What changed
The International Energy Agency reported that electricity use by data centres rose 17% in 2025. Its April 2026 assessment identifies grid connections, equipment supply and approvals as constraints on expansion. This is a reading of that report, not a claim that a new power shortage began today. IEA: Key Questions on Energy and AI, 16 April 2026
Why it matters
The structural question is moving beyond who can buy chips. A computing facility also needs a dependable electricity supply. Our interpretation: a location with available power and a credible connection schedule can be more useful than a larger project that exists only on paper. Hardware orders and usable computing capacity are different milestones.
Efficiency is only half the equation
A simple hypothetical shows the difference between efficiency and total demand. If each task uses half as much electricity but the number of tasks triples, total consumption rises by 50%. This arithmetic is illustrative, not an estimate of current AI use. Lower costs can expand the set of affordable applications without reducing the system’s total resource needs.
Who carries the cost?
For a community considering a new facility, the important questions are specific: who finances additional grid equipment, what happens if the customer leaves, and which costs can reach other electricity users? The answer depends on local contracts and regulation. A data-centre announcement alone does not establish either a household bill increase or a local economic windfall.
What to watch next
Track completed grid connections, operating capacity and the terms used to recover infrastructure costs. For an AI buyer, also watch service availability and contract commitments. These observations test whether investment is becoming a usable service. They are more informative than adding together every announced project as though each will be built on schedule.
Sources & methodology
Sources checked on 25 September 2026. Company and institutional statements are attributed; hypothetical examples are labelled. Interpretations are identified in the text. No original interviews or hands-on product tests are claimed.