The IEA counted 134 data-center grid rules. The AI power fight is now about who pays.

Artificial intelligence has become an electricity planning problem, not simply a question about the wattage of new chips. The International Energy Agency's October 14 policy review, Powering Artificial Intelligence, surveys how regulators and utilities are responding to expanding data centers and the investments they require.

The agency documents 134 policies and regulations guiding the integration of data centers with electricity systems. Almost 40% had been announced or implemented during the previous year. More than 34 policy jurisdictions were analyzed, representing around 90% of the world's data-center capacity. This is a material regulatory shift, although a count of rules does not show that any given rule has worked.

The IEA highlights grid connection queues in which proposals can collectively request capacity exceeding today's system peak demand. Such queues mix projects at different stages, including speculative and repeated applications. Treating every application as an inevitable future load risks expensive, underused infrastructure; dismissing serious projects can leave grids unprepared.

Several jurisdictions are tightening connection requirements so developers demonstrate project readiness and financial commitment. Other approaches allow phased or conditional access to the grid, or ask large data-center customers to help provide dispatchable power or flexibility. These policies differ geographically; no single approach applies to every data center.

The financial issue is central. When a power network is expanded for a new customer and the cost is recovered from all users, existing households or businesses may face larger bills. In other circumstances, a customer using otherwise idle capacity while paying its fair share can help spread fixed costs. Which result occurs depends on the tariff and actual network constraints, not on a general claim that all data centers make power expensive.

The IEA also reports that operators signed 45 TWh of renewable electricity purchase agreements in the first half of 2026, up nearly 20% on the same period of 2025. Contracting clean energy does not necessarily eliminate grid congestion or solve the location and timing of power demand. Proposals to curtail computing during grid stress similarly require credible technical and commercial arrangements.

The accountability question is who assumes the risk if new demand arrives late, uses less electricity than reserved or never arrives. Minimum payments, collateral, phased connections and exit fees are tools regulators can use to guard against stranded assets. Their costs and benefits need transparent regional evaluation.

This story is distinct from Cortist's 2030 energy forecast. The forecast asks whether all US data centers will consume at least 12% of the nation's electricity in 2030. A count of policies across many countries neither resolves that event nor supplies a calibrated probability. Our 54% working-model scenario remains unreviewed, unsigned and unscored.

The useful metrics ahead are real grid hookups, power sold to operational sites, the amounts that customers commit to pay, what happens to unused network investments and published data on cost allocation. These are more informative than the sheer volume of announcements.