For years, the pace of data-centre construction was limited mainly by how quickly server racks could be manufactured and shipped. That constraint has been overtaken by a more stubborn one: getting enough electricity to the site at all. Interconnection queues for new grid capacity in several major data-centre markets now stretch past four years, longer than it takes to design and build the facility itself.
The response has been a scramble for power sources that do not depend on the public grid's timeline. Several large operators have signed agreements to restart or extend the life of existing nuclear plants, securing dedicated capacity in exchange for financing the plants' continued operation. Others are exploring small modular reactors, though none of the announced projects will deliver power before the early 2030s.
The economics are unusual. A data-centre operator paying a premium for guaranteed nuclear power is, in effect, betting that the cost of delay, idle server capacity waiting for grid connection, exceeds the premium many times over. For a facility that can cost billions of dollars and depreciates the moment it stops being state of the art, that bet looks increasingly rational.
Grid operators are ambivalent about the trend. Dedicated power deals reduce congestion on the public network in the near term, but critics worry they let the largest, best-capitalised players jump the interconnection queue while smaller users wait behind them.
Whatever the equity concerns, the direction of travel looks set. Power procurement, once an afterthought handled by facilities teams, has become a strategic function fought over at the same level as chip allocation, a sign of how thoroughly electricity has replaced silicon as the scarcest input in building AI infrastructure.
