How TerraPower’s Natrium Reactor Could Help Power AI Data Centers

AI data centers need enormous amounts of electricity, but their demand is not always smooth. Training, fine-tuning, and inference workloads can create sharp, difficult-to-forecast changes in power use, including fluctuations across large GPU clusters within seconds. 18 That makes a power source valuable not only when it produces a lot of electricity, but also when it can respond to changing demand.

TerraPower’s Natrium design is aimed at that problem. It combines a 345-megawatt electric sodium-cooled fast reactor with a separate molten-salt thermal-storage system. The reactor supplies steady nuclear generation, while stored heat can temporarily increase the plant’s electrical output to as much as 500 MW. 1

The key correction: Natrium is not molten-salt-cooled

Natrium’s name is associated with its sodium-cooled reactor technology. In the reactor, liquid sodium carries heat from the nuclear core through primary and intermediate heat-transport systems. A separate integrated energy system transfers heat to molten salt for storage.

That distinction matters. The molten salt is being used as a thermal-storage medium, not as the reactor coolant. NRC materials describe the thermal-storage and steam-generation systems as operationally independent from reactor-power operations, even though they are integrated through the broader plant design.

The basic operating strategy is:

1. The reactor continues producing heat at a relatively steady rate.
2. When immediate electricity demand is lower, some of that heat charges the molten-salt storage system.
3. When demand rises, the stored heat helps produce additional steam and electricity.

TerraPower describes the design as a way to preserve the high utilization associated with nuclear generation while adding flexibility usually associated with storage or dispatchable generation.

What the storage system could deliver

The nuclear island is designed to provide 345 MWe. TerraPower says the storage system can raise total plant output to 500 MWe for more than five and a half hours when needed.

That higher figure should not be interpreted as the reactor’s continuous nuclear capacity. The additional output comes from stored thermal energy and is therefore limited by the amount of heat in the tanks and the plant’s ability to convert it into electricity. At the stated figures, the temporary peak could represent roughly 850 megawatt-hours of additional output above the reactor’s 345-MW baseline before the storage system needs to be recharged.

This could be useful in two ways:

• For a data center: stored heat could help cover a period of unusually high demand while the reactor continues operating steadily.
• For the grid: the plant could shift more electricity toward peak periods or support a system with substantial wind and solar generation.

The system is not an unlimited 500-MW supply, however. It is a finite-duration flexibility resource, not a replacement for every other layer of data-center power infrastructure.

Why this design fits the AI power problem

Global data-center electricity demand grew 17% in 2025, while electricity use by AI-focused data centers grew 50%, according to the International Energy Agency. 20 The IEA expects total data-center electricity consumption to double by 2030 and AI-focused consumption to triple over the same period. 21

At the same time, AI workloads can be highly variable. A large training run, inference demand, or cluster-level change can alter electricity requirements quickly and make them difficult to predict. 18

A conventional nuclear plant is generally most valuable when it runs at a high utilization rate. Frequently reducing reactor output to follow demand can leave expensive generating capacity underused. Natrium’s proposed answer is to keep the reactor producing heat and use thermal storage plus the steam system to shift when that heat becomes electricity. TerraPower says this can preserve reactor capacity factor while still allowing the plant to respond to grid conditions.

For a behind-the-meter data center, that could mean a stable base of firm generation with a defined amount of additional peak power. On a renewable-heavy grid, it could also help move nuclear heat into periods when variable renewable output is lower or demand is higher. TerraPower specifically presents Natrium as a technology designed to work with grids containing significant variable renewable generation.

The first project: Kemmerer Unit 1 in Wyoming

TerraPower’s first Natrium unit is Kemmerer Unit 1, near Kemmerer, Wyoming, at the site of a retiring coal facility. The project is intended to support PacifiCorp’s Rocky Mountain Power customers in Wyoming and the surrounding region.

The Nuclear Regulatory Commission approved the project’s construction permit in March 2026. TerraPower announced the formal start of construction on April 23, 2026. 14

That is a major licensing milestone, but it is not an operating license. The project still has to complete construction, fuel and commissioning work, and additional regulatory steps before fuel loading and commercial generation. A POWER report cited a planned construction-completion date of February 28, 2031, while other project reporting has described operation around 2030. Those dates should be treated as targets rather than an established delivery date for a first-of-a-kind plant. 613

The schedule is especially important for AI developers. Data centers are seeking power on shorter timelines than nuclear projects typically require, and the IEA has identified constraints involving grid connections, gas turbines, transformers, and other energy-technology supply chains. 21 Natrium could eventually provide firm power, but it cannot solve an immediate power shortage if its plant is not yet licensed and operating.

What is known about data-center customers?

TerraPower has announced several data-center-related relationships, but they do not all carry the same level of commitment.

Meta: up to eight potential plants

In January 2026, TerraPower and Meta announced an agreement to develop up to eight Natrium reactor-and-storage plants in the United States. TerraPower said the projects could provide Meta with up to 2.8 GW of carbon-free baseload energy, with storage capable of raising total output to 4 GW. The company described initial-unit delivery as early as 2032.

“Up to eight” is an important qualification. The announcement is an agreement to support development, not proof that eight plants have already been sited, permitted, financed, built, or connected to Meta data centers. The first units also have a later stated delivery window than the Wyoming demonstration target.

Sabey Data Centers: an exploratory relationship

TerraPower and Sabey Data Centers announced a memorandum of understanding to explore using Natrium plants for current and future data-center operations. An MOU is evidence of a potential commercial relationship, but it does not identify a completed project or guarantee a specific delivery date.

Reporting has also described a planned data-center-focused project whose customer had not yet been publicly identified. That uncertainty means Natrium’s role in powering a named AI data center remains a future possibility rather than an operating fact.

Natrium versus gas turbines and other reactors

Natural gas remains an important competitor because developers are actively considering it for new data-center capacity. Planned gas additions have grown as companies seek to meet AI-related load, and gas projects can benefit from comparatively favorable interconnection economics and completion rates. 24

Gas also has trade-offs: fuel costs and availability become ongoing risks, and generation produces direct carbon emissions. Nuclear offers a different value proposition—firm electricity without operational carbon emissions—but advanced nuclear projects must still prove that they can control first-of-a-kind construction costs, schedules, licensing requirements, fuel supply, and long-term reliability.

Natrium adds another layer of technical and commercial complexity because it combines a sodium fast reactor with a large high-temperature thermal-storage system. Its advantage is the possibility of combining nuclear utilization with dispatchable output; its risk is that neither the reactor nor the integrated storage system has yet demonstrated commercial performance at scale.

What Natrium can—and cannot—solve

Natrium’s strongest case is not that it behaves exactly like a battery or a gas turbine. It is that it could combine three attributes that AI infrastructure increasingly needs:

• steady, firm generation;
• low operational carbon emissions; and
• a limited but substantial reserve of dispatchable peak output.

That combination could be valuable for data centers with large, concentrated loads and for grids adding more variable renewable generation. But the storage duration is finite, the first plant is still under construction, and the customer pipeline includes agreements and exploratory arrangements rather than a fleet of operating reactors.

The central question is therefore not whether Natrium’s architecture is interesting. It is whether TerraPower can deliver the first plant on schedule, secure fuel and supply chains, demonstrate reliable operation, and offer competitive all-in electricity costs. Until that happens, Natrium is a promising power strategy for AI data centers—not yet a proven one.