AI Needs Electricity

Every model response has a physical base: servers, cooling, networks and power plants. The growth of data centres is making that base more visible. The International Energy Agency (IEA) estimates that data centres used around 415 terawatt-hours (TWh) in 2024, or 1.5% of global electricity. In its base case, consumption could rise to around 945 TWh by 2030.

Global figures conceal strong geographic concentrations. In 2024, the United States, China and Europe accounted for 45%, 25% and 15% of data-centre electricity use, respectively, according to the IEA. Although the global share remains relatively modest, local impacts can be significant: computing infrastructure is concentrated in a limited number of hubs and competes for grid connections, available capacity and equipment.

The grid is a practical constraint. The IEA estimates that without action around 20% of planned data-centre projects could face delays. In advanced economies, new transmission lines can take four to eight years to build; waiting times for critical transformers and cables have doubled in three years. An announced data centre is therefore not the same as computing capacity ready to use.

The 945 TWh estimate for 2030 is a scenario, not a certain future measurement. The IEA highlights uncertainty around AI adoption, efficiency gains, the mix of workloads and the pace of grid and power-plant construction. Its scenarios for 2035 range from around 700 to 1,700 TWh of data-centre demand.

The energy response is also mixed: renewables, gas and nuclear contribute in different proportions depending on location and scenario. The key question for a proposed site is therefore local: when can it secure a connection, what mix will supply it and how might it affect prices, grids and other users?

For companies buying AI services, the footprint depends not only on the model but also on how often it is used, the workload and where it runs. Sustainability cannot be inferred from one general consumption figure; comparable measurements by service and activity, with transparent assumptions, are needed.

For infrastructure planners, access to the grid may become as important to competitiveness as capital or chips. Efficiency, flexible loads and locating projects where capacity is available can ease bottlenecks, but they do not replace investment in the grid.

AI is often described as immaterial because people use it through a screen. Its expansion will also depend on very concrete choices about energy, land and infrastructure. Who will bear the local costs of computing, and who will receive its benefits?

These figures are IEA estimates and scenarios published in 2025, not actual measurements for future years. Outcomes depend on explicit assumptions about demand, efficiency and infrastructure delivery.