TerraPower's molten salt secret weapon for AI data centers

Bill Gates' nuclear startup is quietly building the reactor designed to ride out the chaos of AI power demand

Published: 2026-08-22
Category: Quick Take
Sources: TechCrunch

The pitch

Nuclear power startups have spent the last year pitching themselves as the antidote to what ails AI data centers: power that is always available. Bill Gates-founded TerraPower is one of the latest to throw its hat into that ring, with Bloomberg reporting the startup plans to announce its first data center project this year. The customer is undisclosed, though in January Meta agreed to buy eight of its Natrium power plants. The data center project, expected to break ground in 2027, would be TerraPower's second power plant — its first is already under construction in Wyoming.

The engineering edge

Not every nuclear reactor suits data center duty, but TerraPower has one key advantage competitors lack: energy storage. And it's all thanks to renewable power sources like wind and solar. Nuclear reactors work best running at full tilt — they have the highest capacity factor of any generating technology, producing at maximum power 92.5% of the time in the U.S. But they're slow to ramp: existing reactors can only increase or decrease output about 5% of rated capacity per minute. New small modular reactors (SMRs) are faster, at about 10%, but running at reduced capacity makes it hard to make money — especially nuclear, which carries the highest capital expenditure of any generating tech.

Why data centers break the model

For data centers, especially those relying on behind-the-meter power, that's a real problem. Their loads can sink and soar quickly as GPUs respond to training and inference tasks. The swings are so demanding that natural gas turbines have been breaking under the stress; smoothing the curve requires large battery banks, adding cost. TerraPower designed its 345-megawatt molten salt-cooled reactor to work around this. To ramp quickly, it doesn't throttle the reactor — it keeps splitting atoms, and the extra heat gets stored in a giant vat of molten salt. When power demand spikes, the plant taps that reservoir to generate more steam and spin the turbines faster. The expensive equipment keeps working even when demand is low, amortizing the investment over more operational hours.

The take

This is a genuinely clever architecture, and it reframes the nuclear-for-AI conversation. Most SMR pitches lean on cheap manufacturing; TerraPower is betting that operational flexibility — being the reactor that can smooth out a GPU farm's jagged load curve without battery banks — is the real differentiator. Whether the molten salt storage economics hold up at scale is the open question, but the design rationale is sound and the Meta order gives it a credible anchor customer.

Source: TechCrunch, "TerraPower's nuclear reactor has a secret weapon for powering AI data centers," Tim De Chant, August 19, 2026.