Compute and power intelligence / Global
AI's physical bottleneck is increasingly electrical.
Model capability gets the headlines. The limiting path can run through power generation, transmission, substations, cooling, equipment lead times, and local interconnection.
01 / Central scenario
The IEA's 2025 assessment also projected that U.S. data centres would account for nearly half of U.S. electricity-demand growth through 2030. These are modelled outcomes, not metered future facts.
02 / Why it matters
The investable advantage may sit outside the server rack.
For site selection, infrastructure, utilities, cooling, and regional development, the research question is not simply “how much AI demand?” It is where firm power, grid capacity, equipment, water, permits, and construction timelines overlap soon enough to support it.
03 / Boundary
The projection is sensitive to hardware and software efficiency, utilisation, AI adoption, data-centre type, and energy bottlenecks. Global electricity share does not describe a local grid: regional concentration can create material constraints even when the global share remains modest.
04 / Evidence
- International Energy Agency — Key Questions on Energy and AI
Updated central projection · accessed 22 August 2026. - International Energy Agency — Energy and AI executive summary
2025 scenario report · accessed 22 August 2026 · report material identified by IEA as CC BY 4.0.
05 / Next test
Move from global demand to a regional constraint map.
Join utility load forecasts, interconnection queues, generation additions, transmission projects, rate cases, equipment lead times, and announced data-centre capacity. Separate announced, permitted, connected, and operating capacity.