Why One Nuclear Startup's Energy Storage Play Could Solve AI's Power Problem
TerraPower's molten-salt design keeps reactors running at full capacity while buffering wild GPU load swings - a rare combo that could reshape data center economics.
The Load Problem No One Saw Coming
Training runs that swing from idle to peak GPU utilization in seconds. Inference clusters that spike when prompts flood in, then go quiet. The loads AI data centers impose on their power supply are so abrupt that natural gas peaker turbines - historically the go-to technology for fast response - have started to fail mechanically under the stress. At DailyTechWire, we've tracked the power crunch across hyperscale campuses from Northern Virginia to Singapore, and one pattern is clear: the mismatch between how reactors want to operate and how AI compute actually behaves has become the bottleneck no one budgeted for.
TerraPower, the nuclear venture Bill Gates launched more than a decade ago, thinks it has an answer. The company is preparing to announce its first data center customer before year-end, with construction expected to begin in 2027. Meta has already committed to purchasing eight of TerraPower's Natrium reactors, and the firm's first commercial plant is under construction in Wyoming. What sets TerraPower apart is not the reactor itself but the heat battery bolted onto it - a massive tank of molten sodium that decouples electricity generation from fission output.
Full Power, All the Time
Nuclear economics hinge on capacity factor. U.S. reactors run at maximum output 92.5 percent of the time, the highest of any generation technology. That performance is partly necessity: legacy light-water designs can adjust output by only about five percent per minute, according to research from the National Laboratory of the Rockies. Newer small modular reactor designs can ramp twice as fast, hitting ten percent per minute, but even that leaves operators in a bind. Running below rated capacity means spreading capital costs - already the highest in the power sector - over fewer kilowatt-hours. Early SMRs will be expensive; every startup building them acknowledges this. So maximizing uptime is not optional.
Data centers, especially those purchasing behind-the-meter power, present the opposite challenge. GPU clusters training foundation models or serving inference requests can see load swings that would flatten a reactor's economics if the plant tried to follow them directly. Battery arrays can smooth the curve, but they add cost and complexity to a project already stretched thin on capital.
Heat as a Buffer
TerraPower designed its 345-megawatt Natrium reactor with a different problem in mind: complementing wind and solar on grids where renewable penetration is climbing. The design team needed a reactor that could handle intermittency without throttling the fission process itself. The solution was a thermal energy reservoir - a large insulated tank filled with molten sodium. When electricity demand is low, excess heat from the reactor core flows into the tank. When demand spikes, the stored heat generates steam to drive the turbines harder, boosting output without touching reactor power levels.
The result is a power plant that can hold its nuclear island at steady state while output to the grid varies by as much as 500 megawatts peak, well above the reactor's nameplate rating. The approach preserves capacity factor where it matters - inside the reactor - while delivering the load-following capability that both renewable grids and AI data centers require. It is not a battery in the electrochemical sense, but it performs the same economic function: it turns a capital-intensive, inflexible asset into one that can respond to volatile demand.
Why This Matters for Asia's Build-Out
The architecture has implications beyond the U.S. market. Across Seoul, Singapore, and emerging hyperscale corridors in India and Indonesia, data center developers face twin constraints: limited grid capacity and regulatory pressure to decarbonize. Pairing nuclear with on-site thermal storage offers a path to firm, zero-carbon power that does not require overbuild of transmission infrastructure or reliance on diesel gensets for backup.
China has already moved aggressively into SMR deployment for industrial heat applications, and state-owned utilities are exploring co-location models with compute clusters in Inner Mongolia and Gansu. If TerraPower's approach proves out commercially, expect regional players to license or replicate the thermal storage concept. The design is not proprietary in the way reactor physics are; molten-salt thermal systems have been demonstrated at scale in concentrated solar plants. What TerraPower has done is integrate that storage into a nuclear island in a way that aligns reactor economics with data center load profiles.
The Capital Question Remains Unresolved
None of this erases the core challenge: early nuclear plants will be expensive, and mass manufacturing benefits remain hypothetical. TerraPower has not disclosed the capital cost per megawatt for its Natrium units, but industry observers expect first-of-a-kind SMRs to run two to three times the cost of combined-cycle gas turbines on a dollar-per-kilowatt basis. The thermal storage system adds equipment and complexity, though it may reduce the need for external battery installations.
The bet is that running at high capacity factor while serving volatile loads will generate enough revenue to service the debt. If interest rates stay elevated and construction timelines slip - both common in nuclear projects - the math gets harder. Meta's eight-reactor commitment provides some demand visibility, but the broader data center market has shown willingness to pivot quickly when economics shift. Natural gas remains cheap in much of North America, and renewable-plus-storage costs continue to fall.
What We Are Watching
TerraPower's second plant will be the real test. The Wyoming facility benefits from Department of Energy cost-sharing and a supportive state regulatory environment. A privately financed data center project will face different scrutiny from lenders and offtakers. If the company can demonstrate that its thermal storage system delivers the load-following performance it claims without compromising reactor uptime, it will have solved a problem that has stymied the nuclear industry for decades.
The race to power AI is still early, and the technology stack is far from settled. But in a field crowded with reactor designs optimized for baseload dispatch, TerraPower's decision to build flexibility into the heat side of the plant - rather than the nuclear side - offers a model that could scale across both data centers and renewable-heavy grids. Whether it can do so at a price the market will pay is the question that will define the next five years of nuclear deployment.
