00 / SHORT ANSWER
The honest answer starts with 485 TWh — then immediately adds context
The International Energy Agency estimates that all data centres used about 485 terawatt-hours of electricity in 2025. AI is a fast-growing part of that total, yet the world does not meter every AI workload separately.
The IEA’s 2026 central projection puts global data-centre electricity use near 950 TWh in 2030, around 3% of global demand. Electricity used by AI-focused facilities roughly triples over the same period. Those figures describe a sector, not a single chatbot response.
The local picture can feel much larger. A 100 MW AI-focused facility can consume electricity on the scale of roughly 100,000 households over a year, and planned campuses can be far bigger. Concentrating that load behind one grid connection creates a different engineering problem from adding the same energy use across an entire country.
01 / THREE SCALES
One question is hiding three different electricity meters
All data centres
AI shares buildings and power systems with cloud storage, search, streaming, enterprise software, and conventional computing.
An AI-focused campus
Installed accelerators, utilisation, cooling, electrical losses, and PUE turn a nameplate capacity into actual annual energy.
A training run or inference service
Model size, hardware, batch size, response length, and software efficiency can move the answer by orders of magnitude.
Where the load lands
A globally modest percentage can still dominate new demand in a constrained region, especially when projects cluster together.
FACILITY CALCULATOR / SCENARIO MODEL
Turn a nameplate into an annual electricity bill
Set an IT load, average utilisation, and power usage effectiveness (PUE). The calculator converts steady power into annual energy; it does not estimate a specific company or model.
96.0 MW average grid draw
Annual GWh = IT load × utilisation × PUE × 8.76. The household comparison is derived from the IEA’s illustrative 100 MW ≈ 100,000 households scale; household use varies widely by country.
02 / THE PER-PROMPT TRAP
A per-prompt number can be real and still travel badly
Google measured a median Gemini Apps text prompt at 0.24 Wh in May 2025. That is a useful, unusually transparent measurement for one production system at one point in time. Google also reported a 33-fold fall in energy per median prompt over the preceding year.
Copying 0.24 Wh onto every model, image request, long-running agent, or training job would erase the variables that matter. A short batched text completion and a multi-minute reasoning workflow occupy very different slices of hardware. Product-level measurements need a model name, date, request mix, system boundary, and methodology attached.
Global scale tells us where the grid is heading. Facility scale tells us what must be built. Per-request scale tells us how efficiently a product runs. Keep all three meters visible.
03 / SOURCES
Primary reports used for this answer
International Energy Agency
Key Questions on Energy and AI — Executive summary
Updates the global estimate from about 485 TWh in 2025 to about 950 TWh in 2030, while AI-focused data-centre use roughly triples.International Energy Agency
Energy demand from AI
Breaks down servers, storage, networking, and cooling, and explains the annual scale of a 100 MW AI data centre.Lawrence Berkeley National Laboratory