Powering Artificial Intelligence (IEA, October 2026): hyperscaler capex tops $700B in 2026, straining grids

Requested by France's G7 presidency, this IEA report surveys how 34 jurisdictions, about 90% of global data centre capacity, connect AI data centres to the grid. It counts 134 policies and regulations on the topic, nearly 40% of them announced or implemented in the last 12 months, and notes that hyperscaler capital spending will top $700 billion in 2026, most of it on new data centres. Connection queues in Texas, Virginia, Finland and the UK now exceed today's peak demand, much of it speculative, so planners are tightening queue entry rules and moving away from first-come-first-served.

Operators signed 45 TWh of power purchase agreements in the first half of 2026, up nearly 20%, with 70% in the US, and a growing number of onsite gas, geothermal, small nuclear and storage deals. The IEA estimates non-firm, curtailable connections could unlock 750 to 900 GW of additional hosting capacity worldwide.

For anyone building AI infrastructure or compute-heavy products, this is the clearest map of where power will be available, on what terms, and how fast permitting is likely to move.