Data-centre electricity demand grew 17% in 2025, says IEA

The IEA's April 2026 update estimates that global data-centre electricity consumption rose 17% in 2025, against 3% growth in electricity demand overall. Its central outlook puts data centres at roughly 485 terawatt-hours in 2025 and 950 TWh in 2030. The latter is a projection across all data-centre workloads, including AI.

That growth is occurring while individual AI tasks become more energy-efficient. Hardware and software improvements reduce the resources needed for a given task, but adoption and the mix of tasks change at the same time. Video generation, extended reasoning and agents that make repeated model calls can have very different consumption profiles from a short text response. The report therefore tracks both efficiency and the expansion of use.

The physical constraint is increasingly the sequence in which infrastructure becomes available. The IEA identifies pressure on transformers, turbines, chips and grid approvals. A developer may secure one part of the system while waiting for another. An accelerator order, a building permit and a power agreement each remove a different obstacle; none establishes that the complete facility can start operating on schedule.

For readers following AI investment, this changes what counts as progress. Announced spending describes a commitment. An energised, commissioned site describes capacity that can begin earning a return. The distance between the two matters when comparing buildout plans.

Strategic impact

Impact

High

Horizon

2026–2030

Regions

Global

Affected sectors

Electricity · AI infrastructure

Key players

IEA · Grid operators · Data-centre developers

The commercial implication is that delivery confidence deserves a price alongside electricity itself. A site with a low tariff can become expensive if equipment sits idle, financing costs accumulate or customers must be served elsewhere while connection work continues. That exposure should appear in project economics before procurement locks in the schedule.

Power arrangements also need to be read closely. Annual energy purchases, physical connection capacity and continuous availability answer different questions. A contract can address the first while leaving the other two dependent on network upgrades or additional generation. This makes the quality of the development plan harder to capture in a single headline megawatt figure.

For cloud customers, infrastructure bottlenecks could show up as restricted regional availability or a narrower choice of systems. For developers and their financiers, the relevant comparison is the cost of usable capacity over time. Our assessment is that projects with explicit dependencies, staged commitments and credible commissioning milestones are easier to evaluate than plans built around a single ambitious opening date.

What to watch next

Follow the conversion from announced projects to connected facilities, then from installed equipment to sustained utilisation. Delays at each step have different causes and require different remedies; combining them into one capacity figure conceals where the constraint sits.

The next IEA update will also help distinguish a temporary equipment shortage from a longer-lived limit on growth. Watch whether additional manufacturing capacity and grid investment shorten delivery timelines, and whether demand shifts toward more intensive workloads faster than efficiency improves. The 2030 estimate should be read as a scenario to update as those variables move.