The gigawatt question: AI's collision with the power grid

For the first years of the AI buildout, the scarce resource was silicon. Allocation of accelerators decided who could train what, and the industry organized itself around chip supply. That constraint hasn't vanished, but a second one has grown up beside it and, in some regions, overtaken it: electricity.

Training clusters and the inference fleets behind popular AI products draw power at a scale utilities describe in the units of heavy industry. Datacenter operators now negotiate for hundreds of megawatts per campus, and the queue of interconnection requests sitting in front of grid operators has become one of the most-watched indicators in the sector.

Two clocks that don't match

The core tension is temporal. A datacenter can be built in eighteen months. The transmission lines, substations, and generation capacity to feed it can take the better part of a decade to permit and construct. The industry's planning horizon and the grid's planning horizon are simply different clocks, and the gap between them is where most of the drama lives.

The responses are already visible. Operators site new campuses wherever surplus power exists rather than wherever engineers want to live. Long-term power purchase agreements — including a conspicuous wave of interest in nuclear capacity — have become standard boardroom strategy. And efficiency, once a sustainability slide, is now a siting decision: every watt saved per query is capacity you don't have to find.

What to watch

Three signals tell you where this is heading. Interconnection queue times: if they lengthen, power replaces chips as the industry's rate limiter. Co-location deals between AI firms and generation owners: each one is a bet that the public grid can't move fast enough. And regulatory attention: when residential ratepayers start asking who pays for grid upgrades that serve datacenters, the politics of compute change quickly.

The uncomfortable truth for an industry that iterates in weeks is that its ceiling is now set by infrastructure that iterates in decades. How that mismatch resolves — more generation, radical efficiency, or slower growth — is one of the defining questions of the next phase.