatility on the demand side instead of the supply side.

(One correction worth noting: an earlier version of the TechCrunch piece said molten sodium; the storage medium is molten salt. The sodium loop cools the reactor; the salt stores the heat.)

The Meta connection

The hyperscaler interest is no longer hypothetical. In January 2026, Meta announced a series of nuclear agreements totaling up to 6.6 gigawatts of firm power, part of what the company framed as “powering American AI leadership.” The TerraPower piece of that puzzle supports the development of two new Natrium units capable of generating up to 690 megawatts, within a framework allowing up to eight Natrium plants — roughly 2.8 gigawatts — in the United States.

For Meta, the calculus is straightforward. AI infrastructure plans announced across the industry imply tens of gigawatts of new demand this decade, and grid interconnection queues stretch for years. Nuclear — whether built behind the meter or contracted through utilities — offers firm, carbon-free capacity that hedging with gas turbines cannot match reputationally. TerraPower, for its part, gets something every advanced nuclear startup desperately needs: an anchor customer with a balance sheet capable of de-risking first-of-a-kind construction.

TerraPower is also expanding internationally. The Natrium design recently entered the United Kingdom’s Generic Design Assessment process, and the company has discussed deployments beyond U.S. borders.

What to watch

The 2027 groundbreaking for the data-center-linked plant is the milestone to circle. Advanced nuclear in the U.S. has a long history of cost overruns and cancellations, and TerraPower has already absorbed delays — its Wyoming timeline slipped partly due to fuel supply issues tied to Russia’s invasion of Ukraine, which forced a pivot to domestic high-assay low-enriched uranium (HALEU) fuel. Whether the second plant breaks ground on schedule will say a lot about whether the AI-nuclear alliance is real infrastructure strategy or another round of speculative announcements.

The economics question also remains open. Natrium’s storage advantage only pays off if data center operators actually value flexibility enough to price it in — and if TerraPower can deliver first-of-a-kind costs anywhere near its targets. SMR competitors like Oklo, X-energy, and Kairos are courting the same customers with different technical bets.

But the underlying logic is sound, and it marks a subtle shift in how the industry thinks about AI power. The question is no longer just “how many gigawatts can you build?” It is “how fast can you turn them up and down?” For a decade, nuclear’s inflexibility was a bug. TerraPower’s bet is that for AI-era grids, wrapped in molten salt storage, it can become a feature.