TerraPower will announce its first AI data center project this year, with groundbreaking expected in 2027, according to Bloomberg. The Bill Gates-founded reactor startup is pitching a 345-megawatt molten salt-cooled design whose thermal storage system is engineered to match the erratic load swings of modern GPU clusters. It would be TerraPower's second power plant, following one already under construction in Wyoming.
The customer has not been named. In January, TerraPower announced that Meta had agreed to buy eight of its Natrium power plants, making Meta the largest disclosed off-taker in the company's pipeline. Whether this year's project is tied to that Meta agreement or a new counterparty is unclear.
The pitch rests on a specific technical claim about ramp rates. Existing US nuclear reactors run at a 92.5% capacity factor, the highest of any generation technology, but they are slow to modulate output — roughly 5% of rated capacity per minute, according to the National Laboratory of the Rockies. Small modular reactors under development by TerraPower's competitors ramp at about 10% per minute. Neither number keeps up with a GPU training cluster that can slam load from idle to peak in seconds.
Key facts
- 01TerraPower plans to announce its first AI data center project in 2026, with groundbreaking expected in 2027.
- 02Meta agreed in January to buy eight of TerraPower's Natrium power plants; the data center customer has not been named.
- 03The Natrium reactor is a 345-megawatt molten salt-cooled design that stores excess heat in molten sodium for on-demand ramping.
- 04US nuclear reactors run at 92.5% capacity factor but ramp only 5% of rated output per minute; SMRs ramp about 10% per minute.
- 05TerraPower's first plant is already under construction in Wyoming; the data center facility would be its second.
TerraPower's workaround is to stop ramping the reactor at all. The reactor keeps splitting atoms at full power around the clock, and excess heat is dumped into a large vat of molten sodium. When a data center's demand spikes, that stored heat gets pulled to generate additional steam, spinning the turbines harder without touching the fission rate. The original design goal was pairing nuclear with wind and solar, but the same mechanic maps onto AI training loads.
The economics matter as much as the physics. Nuclear has the highest upfront capital expenditure of any generation technology, and running a reactor below peak output is the fastest way to blow up its unit economics. As the source notes, the expensive equipment keeps working even when demand is low, allowing the company to amortize its investment over more operational hours — the case for thermal storage in one sentence.
Behind-the-meter power arrangements have become the AI industry's preferred way to sidestep grid interconnection queues that stretch years in most US regions. But behind-the-meter setups also strip away the grid's ability to absorb load swings. Natural gas turbines pressed into that role have been breaking under the mechanical stress of constant cycling, and battery banks large enough to smooth the curve add meaningful cost on top of already expensive gas peakers.
Meta's eight-plant agreement, along with hyperscaler deals struck by other reactor developers over the past 18 months, signals that AI operators are no longer treating nuclear as a decade-out option. The compute buildout is running ahead of grid capacity in Virginia, Texas, and Arizona, and the operators writing the checks for GPUs are the same operators writing power-purchase agreements with reactor startups that haven't yet built a commercial unit.
The counterweight is that TerraPower has not yet delivered a working Natrium plant. The Wyoming facility is under construction, not operating, and first-of-a-kind nuclear projects have a long history of schedule slippage and cost overruns. A 2027 groundbreaking on the data center project means first power is realistically late-decade at earliest, which is a long runway when AI capex cycles are measured in quarters. Small modular reactor economics also remain unproven; the industry is betting that mass manufacturing brings capex down, but that argument could take a decade to validate.
TerraPower's differentiation is real on paper: thermal storage genuinely does decouple reactor output from grid or data-center demand, and that is a harder problem to solve with an SMR that lacks a heat reservoir. If the design ships on schedule and on budget, TerraPower gets to sell the same asset to two different customer profiles — renewable-heavy utilities and hyperscaler campuses — without redesigning the plant. That optionality is worth something in a market where hyperscaler power demand forecasts keep getting revised upward and grid operators keep pushing interconnection dates further out.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
