Power, Not Chips, Is Now the Binding Constraint on AI Data Centre Expansion

The bottleneck narrative around AI infrastructure has shifted decisively from semiconductor supply to electricity. Accelerator availability, which dominated planning conversations during the acute shortage, has eased relative to demand; grid capacity has not. Operators describe a situation where the constraint on bringing capacity online is neither hardware nor capital but the ability to secure a firm power connection at the required scale, in a location where the grid can actually deliver it, on a timeline that matches the rest of the project. The mismatch is one of planning horizons. A data centre can be built in roughly a year to eighteen months. A transmission upgrade, and the interconnection study and queue process that precedes it, runs on a multi-year cycle in most major markets, because the transmission planning process was designed around load growth that was predictable and gradual. A single large AI campus can request a connection comparable to a small city's peak demand, arriving as a step change in a queue built for increments. The consequence is visible in siting decisions, which have reoriented around existing generation and transmission capacity rather than around network latency, tax treatment, or proximity to customers. The responses operators have adopted are varied and largely additive rather than alternative. Long-term power purchase agreements with generators have become standard for large facilities. Nuclear interest, including restart of retired capacity and forward commitments to small modular reactor projects, has moved from press release to signed agreement in several cases, though the delivery timelines mean this addresses the end of the decade rather than the current constraint. On-site generation and grid-interactive designs that curtail load when the system is stressed have moved from experimental to contractual, because demonstrating flexibility can move a project forward in an interconnection queue in a way that being a well-capitalized customer cannot. The dimension attracting the most political attention is who pays for the network upgrades a large new load requires. Utility regulators in several jurisdictions have opened proceedings on cost allocation, on the argument that existing ratepayers should not subsidize transmission built for a single industrial customer, and on whether large loads should be required to demonstrate they can be curtailed during system stress. The outcomes vary by jurisdiction, but the direction is consistent: the assumption that a large new load simply connects and pays the standard tariff is being replaced by bespoke arrangements with obligations attached, which adds negotiation time to projects whose economics were modelled on faster deployment.