How much electricity does AI use? Per-query, per-year and per-country numbers, explained

A single AI text prompt uses about a quarter of a watt-hour; data centres used 1.5% of the world's electricity in 2024 and 2% of Great Britain's. Here are the credible figures, why they vary, and what they mean for a UK home or small firm.

Business & home By suj · · 8 min read · 1,800 words

Search the blog

Ask ten people how much electricity an AI chatbot uses and you will get answers from "nothing, it's just software" to "a kettle every time you press Enter". Both are wrong, and the honest answer has two halves. Per question, the cost is tiny: Google's own measurement puts a typical Gemini text prompt at 0.24 watt-hours, the energy a television uses in under nine seconds. Added up across billions of prompts, plus training runs and everything else data centres do, the total is not tiny at all: the International Energy Agency (IEA) estimates data centres used 415 terawatt-hours (TWh) in 2024, about 1.5 per cent of the world's electricity, and expects that to more than double by 2030.

This article sets out the figures that come from measured data or peer-reviewed work, explains why per-query estimates differ so widely, and puts them next to things you already pay for: a phone charge, a day at home, and Great Britain's total demand.

What a data centre is, and where the electricity goes

A data centre is a building full of servers: computers that store data and run software, including the models behind AI assistants. AI servers carry specialised chips (graphics processing units, or GPUs, and similar accelerators) that do the arithmetic. Servers account for around 60 per cent of a modern data centre's electricity; storage takes about 5 per cent, networking up to 5 per cent, and cooling anywhere from 7 per cent in an efficient hyperscale site to over 30 per cent in a less efficient corporate one (IEA, 2025b).

Google reports a fleet-wide power usage effectiveness (PUE) of 1.09, meaning only 9 per cent goes on cooling, power distribution and everything that is not the computers themselves (Google, 2025).

How much energy does one AI query use?

The best measured figure comes from Google's technical report of 21 August 2025, which used production data rather than a lab test. Its headline: the median Gemini Apps text prompt uses 0.24 Wh of energy, emits 0.03 grams of CO2-equivalent and consumes 0.26 millilitres of water, "about five drops" (Google, 2025). Over the previous twelve months the energy per prompt fell 33-fold and the carbon footprint 44-fold, which is a useful reminder that any figure older than a year is probably stale.

Why do other estimates come out higher or lower? Mostly because of what is counted:

  • Boundary. Google shows that counting only the active AI chips gives 0.10 Wh; adding idle capacity held in reserve, the host CPU and memory, and data-centre overhead brings it to 0.24 Wh. Many published calculations "only include active machine consumption" and so describe theoretical rather than real efficiency (Google, 2025).
  • Task. Luccioni, Jernite and Strubell (2024) benchmarked 88 models across 10 tasks on a single GPU. Text classification averaged 0.002 kWh per 1,000 inferences, text generation 0.047 kWh, and image generation 2.907 kWh per 1,000: roughly 2.9 Wh per picture. Generative tasks are far more energy-hungry than classifying or sorting, and the least efficient image model used the energy of 522 smartphone charges for 1,000 images.
  • Model size and output length. A long essay costs more than a yes/no answer; a large model costs more than a small one. The same paper found a spread of over 1,450 times between models doing the same job.
  • Age of the estimate. De Vries (2023) estimated ChatGPT could need 564 MWh a day in early 2023 and projected AI could add 85 to 134 TWh a year by 2027, comparable to the Netherlands, Argentina or Sweden. Those were calculations from hardware shipments, not measurements, and predate the efficiency gains Google describes.

The sensible reading: a plain text question costs a fraction of a watt-hour; an image costs a few watt-hours; video, long documents and "reasoning" modes that think for minutes cost more again, and nobody has published a measurement of those as rigorous as Google's text figure.

Putting a prompt next to everyday electricity

Ofgem's Typical Domestic Consumption Values, in force from 1 July 2026, put a medium-use British household at 2,500 kWh of electricity a year (low 1,600 kWh, high 3,800 kWh), about 6.85 kWh a day (Ofgem, 2026). Luccioni and colleagues put a full smartphone charge at 0.022 kWh, or 22 Wh.

Table 1. One AI text prompt compared with everyday electricity use. Sources: Google (2025); Luccioni, Jernite and Strubell (2024); Ofgem (2026); DESNZ (2026a); IEA (2025a). Ratios are our arithmetic from the cited figures.
ItemEnergyEquivalent in median Gemini text prompts (0.24 Wh each)
Median Gemini Apps text prompt (equal to under 9 seconds of television)0.24 Wh1
One AI-generated image (mean of models tested, GPU only)about 2.9 Whabout 12
Full smartphone charge22 Whabout 90
Typical UK home, one day (2,500 kWh a year)about 6.85 kWhabout 28,500
Typical UK home, one year2,500 kWhabout 10.4 million
All data centres in Great Britain, 20244.5 TWhabout 1.8 million typical homes for a year

A household would have to send roughly 28,000 text prompts a day to match its own electricity use; nobody does. But an industry serving billions of people, and training the models in the first place, draws power on the scale of whole countries.

Global and UK data-centre electricity

The IEA's Energy and AI report is the most thorough independent estimate. Its key numbers (IEA, 2025a):

  • Data centres used around 415 TWh in 2024, about 1.5 per cent of global electricity, and consumption has grown around 12 per cent a year since 2017, more than four times faster than electricity use overall.
  • The United States accounted for 45 per cent of that, China 25 per cent and Europe 15 per cent.
  • Consumption is "set to more than double to around 945 TWh by 2030", slightly more than Japan uses today.
  • A typical AI-focused data centre "consumes as much electricity as 100 000 households", and the largest under construction will use twenty times that.
  • The range is wide: by 2035 the IEA's cases span 700 to 1,700 TWh, depending on adoption, efficiency and grid bottlenecks.

For Britain the government now has measured data rather than estimates. A Department for Energy Security and Net Zero (DESNZ) analysis of meter-level readings, published on 30 June 2026, found that data centres consumed an estimated 4.5 TWh from the grid in 2024, 2 per cent of the 249.2 TWh Great Britain drew from the grid, and that consumption rose 1.3 TWh (41 per cent) between 2020 and 2024 (DESNZ, 2026a). The geography is striking: 77 per cent of that electricity was used in the South East (1.8 TWh) and London (1.7 TWh), and Slough alone accounted for 1.3 TWh, almost two-thirds of all electricity consumed in the town. For scale, UK households used 96.2 TWh in 2025 out of a national final consumption of 275.5 TWh (DESNZ, 2026b). Two per cent is modest; the growth rate and the concentration around London are what give grid planners pause.

Water

Data centres "often consume water for cooling", as Google puts it, and its figure of 0.26 mL per median text prompt covers that direct use in its own sites; making the computing more energy-efficient reduces the water too (Google, 2025). The per-prompt amount is trivial. Multiplied by a large site in a dry region it is not, which is why siting and cooling design have become planning issues in several countries. We could not find a measured UK-wide figure for data-centre water use from an official source, so we have not given one.

What is being done about it

Better chips (Google says its latest accelerator is 30 times more energy-efficient than its first), model designs that activate only part of a model per question, and better data-centre engineering explain the 33-fold fall per prompt in a year (Google, 2025). On supply, the IEA expects half of the growth in data-centre demand to 2030 to be met by renewables, with natural gas adding 175 TWh, mainly in the United States, and nuclear about as much again (IEA, 2025a). Pushing the other way is sheer usage: as each answer gets cheaper, people ask for more, so total demand keeps rising even as per-query cost collapses.

What this means for you

For a household, your own use of AI assistants is a rounding error on the electricity bill, and the water in a prompt is a few drops. Heating, hot water and an old fridge matter far more than rationing questions to a chatbot. If you want to trim your footprint, skip images and video you do not need; they cost ten to a hundred times more than text.

For a small business there are two practical edges. First, if a tender or supplier questionnaire asks about the footprint of the tools you use, the figures above give you a defensible answer, and the cloud services you already run on fall under the same accounting. Second, the electricity story is regional: the concentration of data centres in Slough and west London is worth knowing before you plan a server room, an office move or a switch to hosted services. We are happy to talk that through; see business IT support or get in touch.

The most useful habit is to check the date on any AI energy figure you are given. In this field, a number from 2023 is a historical document.

Sources

Want this handled for your business?

A short conversation with an engineer — not a salesperson — is the fastest way to find out what you actually need.

Talk to an engineer 020 3457 6267
Vision House, 3 Dee Road, Richmond TW9 2JN Registered UK company no. 09064078 No cookies, no trackers on this site