A Gates-backed firm plans a second U.S. reactor to power an AI data center

The nuclear developer founded by Bill Gates is preparing to announce a project aimed squarely at the electricity appetite of artificial intelligence. The company says it intends to build a reactor dedicated to serving a data center, marking a shift toward a customer base that barely existed when the venture was formed.

TerraPower has signaled that it will reveal the deal before the end of the year and that construction on the unit could begin in 2027. Executives have not yet named the data-center customer, but the direction reflects how quickly power-hungry computing has become a driver of new energy investment across the United States.

A reactor sized for computing loads, not cities

TerraPower’s flagship design, called Natrium, is a small modular reactor rather than a traditional gigawatt-scale plant. It is engineered to deliver a steady 345 megawatts of baseload power, a fraction of the roughly one gigawatt produced by conventional reactors that were built to supply entire regions.

That smaller footprint is part of the pitch to technology firms, which increasingly want generation located near their facilities rather than delivered across congested transmission lines. A reactor scaled to a single large customer can be sited and financed differently than a plant meant to serve a broad utility grid. Building smaller units in a repeatable, factory-style design is also central to the small-modular strategy, since standardized components are meant to shorten construction timelines and hold down the per-unit cost that has historically plagued large one-off reactors.

The molten-salt battery that sets it apart

The feature TerraPower is highlighting for AI operators is an integrated molten-salt energy storage system that acts like a thermal battery. According to reporting on the design, the reactor can hold its steady output and then ramp up to 500 megawatts for more than five hours when demand spikes.

That flexibility matters because data centers do not draw power evenly. Training runs, cooling loads, and periods of peak usage can push consumption well above a baseline, and a plant that can surge for several hours offers a hedge against those swings without a separate generator or grid purchase. The storage layer also lets the reactor run at a constant, efficient output around the clock while releasing extra power only when it is needed, a design that decouples steady generation from uneven demand.

Storing heat rather than electricity is what gives the system its distinctive economics. Molten salt can hold thermal energy at high temperature for hours, and that reservoir can be tapped to boost turbine output during a spike, effectively turning a steady nuclear plant into one that behaves more like a dispatchable resource. For an operator whose load can jump quickly, that behavior is closer to what a natural-gas peaking plant provides, but without the associated fuel combustion.

A first plant already rising in Wyoming

The data-center reactor would be TerraPower’s second power plant. The company’s first Natrium project is already under construction in Wyoming, a build that has served as the proving ground for the design and its storage approach.

Placing a second unit behind a named AI or industrial customer would represent a different commercial model than the utility-anchored first plant. It would tie a reactor’s economics to a private buyer’s demand rather than to a regulated utility’s rate base, a structure that is still uncommon in the U.S. nuclear industry. Lessons learned building and licensing the Wyoming unit — from supply-chain arrangements to the regulatory path for the design — would carry directly into a follow-on plant, potentially reducing the uncertainty that surrounds any first-of-a-kind reactor.

Why AI is reshaping the reactor market

The surge of interest from technology companies has opened a market that small-reactor developers spent years trying to find. As reporting on the project notes, AI applications and industrial facilities such as semiconductor plants are now exploring on-site generation, creating fresh demand for developers of compact reactors.

Data-center operators have pursued a range of power sources in recent years, including large solar and wind purchases and long-term deals with existing nuclear plants. What distinguishes the newer wave is the willingness to underwrite construction of dedicated generation, shifting some builders from selling into a grid to building for a single anchor tenant.

The scale of computing expansion is the backdrop. Firms racing to build and operate ever-larger AI systems have found electricity supply, not just chips, to be a binding constraint, and that pressure is pulling nuclear developers into commercial conversations that would have been unlikely a few years ago. Nuclear power carries a particular appeal for these buyers because it produces carbon-free electricity around the clock, matching a data center’s need for constant, reliable power in a way that intermittent sources cannot on their own.

What remains unresolved

Several key details are still outstanding. The customer has not been disclosed, the exact site for the second plant has not been confirmed publicly, and a 2027 construction start leaves years of engineering, licensing, and financing work ahead before any power flows.

Even so, the announcement underscores a broader trend in the energy sector: the buildout of artificial intelligence is reordering where new generation gets built and who pays for it. A reactor designed around a data center’s load profile, complete with a storage system tuned to demand spikes, is a concrete example of how that reordering is beginning to take physical form.

This article was produced with the assistance of AI and reviewed by Morning Overview editors prior to publication.

More from Morning Overview