The AI industry's power crunch has turned nuclear restarts into a hyperscaler procurement strategy

The continued progress of the restarted Three Mile Island reactor supplying dedicated power to Microsoft's data-centre operations, alongside a growing list of comparable nuclear power-purchase agreements signed by Google, Amazon and Meta with both restarted legacy reactors and small-modular-reactor developers still years from commercial operation, has confirmed that electricity supply, not chip availability, has become the binding constraint on how quickly hyperscalers can bring new AI compute capacity online. Each of these deals represents a hyperscaler effectively underwriting nuclear-power-plant economics - either by restarting mothballed capacity or by providing the offtake commitments that make new small-modular-reactor projects financeable - in exchange for guaranteed, carbon-free electricity supply that grid-scale renewable and natural-gas alternatives cannot match on reliability terms for the always-on load that AI data centres require. Grid operators in several US regions have flagged data-centre electricity demand as the single largest driver of near-term transmission and generation capacity planning, a dynamic that has created friction with residential and industrial electricity customers in the same grid regions who have seen electricity price increases partly attributed to AI-driven demand growth, and several state utility regulators have opened proceedings specifically examining whether data-centre customers should bear a larger share of grid-upgrade costs rather than spreading them across the broader ratepayer base. The nuclear-restart and small-modular-reactor commitments carry execution risk that is meaningfully different from the chip-supply risk the industry has grown accustomed to managing - nuclear projects face regulatory approval timelines, construction-cost overrun risk and safety-review processes that operate on multi-year timescales largely outside any technology company's direct control, meaning hyperscalers making long-dated power commitments are accepting a form of infrastructure risk quite different from the compute-procurement risk they manage elsewhere in their AI build-out. India's own AI infrastructure build-out faces a comparable, if less acute, power-availability question, and Indian data-centre operators and the broader power sector have watched the US nuclear-restart wave as a possible template, even as India's own nuclear capacity-expansion plans move on a timeline that is unlikely to intersect meaningfully with the compute build-out happening over the next several years. What to watch: whether any of the announced small-modular-reactor projects hit a construction or regulatory delay significant enough to affect the data-centre capacity plans tied to them, how state utility regulators resolve cost-allocation disputes between data-centre and residential ratepayers, and whether India's power sector begins exploring comparable dedicated-generation arrangements for its own AI data-centre ambitions.

Original source: Reuters