TerraPower’s Data Center Plan Turns Nuclear Heat Into a New AI Power Play

Updated August 19, 2026 6:54 pm

TerraPower is preparing to announce its first data center power project this year, a move that could position Bill Gates-founded nuclear startup as a serious contender in the race to supply AI infrastructure with around-the-clock electricity. The company’s edge is a reactor design that can pair nuclear generation with built-in energy storage, giving it more flexibility than most atomic plants.

The project, reported by Bloomberg and expected to begin construction in 2027, would be TerraPower’s second power plant. Its first reactor is already under construction in Wyoming, and the new deal would extend the company’s ambitions beyond the traditional utility market into the high-demand world of AI compute.

That matters because data centers are becoming one of the most power-hungry industries on the planet. As companies race to train larger models and serve increasingly heavy AI workloads, they need electricity that is not just cheap, but consistently available and capable of handling steep swings in demand. TerraPower’s design aims to meet that need in a way most reactors cannot.

Why TerraPower is entering the data center race now

TerraPower’s move reflects a broader shift in the energy market: AI companies are scrambling for reliable power, and nuclear startups are trying to prove they can deliver it at scale. The pitch is simple but urgent. Data centers cannot wait years for grid upgrades, and they cannot rely on intermittent sources alone when GPUs are consuming massive loads day and night.

Nuclear power offers a compelling answer because it can run continuously at high output. But traditional reactors are not particularly good at following rapid changes in demand, which is a problem for data centers whose electricity use can swing sharply depending on workload.

By targeting the data center market, TerraPower is aiming at one of the few customer classes willing to pay a premium for guaranteed, behind-the-meter generation. If the company can make that model work, it could create a new revenue stream for advanced nuclear projects that have historically depended on utilities and government support.

What makes AI workloads such a difficult power problem?

AI workloads are hard to serve because they can ramp up and down very quickly. Training runs, inference spikes, and large batch jobs all create unpredictable demand patterns, especially when multiple customers share the same facility. That volatility can strain generation equipment and force operators to add battery storage or other buffering systems.

In some cases, even gas turbines have struggled under the stress of these rapid fluctuations. The result is a power challenge that is not solved by generation alone. It requires a system that can absorb sudden changes without sacrificing efficiency or reliability.

• AI clusters often draw very large loads in short bursts.
• Demand can shift dramatically from one minute to the next.
• Data center operators need stable electricity to protect uptime.
• Extra smoothing infrastructure can add major cost.

How TerraPower’s reactor design is different

TerraPower’s advantage is not that its reactor can dramatically speed up and slow down like a battery. Instead, it keeps producing heat at a steady rate and stores some of that thermal energy in molten sodium. When the system needs more power, it taps that stored heat to generate additional steam and spin the turbines harder.

This design allows the plant to keep its expensive nuclear equipment operating efficiently while still responding to changing electricity needs. That makes it useful in settings where demand is not perfectly flat, including grids with a high share of wind and solar, and potentially data centers with variable loads.

The company’s Natrium system is a 345-megawatt molten salt-cooled reactor paired with energy storage. TerraPower originally built the concept to support renewable-heavy grids, where it could help balance intermittent sources by storing heat and dispatching extra power when needed.

That same flexibility now looks valuable for AI infrastructure. Instead of forcing the reactor to chase load in real time, TerraPower can keep the core doing what it does best while the storage layer handles short-term demand changes.

TerraPower’s core idea is to let the reactor stay productive while a thermal storage system absorbs the fluctuations that would otherwise make nuclear power awkward for fast-changing loads.

Why most nuclear plants are not ideal for data centers

Most nuclear stations are built for steady baseload operation, not rapid ramping. In the United States, they already operate at exceptionally high capacity factors, meaning they spend most of their time generating close to maximum output. That strength, however, comes with a limitation: traditional reactors are not built to move up and down quickly.

According to national laboratory data cited in the source material, existing reactors can generally adjust output by only about 5% of rated capacity per minute. New small modular reactor concepts are expected to improve on that, reaching roughly 10% per minute, but even that can be too slow or too inefficient for a highly dynamic data center load profile.

For nuclear startups, operating below full capacity can be economically painful. Nuclear projects carry some of the highest upfront costs in the power sector, and every hour a reactor is not producing sells fewer kilowatt-hours to recover that investment. That is one reason advanced reactor developers continue to emphasize scale and high utilization.

What TerraPower gains from storage

Storage is the secret ingredient that makes TerraPower’s reactor more suitable for AI infrastructure. By decoupling heat production from electricity delivery, the company can keep the reactor running efficiently even when immediate power demand falls. That reduces the economic penalty of serving a load that moves around.

In practical terms, the storage tank acts as a buffer. When demand is low, the reactor keeps heating the sodium. When demand rises, the plant can draw on that reservoir to produce more steam and electricity without having to rev the reactor itself up and down.

This approach gives the plant a better chance of fitting into a data center partnership, where the customer wants predictable power but may not have perfectly steady consumption. It also strengthens TerraPower’s case in markets with substantial wind and solar generation, where the grid itself experiences frequent variation.

How does this compare with batteries?

It compares favorably in certain roles because thermal storage is built into the plant rather than bolted on afterward. Batteries can still play an important part in smoothing loads and providing short-duration backup, but they add expense and are not always the most efficient answer for large, sustained energy shifts.

TerraPower’s system is designed to use nuclear heat as the primary energy source and storage as the balancing layer. That differs from a battery-heavy model, in which electricity is stored and later discharged, often at significant cost for very large installations.

• Thermal storage is integrated into the power plant design.
• It lets the reactor keep running at a steady pace.
• It can reduce the need for oversized battery banks.
• It may improve economics for long-duration flexibility.

How significant is the reported data center project?

The project is significant because it could be TerraPower’s first step into a market that is expanding much faster than the traditional electricity sector. AI infrastructure is forcing new conversations about siting, generation, transmission, and reliability, and nuclear firms want to be part of that solution set.

According to the reporting, TerraPower has not identified the customer publicly. That leaves open questions about whether the plant will serve a hyperscaler directly, a colocation operator, or another industrial partner with major compute needs.

Still, the timing is notable. TerraPower announced in January that Meta had agreed to buy eight Natrium plants, underscoring the growing willingness of major technology firms to look at nuclear as a long-term energy strategy. A data center-specific project would move that relationship from procurement interest into physical deployment.

The company’s January deal with Meta signaled that large technology buyers are willing to back advanced nuclear capacity, but a dedicated data center project would show how that appetite translates into actual infrastructure.

What TerraPower still has to prove

TerraPower still faces the same challenges confronting much of the advanced nuclear sector: construction risk, cost control, regulatory hurdles, and a long timeline to commercialization. Its first plant, now under construction in Wyoming, remains the key proving ground.

That matters because every planned data center partnership depends on confidence that the technology can be delivered on time and at the expected cost. Nuclear startups often speak about serial manufacturing and lower costs from standardization, but those benefits typically arrive only after the first units are built and operated successfully.

For TerraPower, the question is not only whether the reactor can work technically, but whether it can be packaged into a business model attractive to data center customers. Those buyers want certainty, and they may not be inclined to wait through a long construction cycle unless the power deal is compelling enough to justify it.

Who benefits if the model works?

If TerraPower’s approach succeeds, the winners would include AI operators, utilities, and potentially renewable developers seeking firm backup power. Data center customers would gain a stable source of electricity. Grid operators could benefit from a more flexible zero-carbon resource. TerraPower, in turn, could tap a high-value market with steady demand.

There is also a broader climate and industrial-policy angle. A nuclear plant that can better match variable load could strengthen the case for advanced reactors as a practical tool in decarbonizing both the grid and energy-intensive digital infrastructure.

1. AI companies get steadier, lower-risk power access.
2. Nuclear developers gain a new commercial use case.
3. Renewable-heavy grids get a more flexible firm-power resource.
4. Investors may see a clearer path to monetizing advanced reactors.

The bigger picture for AI and nuclear power

TerraPower’s reported move is part of a broader convergence between the AI boom and the nuclear revival. As model training and inference expand, energy demand is becoming one of the defining constraints on the next phase of AI growth. At the same time, advanced nuclear developers are searching for anchor customers and workable deployment models.

That overlap is creating a new market narrative: the power source that once seemed slow, expensive, and politically difficult is being recast as infrastructure for the digital economy. Whether that narrative becomes a lasting business reality will depend on execution, financing, and the ability of reactor designs to do more than simply generate electricity.

TerraPower’s storage-enabled reactor may be one of the clearest attempts yet to answer that challenge. Instead of asking nuclear to behave like a battery, it uses storage to make nuclear more useful on a changing grid and more compatible with the erratic energy demands of modern data centers.

If the company can translate that technical logic into a bankable project, it could help define a new category of power plant: one built not just for the grid of the past, but for the AI economy of the future.

Key dates and milestones

The following timeline summarizes the developments that frame TerraPower’s latest move into the data center market.

In an industry obsessed with computing speed, TerraPower is betting that the winning power system may be the one that can move slowly, store heat, and still respond fast enough when the machines surge. That is a distinctly nuclear answer to an AI-era problem.