AI data centers now use about 0.5% of global electricity

Hannah Ritchie’s piece for Our World in Data injects much‑needed quantitative rigour into a debate that often swings between AI‑doom energy narratives and dismissive hand‑waving. Putting AI data centres at roughly 0.5% of global electricity in 2025, and all data centres at about 1.5%, frames the problem: AI isn’t yet an energy monster in absolute terms, but demand is concentrated in a few grids and projected to grow steeply.

For AGI watchers, the key takeaway is that energy is unlikely to be a hard global cap in the near term, but it will be a serious local and political constraint. When 5% of US electricity already goes to data centres and some regions like Ireland or Northern Virginia approach double‑digit shares, adding another few hundred terawatt‑hours for AI alone becomes a siting and permitting challenge as much as a technical one. The article’s breakdown of per‑query energy use — on the order of 0.3 Wh for a typical LLM call and tens of Wh for very long or agentic queries — also helps calibrate individual behaviour against system‑level impacts.

In practice, this analysis suggests that the AGI race will hinge less on the raw availability of electrons and more on the speed at which operators can build low‑carbon, high‑capacity grids in a handful of hotspots. It also underlines that decarbonising those grids matters more for climate impact than shaving a few percent off model energy use.