Products, laws and numbers here change fast. Checked October 2026.
In plain words
Every AI answer runs on chips in a data center, and those chips run on electricity. Nearly all of it ends up as heat, so the building also needs cooling, which uses more power and, in many places, water.
For one question the amount is small. Google says a typical text prompt to its Gemini app used about 0.24 watt-hours in 2025, less than watching TV for nine seconds. Outside estimates for ChatGPT are similar, around 0.3 watt-hours. Long documents, images, video and reasoning models use much more.
The total is big because the numbers multiply: billions of prompts a day, plus training new models. The International Energy Agency estimates all data centers used about 1.5% of the world's electricity in 2024, and that this could more than double by 2030, with AI the main driver.
Everyday example
One more ceiling fan
One ceiling fan barely moves an electricity bill. But if every home in Mumbai switched on one extra fan, the city's power demand would jump. AI is like that. A single question uses about as much electricity as a few seconds of TV. The concern is not your one question but billions of them every day, plus the large one-off cost of training each new model.
Where the example breaks: Extra fans spread across a whole city. AI's demand lands wherever the data centers are, often concentrated in a few places, so the strain on local grids and water can be much sharper there.
How it actually works
AI's footprint has two parts: training, a large one-off cost per model, and inference, a small cost per answer repeated billions of times.
Where the electricity goes
Most goes to the AI chips. The rest powers host computers and memory, spare chips kept ready on standby, and the building's cooling and power losses. Google's 2025 figure counts all of these; counting the AI chips alone would give less than half.
Why per-prompt figures vary
Results depend on the model, the chips, the length of prompt and answer, and what is counted. Epoch AI estimates a typical ChatGPT query at about 0.3 Wh, but almost 40 Wh for a 100,000-token input. Company figures, such as Google's 0.24 Wh and the 0.34 Wh average given by OpenAI chief Sam Altman, are self-reported.
Water
Many data centers cool with water, part of which evaporates. Google puts a median Gemini prompt at 0.26 millilitres, about five drops. Totals depend on climate: in cool places outside air does most of the work, and newer designs reuse water in closed loops.
Efficiency up, demand up faster
Chips and software keep improving: Google reported a 33-fold fall in energy per median prompt in the year to May 2025. But use is growing faster. The IEA said in April 2026 that data center electricity use rose about 17% in 2025, and about 50% at AI-focused ones.
The effects are local
A typical AI-focused data center uses as much electricity as 100,000 homes, and the largest being built, 20 times more. Grids, local prices and water supplies near them feel it first. Whether the power comes from coal or solar decides the carbon cost.
- All data centers, 2024
- About 415 TWh of electricity, around 1.5% of the world's use (IEA, April 2025)
- Forecast for 2030
- Around 945 TWh, more than double, roughly what Japan uses in a year (IEA)
- Growth in 2025
- Data center use up about 17%; AI-focused data centers up about 50% (IEA, April 2026)
- One text prompt
- About 0.24 Wh and 0.26 mL of water for a median Gemini prompt (Google, August 2025); about 0.3 Wh for a typical ChatGPT query (Epoch AI estimate, February 2025)
A terawatt-hour (TWh) is a billion kilowatt-hours, the units on a home electricity bill.
Is AI's energy use a serious problem?
One view
A growing strain
Demand is rising fast, is concentrated in a few places, and can push up local power prices, water use and emissions, especially where grids still rely on fossil fuels.
The other view
Often overstated
Each prompt uses very little and the figure keeps falling. Data centers are still a small share of world electricity, and efficiency gains are large.
Most likely: Both are partly right. Per-answer energy is small and falling, but total demand is growing quickly and lands in particular places, so the real pressure points are local grids, water and how clean the power is.
Questions people ask
How much energy does one ChatGPT question use?
Roughly 0.3 watt-hours for a typical text question, by Epoch AI's 2025 estimate, close to the 0.34 that OpenAI chief Sam Altman gave in June 2025. That is about what an LED bulb uses in a few minutes. Very long inputs, images, video and reasoning models can use many times more.
Does AI use a lot of water?
Per question, very little: Google estimates about five drops for a median Gemini text prompt. In total it adds up, because many data centers cool themselves by evaporating water. How much depends on climate and design; some sites use outside air most of the year, and newer ones recycle water.
Which uses more energy, training AI or using it?
Training a large model is a big one-off cost. Using it is tiny per question but repeated billions of times, for as long as the model is in service. Companies rarely publish the split for their own models, so the exact balance is uncertain.
How much electricity will AI use by 2030?
The International Energy Agency projects that all data centers together will use around 945 TWh by 2030, more than double 2024 and roughly what Japan uses today, with AI the biggest driver. Forecasts like this are uncertain and depend on how fast AI use grows.