The data center industry has an infrastructure problem that solar panels and battery storage cannot solve. Artificial intelligence workloads — training runs, inference at scale, continuous availability — require firm power: always-on, weather-independent, carbon-free. Fission, the process at the core of every commercial nuclear plant operating today, is the only proven technology that meets all three criteria simultaneously. In our view, the question for capital allocators is not whether fission matters to the AI buildout. It is which fission pathway delivers capacity in time to matter — and the answer points clearly toward restarting and upgrading existing plants, not waiting on small modular reactors.
The Load Is Real and the Gap Is Now
Global electricity generation to supply data centers is projected to grow from 460 TWh in 2024 to over 1,000 TWh in 2030 and 1,300 TWh in 2035. That is not a gradual curve — it is a near-vertical ramp. Deloitte estimated U.S. data center power capacity would rise 24% from 33 GW in 2024 to 41 GW in 2025, before tripling to 120 GW by 2030 and rising further to 176 GW by 2035. The grid is already showing strain under this demand. As of 2025, 2,600 GW of generation capacity — more than twice the entire installed capacity of the U.S. power plant fleet — was waiting for grid connection, with some data center projects facing delays of up to 12 years.
Renewables fill a critical part of this gap, but they do not answer the firmness problem. A frontier AI training run cannot be scheduled around cloud cover. Google's own senior energy leadership put it plainly: "We also need firm, dispatchable, carbon-free electricity technologies to cost-effectively decarbonize our electricity consumption." That is not a marketing statement. It is a procurement constraint, and fission is what satisfies it.
Major technology companies have reached the same conclusion and are acting on it: Microsoft, Google, Amazon, and Meta have all signed nuclear power agreements or invested directly in reactor development. The question is what form that nuclear power takes — and when it arrives.
Why SMRs Are Not the Near-Term Answer
Small modular reactors occupy most of the headlines in the nuclear-plus-AI conversation. The logic is intuitive: factory-built, modular units should be faster to deploy and easier to finance than decade-long megaprojects. SMRs offer factory fabrication, shorter construction timelines of three to five years versus seven to ten for large reactors, and scalable deployment from 50 MW to 300 MW per unit. But the commercial reality has not matched the theoretical appeal.
Total announced investment in SMR technology exceeded $30 billion between 2020 and 2025, yet only four SMR units were operating commercially worldwide as of early 2026 — all at a Russian floating plant and a Chinese demonstration reactor. No U.S. small modular reactor is in commercial operation. The first U.S. SMR commercial deployment is currently expected in 2028 — and that projection carries significant schedule risk given the sector's history. First-of-a-kind builds continue to face cost overruns, regulatory delays, and financing uncertainty that have historically plagued large nuclear projects.
The economics are equally sobering. Levelized cost of electricity estimates for first-of-a-kind SMRs range from $100 to $180 per MWh, significantly higher than existing nuclear at $30 to $60 per MWh. Analysts project SMR costs to decline only after 10 or more GW of cumulative deployment — a threshold that is nowhere in sight for U.S. projects. The gap between promise and reality does not disqualify SMRs as a long-term asset class; it simply means that any capital thesis anchored on SMR capacity arriving this decade should be tested against a skeptical timeline.
Restarts and Uprates: The Fastest Path to New Zero-Carbon Capacity
The most direct route to incremental nuclear generation in the United States is the fleet that already exists. Getting more output from the existing fleet is the most rapid and cost-effective method to add large-scale nuclear electricity generation capacity. The Nuclear Energy Institute's 2025 Future of Nuclear Power survey — conducted among 21 member companies operating 94 reactors — quantified the opportunity: uprates, restarts, longer fuel cycles, and other output increases together could add over 8 GWe of new carbon-free nuclear capacity over the coming decade.
The cost differential relative to new construction is substantial. Deloitte estimates that restarting retired plants can cost approximately $6.2 billion for three plants with 2 GW of capacity, versus $37 billion for the same capacity in new construction. That is a six-to-one capital efficiency advantage before accounting for the time value of getting capacity online years earlier.
Policy has aligned with economics. In May 2025, President Trump signed a series of executive orders targeting 400 GWe of U.S. nuclear capacity by 2050, with 5 GWe of power uprates and 10 large reactors under construction by 2030. The DOE's UPRISE initiative is the operational mechanism: UPRISE aims to add 2.5 GW of nuclear capacity by 2027, and 5 GW by 2029 — the equivalent of five new nuclear reactor years sooner than any other nuclear technology.
The transactions are already moving. The former Three Mile Island Unit 1 has a $1 billion DOE loan to support a roughly $1.6 billion restart of the 835-MW unit, targeted for return to service as soon as 2027 under a long-term power contract with a major technology company. A plant in Iowa is also advancing as a restart candidate, with a 25-year power purchase agreement announced in October 2025 to support a restart of the roughly 600-MW facility by 2029, in a project expected to cost more than $1.6 billion. A third restart project, an 800-MW plant in Michigan, is backed by a DOE loan guarantee of up to $1.52 billion, with the operator targeting power production in 2026 under the plant's renewed operating license.
None of these projects are without execution risk. The Michigan restart has already encountered schedule delays relative to its original target dates, a reminder that even the fastest nuclear pathway is not fast by conventional infrastructure standards. But the scale and capital commitment — federal loan guarantees, long-term offtake from investment-grade counterparties — represent the clearest current signal of where new zero-carbon firm power is actually arriving this decade.
The Fuel Supply Constraint No One Is Pricing Adequately
The strongest near-term check on the SMR buildout — and a genuine risk to the optimistic timelines — is fuel supply. Existing large reactors primarily use conventional low-enriched uranium (LEU), while many advanced reactor designs require high-assay low-enriched uranium (HALEU). In the U.S., only LEU is available on a commercial scale, with HALEU produced only in pilot quantities.
The demand trajectory for HALEU is steep. DOE projections cited by the American Nuclear Society indicate that advanced reactor deployment could require more than 40 metric tons of HALEU by 2030. In 2024, only about 900 kilograms of HALEU were produced domestically — a fraction of what the advanced reactor pipeline will require. Access to domestically sourced HALEU remains one of the central constraints facing the advanced nuclear sector at precisely the moment when demand is escalating most rapidly.
This constraint reinforces the restart-and-uprate thesis rather than undermining it. Existing plants run on conventional LEU, which is available at commercial scale today. The fuel supply bottleneck that threatens to slow advanced reactor timelines does not apply to the operating fleet or to plants being brought back online. That asymmetry matters for anyone evaluating which part of the nuclear buildout carries more near-term execution risk.
Fission This Decade, Fusion the Next
Fusion — joining atoms rather than splitting them — is the technology that eventually closes every energy equation. Private investment in fusion has accelerated sharply, with over $3 billion committed to the leading commercial programs globally.1 The physics case is not in doubt. The timeline is.
No fusion reactor has yet produced net electricity commercially. The gap between laboratory milestones and grid-connected commercial generation almost certainly extends past this decade. Fusion is the right bet for the 2030s and beyond; it is not the answer to a data center that needs firm power by 2028.
That leaves fission — proven, operating, politically supported at levels not seen since the 1970s, and structurally differentiated from intermittent renewables by its firm-power characteristics. In our view, the near-term opportunity within that thesis sits with the existing fleet: plants being restarted under DOE-backed loan guarantees with technology company offtake, and operating plants being uprated to squeeze additional output from infrastructure that has already cleared the hardest regulatory and construction hurdles.
SMRs are a real option class and will matter meaningfully once the first commercial U.S. units operate and demonstrate replicable cost curves. The investment thesis for SMRs is a 2030s thesis, not a 2020s one.
What would change this view: accelerated HALEU supply chain buildout that brings advanced reactor fuel costs in line with conventional LEU; demonstrated on-time, on-budget SMR completions that break the first-of-a-kind cost pattern; or a regulatory reform at the NRC that compresses licensing timelines sufficiently to make new large-reactor construction competitive with restart economics. Any of those developments would shift the calculus. Until then, the fastest path to fission-powered AI infrastructure runs through the plants that are already in the ground.
Footnotes
-
Fusion Industry Association, Global Fusion Data 2026 ↩