AI Is Rewriting the World's Electricity Map

The most consequential constraint on artificial intelligence is not chips or talent. It is electricity, and the industry's demand is arriving faster than power systems can respond.

The scale of the problem:

A single large AI data centre can draw as much electricity as a mid-sized city, and the largest announced projects are measured in gigawatts, comparable to the output of major power stations.

Utilities in several markets have revised long-term demand forecasts upward sharply after a decade of flat or declining consumption. Planning systems, regulatory approvals and construction timelines were built for gradual change, not step changes.

The interconnection problem:

In most developed markets, the binding constraint is not generation but connection. Queues to connect large loads or new generation to the grid run for years, in some regions approaching a decade.

This is why Texas has attracted disproportionate investment: its independent grid connects large loads faster than most alternatives. It is also why Ireland and parts of the Netherlands effectively stopped new connections, and why Northern Virginia's utility faces politically fraught transmission expansion.

Who pays:

The central political question is cost allocation. Grid expansion driven by industrial demand is expensive, and if those costs enter general tariffs, ordinary consumers subsidise data centres.

Regulators in several jurisdictions are developing special tariff structures for very large loads, requiring them to fund the infrastructure they necessitate and to commit to minimum consumption so that stranded investment risk does not fall on other customers.

Public opposition has grown where bills rise visibly, and this is now a genuine electoral issue in several regions.

The generation response:

Operators have moved from buying renewable certificates toward procuring actual dispatchable capacity, because AI workloads run continuously and intermittent generation alone cannot serve them without storage.

Nuclear has returned to the conversation seriously, with agreements to restart shuttered plants, contract existing output and fund small modular reactor development. Timelines remain long and costs uncertain.

Gas generation is being built in several markets as the fastest route to firm capacity, which sits awkwardly with corporate emissions commitments.

Batteries, demand response and load flexibility are increasingly part of contracts, with data centres agreeing to reduce consumption during system stress.

Where this pushes investment:

The effect is geographic. Compute is migrating toward places with available power: Texas, the Nordics, Spain, France, the Gulf, Malaysia, and anywhere with surplus hydro or nuclear generation.

Countries with cheap, clean, dispatchable electricity have discovered an unexpected export: Norway, Iceland, Paraguay, Quebec and France all sit on assets that have become more valuable because of AI.

The efficiency question:

Chips and models are becoming more efficient per unit of computation, substantially so. Historically, efficiency gains in computing have been met with expanded use rather than reduced consumption, and current demand growth suggests the same pattern.

The outlook:

Electricity has become the strategic input of the AI economy, and energy policy has become AI policy whether governments intend it or not.

The countries that can deliver large amounts of firm, clean power quickly will host the infrastructure. Those that cannot will rent it from those that can, which is a dependency with economic and political consequences extending well beyond technology.