AI data centres run up against cooling constraints as nuclear reactors fail to save power grids alone

Major technology companies and governments worldwide are turning to nuclear power in search of vast, continuous electricity supplies to feed artificial intelligence data centres. Yet these bets are running up against climatic and economic variables that are redrawing the limits of reactor reliance. Nuclear energy, promoted as a direct carbon-free pillar capable of round-the-clock operation, faces environmental constraints tied directly to the physics of its operation, as reactors require massive volumes of river water for cooling and dissipating heat generated during power generation.

Recent heatwaves and droughts across Europe have exposed the vulnerability of this operating model, forcing reactors at six sites to curtail output or shut down temporarily. France was the hardest hit after its nuclear generation fell by up to 20 percent at the peak of the disruptions, removing roughly 1.7 gigawatts from the grid in late July and early August and driving up electricity prices. The pattern repeated across Central and Eastern Europe as the Danube dropped to historic lows, taking Unit 3 offline at Hungary's Paks plant, disconnecting a unit at Romania's Cernavoda facility, and affecting Switzerland's Beznau site, the world's oldest operating nuclear power plant.

These shutdowns stem from strict environmental regulations that prohibit discharging heated water into rivers during periods of elevated water temperatures and low flow to protect aquatic ecosystems. This forces plants to curtail capacity precisely when electricity demand peaks from air conditioning and computing infrastructure. Even though European reliance on nuclear power has declined from roughly a third of total generation in 1990 to around 15 percent today, revival efforts are gathering pace in Italy, Belgium, Sweden, and the United Kingdom, alongside discussions over advanced designs in Greece, driven by mounting demand from the AI sector.

Time and cost impose a different reality on rapid expansion ambitionsBuilding a new nuclear power plant takes an average of roughly 14 and a half years from planning to commercial operation, as seen with Finland's Olkiluoto 3 reactor, proposed in 2000 and entering commercial service only in 2023 after delays and multiplied costs. In contrast, wind and solar projects come online within months or a few years. Levelised cost of electricity data published by Lazard in 2018 underscores this gap, with the cost per megawatt-hour from new nuclear plants reaching around $155, compared with $43 for onshore wind and $41 for utility-scale solar.

Nuclear calculations are further complicated by long-term radioactive waste, which requires hundreds of millions of dollars annually to manage amid a scarcity of permanent deep geological repositories beyond limited projects such as Finland's Onkalo repository. While the energy sector promotes small modular reactors as a faster-to-build, lower-cost solution for data centres, most of these designs remain unproven at commercial scale in the West. Indeed, a 2022 study by researcher Lindsay Krall and colleagues at Stanford University, published in the Proceedings of the National Academy of Sciences, showed that certain modular reactor designs may produce more radioactive waste per unit of energy than conventional reactors due to neutron leakage. The November 2023 cancellation of NuScale's project in Idaho, after estimated electricity costs rose from $58 to $89 per megawatt-hour, underscored the widening gap between theoretical promises and practical execution, making reliance on nuclear power alone as an immediate fix for AI's energy crunch an incomplete bet that demands supply diversification rather than pinning hopes on a single path.