The International Energy Agency published its first dedicated report on artificial intelligence and energy this month. Two findings stand out, and they are not equally firm. The 2024 baseline is well established: global data centres consumed 415 TWh of electricity, around 1 to 1.5% of the world total. The trajectory is softer terrain: consumption rose 17% in 2025, AI-specific facilities surged 50%, and the IEA projects a near-doubling of data centre demand by 2030. Competing 2030 estimates range from 700 to over 1,800 TWh, so the central projection is best read as one scenario among several rather than a prediction.
Sources: IEA Energy and AI (2026); Carbon Brief. Highlighted squares are rounded for display: data centres represent approximately 1.4% (415 TWh of roughly 29,000 TWh global total), Bitcoin approximately 0.7% (~200 TWh).
The 415 TWh consumed by all global data centres in 2024 represents approximately 1 to 1.5% of total global electricity use, with an associated CO2 contribution of roughly 0.5% of global emissions, according to Carbon Brief analysis of the IEA data. Bitcoin mining consumed an estimated 173 to 228 TWh of electricity in 2024, roughly half the data centre total, despite receiving considerably more sustained public attention on energy grounds.
Projections for 2030 data centre demand vary by a factor of nearly three: from approximately 700 TWh to over 1,800 TWh across published analyses. The IEA's central estimate of 945 TWh, a near-doubling from the 2024 baseline, sits toward the lower end of that range and implies roughly 15% annual growth. Straight-line projections of this kind tend to be widely cited and weakly predictive. They assume continued buildout at recent rates and predictable efficiency gains, when in practice the 2030 figure will be shaped by grid bottlenecks, chip supply, water constraints, model efficiency improvements, and demand-side rebound effects, none of which currently have well-established trajectories.
The comparative shape of growth is more useful than any single point estimate. The same IEA analysis projects that electric vehicles will add 838 TWh of new electricity demand by 2030, air conditioning 651 TWh, and industrial sectors 1,936 TWh. Under those projections, data centres represent roughly 8% of the total projected increase in global electricity demand through 2030. Each of these forecasts carries similar uncertainty, but together they describe a structural fact: electricity demand is rising broadly across many sectors simultaneously, and data centres are one contributor among several rather than the singular driver they are often presented as.
The most specific concern in the IEA report is not the absolute scale of consumption but the fuel mix supplying it. Approximately 27% of global data centre electricity currently comes from renewable sources. In the United States and China, where the majority of data centres are located, the share sourced from fossil fuels remains above 70%. The IEA projects that in both countries, most of the additional demand through 2030 will be met by fossil fuel generation rather than dedicated new renewable capacity.
This sits in tension with the commitments made by the sector's largest operators. Google reports operating at approximately 90% carbon-free energy on an hourly basis, with a commitment to 100% by 2030. Amazon committed to 100% renewable energy for its operations by 2025. Meta reports having matched 100% of its consumption with renewable sources since 2020. These figures are met primarily through power purchase agreements: contracts that fund renewable capacity on the broader grid without guaranteeing that any specific facility draws clean electricity at any given moment. On that accounting basis, data centres sign roughly 40% of all corporate renewable power purchase agreements globally, making the sector a material driver of clean energy investment even where consumption and generation are not directly paired.
Straight-line projections do not capture efficiency improvements, which have been substantial. The IEA notes that AI inference efficiency improved approximately tenfold between early GPT-4 and GPT-4o. If algorithmic and hardware improvements continue at comparable rates, total energy consumption may grow more slowly than demand-side projections imply. In previous technology sectors, efficiency gains have often been absorbed by expanded use rather than reduced total consumption, a pattern sometimes called the rebound effect. The IEA's base case assumes efficiency improvements will continue but not reverse the overall trajectory of demand growth.
The IEA's stated concern is narrower than much of the reporting that followed publication. The agency found that AI's energy appetite is growing faster than the deployment of AI-based tools within the energy sector itself: grid optimisation, predictive maintenance, demand forecasting, and infrastructure planning are all areas where AI could materially reduce overall system energy consumption. That deployment is proceeding slowly. The IEA does not characterise the current situation as a fixed outcome, but as a gap that is widening, and whose direction could change depending on investment decisions and policy choices made over the next four years.