The computers behind the artificial-intelligence boom are hungry in a way the world’s power grids were never built to feed. Training and running the large models that now answer questions, write code and generate images consumes staggering amounts of electricity, and the data centers that house them have quietly become one of the fastest-growing loads on the planet. Their appetite has grown so large that the industry’s total draw now rivals the annual electricity use of an entire industrialized country — and the search for a steady, carbon-free way to power it keeps circling back to nuclear reactors.
An appetite the size of a nation
The scale is easiest to grasp through comparison. Analysts at the International Energy Agency have tracked data-center electricity demand climbing toward levels that match the consumption of a major European economy, with France — a country of roughly 68 million people — serving as a common yardstick. France uses on the order of 450 terawatt-hours of electricity in a year, and global data-center demand is closing on or surpassing that figure as AI workloads pile onto an already-growing base of cloud computing and streaming.
The trajectory is steeper still. The agency reported that data-center electricity use surged by about 17 percent in 2025 alone, with AI-focused facilities climbing even faster than the broader category. Its longer-range analysis projects that data-center consumption could roughly double by the end of the decade, and that the slice devoted specifically to AI could triple. That kind of growth turns a rounding error in national energy statistics into a strategic problem for utilities.
Why the grid is straining
The trouble is not only how much power these facilities want but how they want it. A hyperscale data center runs around the clock, demanding a constant, uninterrupted supply measured in hundreds of megawatts — the equivalent of a small city that never sleeps and never dims. Wind and solar, for all their momentum, deliver power that rises and falls with the weather, which makes them an awkward match for a load that cannot tolerate flicker or interruption.
That mismatch has pushed technology companies to hunt for firm, always-on generation that also happens to be low-carbon, so their expansion does not blow past their own climate pledges. Natural gas can fill the gap but carries emissions. Batteries help smooth supply but do not generate anything. The remaining option that checks both the always-on and the clean box is nuclear fission, and that is why the reactor has re-entered a conversation it had largely exited a generation ago.
Argonne’s case for the reactor
Researchers at Argonne National Laboratory, a U.S. Department of Energy research center, have argued that nuclear plants are well suited to the digital economy precisely because they produce large amounts of steady, carbon-free electricity that matches a data center’s flat, unrelenting demand curve. A reactor’s output does not sag at night or fade when clouds roll in, which is exactly the profile a server farm needs.
The laboratory’s analysis frames nuclear less as a silver bullet than as a natural partner for a load that behaves like heavy industry. Existing plants can supply verified clean power today, and a new generation of smaller designs promises to place generation closer to where the computing actually happens. That pairing — constant demand met by constant supply — is the core of the argument for wiring data centers to reactors.
The scramble for small reactors
The clearest sign that the idea has moved from theory to commerce is the rush of deals. According to the International Atomic Energy Agency, data centers, AI operators and cryptocurrency miners are increasingly eyeing advanced nuclear to meet their growing power needs. The pipeline of conditional agreements between data-center operators and small modular reactor projects swelled from about 25 gigawatts at the end of 2024 to roughly 45 gigawatts within a year — a near-doubling of proposed capacity in a matter of months.
Small modular reactors, or SMRs, are the technology at the center of that scramble. Built in factories and shipped in sections, they are meant to be cheaper and faster to deploy than the sprawling conventional plants of the past, and small enough to sit beside a data-center campus rather than on a distant coastline. Technology firms have also snapped up a large share of corporate power-purchase agreements for renewables, but the reactor deals are what signal a bet on firm, dispatchable clean power for the long haul.
The catch is time
For all the momentum, the nuclear fix runs headlong into the calendar. Reactors — even the smaller, modular kind — take years to license, build and connect, while AI demand is growing quarter by quarter. Many of the announced small-reactor projects will not deliver electricity until the 2030s, well after the current surge in computing power has to be supplied by something. In the meantime, operators are leaning on the existing grid, on natural gas and on whatever firm capacity they can secure.
That gap between a decade-long build-out and an immediate, exploding demand is the defining tension of the moment. The technology that promises to solve the AI industry’s energy problem is real and advancing, but it cannot arrive fast enough to power the machines already being switched on. How the world bridges those years — and how much carbon it burns doing so — may matter as much as the reactors themselves.
This article was produced with the assistance of AI and reviewed by Morning Overview editors prior to publication.
More from Morning Overview
