AI data centers need electricity that can stay available while demand moves sharply up and down. TerraPower, the Bill Gates-founded nuclear power startup, is entering that conversation with a reactor design built around a feature that could matter as much as the reactor itself: energy storage.
According to Bloomberg reporting cited by TechCrunch, TerraPower plans to announce its first data center project this year. The company has not named the customer. The project is expected to break ground in 2027 and would be TerraPower’s second power plant, while its first is already under construction in Wyoming.
Nuclear Power Meets AI Demand
Nuclear power startups have been presenting themselves as a fit for AI data centers because they can provide power that is always available. That matters for facilities running heavy computing workloads, especially when GPUs shift quickly between training AI systems and responding to prompts.
Those workload changes create a difficult power profile. Demand can sink and soar quickly, and the source article notes that the swings have been severe enough that natural gas turbines have been breaking under the stress. To reduce those swings, data centers often need large banks of batteries, which adds more cost.
TerraPower’s pitch is not simply that nuclear power can run steadily. Its more specific advantage is that its reactor design can keep producing heat while the plant’s output to the grid or a data center changes. That distinction is central to why the company’s design may be useful for AI infrastructure.
Why Traditional Nuclear Plants Are Hard To Throttle
Nuclear reactors generally work best when they run at full output. In the U.S., nuclear reactors generate at maximum power 92.5% of the time, giving them the highest capacity factor among power plant types referenced in the source article.
That high utilization is valuable, but it also creates a challenge. Existing reactors are slow to ramp up or down. According to the National Laboratory of the Rockies, they can increase or decrease output by only about 5% of their total rated output per minute.
New small modular reactors, or SMRs, can respond more quickly. The source article says many startups are pursuing SMRs, and that they can adjust at about 10% of their rated output per minute, per NRL. But even if a reactor can slow down, running below capacity is not ideal.
The business reason is straightforward. Power plants earn revenue by generating electricity. Nuclear has the highest capital expenditures of any generating technology mentioned in the source article, so early projects need as many productive operating hours as possible.
Startups hope that mass manufacturing of SMRs will bring capital expenditures down. But that has not yet been proven, and the source article says that even if it works, it could take a decade or more to see those benefits. In the meantime, early nuclear plants are expected to be expensive.
TerraPower’s Storage-Based Workaround
TerraPower designed its 345-megawatt molten salt-cooled reactor with flexibility in mind. The original target was not data centers. The design was intended to complement intermittent electricity from wind and solar, where power supply can rise and fall.
Data centers present a related but different problem. Instead of supply changing because the wind or sun varies, demand changes because computing workloads rise and fall. In both cases, the power plant benefits from a way to handle rapid changes without forcing the reactor itself to constantly adjust.
TerraPower’s answer is to keep the reactor running and store extra heat. Rather than increasing or decreasing the reactor’s power output, the plant continues splitting atoms. When electricity demand is lower, extra heat goes into a large reservoir of molten sodium.
When demand rises, the plant can draw from that stored heat to produce more steam and spin turbines. That lets the power plant respond to demand spikes while the expensive reactor equipment keeps operating. The approach helps spread the investment over more operational hours.
Why This Could Matter For Data Centers
The key point is that TerraPower’s design combines two things that data centers want but do not always get from one source:
- High utilization: the reactor can keep operating near its preferred steady state.
- Fast response: stored heat can help the plant meet sharp changes in electricity demand.
- Lower stress on external backup: smoother output could reduce reliance on large battery banks, based on the logic described in the source article.
This does not mean every nuclear reactor is well suited for data center duty. The source article is clear that not all reactors are a natural match. The important difference is whether the power plant can serve a volatile load without sacrificing the economics that nuclear plants depend on.
TerraPower’s Natrium plants are already part of the data center power discussion. In January, Meta agreed to buy eight of TerraPower’s Natrium power plants. The newly reported project is separate from that disclosure in the source article, and TerraPower has not identified the customer.
The Race To Power AI
AI growth is making electricity supply a strategic issue for data center operators. The problem is not only how much power they need, but how that demand behaves minute by minute. A steady power source can still struggle if the customer’s load is unstable.
TerraPower’s design tries to solve that mismatch by separating reactor operation from immediate electricity output. The reactor can keep doing what nuclear reactors do best, while the molten sodium storage system gives the plant a way to respond when demand changes.
That makes TerraPower’s energy storage system more than a technical detail. It may be the feature that determines whether its nuclear plants can fit the operating rhythm of AI data centers, renewable-heavy grids, or both.