Grid Operators Meet the AI Industry

Electricity systems are planned on decade-long horizons with highly predictable demand, and utilities are staffed by people who are good at exactly that. Data centre developers arrive with requests for hundreds of megawatts, a construction schedule measured in months, and a willingness to pay above market rates.

The mismatch between those two cultures is one of the most consequential infrastructure stories of the decade. It is not solved by a better filing form, and it is not solved by longer hours at the same commission.

Seven years behind a project that has not been built

Connecting a large load to the grid requires studies of whether the network can support it, which upgrades are needed, and who pays for them. Queues have grown long, driven partly by renewable generation projects and partly by new data centre demand. Wait times of several years are common.

• Large loads move to the front of the queue in some jurisdictions, raising fairness questions.
• Upgrade costs are allocated through rules that were not designed for single enormous customers.
• Demand forecasts must be built from announcements that may never be built.

Five-year forecasts built out of press releases

A grid planner needs to know how much power a region will need in five years. Data centre demand depends on capital markets, model architectures and corporate strategy, none of which are stable enough to underwrite a transmission line.

Interrupting a training run, by contract

The likely outcome is a set of arrangements that would have seemed strange a few years ago: co-located generation, contracts that allow a utility to interrupt a training run, and large customers funding network upgrades directly.

Training loads behave. Inference loads do not.

Training loads are large and relatively constant, which is easier for a grid to serve than a volatile industrial load. Inference loads track user demand, which means they fluctuate and can stress the network at exactly the times it is already strained — the early evening peak, for instance.

Utilities have responded with contracts that allow curtailment: the operator agrees to reduce demand during peak periods in exchange for favourable rates or expedited connection. Whether an AI facility will genuinely reduce its consumption on request has not been tested at this scale before.

Who pays for the line to one customer

Upgrading a transmission network to serve one large customer costs money that is recovered through rates. Whether those costs should fall entirely on the customer that caused them or be shared across all ratepayers is contested, and the answer varies by jurisdiction.

This is the issue most likely to generate public conflict, and it is already generating some. Electricity bills are visible to every voter, and a data centre’s consumption is easy to characterise as an imposition on households.