Few questions about AI generate more heat, in both senses, than its energy use. Headlines range from reassuring to apocalyptic, often citing figures that are years old or that measure different things. Here is what can be said with reasonable confidence, and where each number comes from.
Updated September 2026: we added sources for every figure, Google’s published per-prompt measurements, and details of the nuclear power agreements mentioned below.
Where AI’s energy goes
AI uses electricity in two main phases:
- Training, when a model learns from data. Training a frontier model runs thousands of specialized chips for weeks or months. It is energy-intensive but happens relatively rarely per model.
- Inference, when the model is used to answer requests. Each individual request uses a small amount of energy, but there are billions of them. As AI features spread into search, office software and phones, inference is widely expected to account for a large and growing share of AI’s total energy use.
The chips also need cooling, which adds to electricity use and, in many facilities, consumes water.
The big picture
The International Energy Agency’s Energy and AI report estimates that data centers used about 415 terawatt-hours of electricity in 2024, around 1.5% of the world’s total. It projects that figure will more than double to around 945 TWh by 2030, with AI the most important driver of growth. The United States accounted for 45% of data center consumption in 2024, and the IEA expects data centers to make up nearly half of US electricity demand growth to 2030.
| Measure | Figure | Source |
|---|---|---|
| Data center electricity use, 2024 | About 415 TWh (about 1.5% of global use) | IEA |
| Projected data center use, 2030 | About 945 TWh | IEA |
| Median Gemini app text prompt | 0.24 Wh, 0.26 ml of water | Google (company-reported) |
Data centers are not all AI, since they also run streaming, cloud storage and ordinary web services, but AI is the fastest-growing part.
The impact is concentrated. In some regions and local grids, new data center demand is large relative to available supply, which is why utilities, regulators and local communities are paying close attention.
What a single prompt costs
Per-query figures are where public debate has been weakest, because companies have disclosed little. One exception: in August 2025 Google published its methodology and reported that the median text prompt in its Gemini apps used 0.24 watt-hours of energy, about the same as watching television for under nine seconds. It also reported 0.03 grams of CO2-equivalent emissions and 0.26 milliliters of water. Google said energy per prompt had fallen 33-fold over the previous 12 months.
These figures are useful but come with caveats. They are self-reported, cover one company’s text prompts rather than images, video or long reasoning tasks, and describe a median rather than the heaviest uses.
Why estimates vary so widely
- Companies disclose limited data about individual models and their energy use.
- Efficiency is improving quickly, so figures from a few years ago can be badly outdated. Google’s reported 33-fold drop in one year shows how fast they move.
- Different tasks cost very different amounts. A short text reply uses far less energy than generating a video or running a long reasoning process.
- Measurement choices matter. Counting only the AI chips, and leaving out cooling, idle capacity and other hardware, gives much lower numbers. Google said that narrower method would have understated its figure by more than half.
- Hardware and location matter. The same computation has a different footprint depending on the chips used and the local energy mix.
Read numbers carefully When you see a claim such as “one AI query uses X times more energy than a web search,” check the date, the model, the type of task and the source. Many widely shared comparisons rest on early, rough estimates.
What is being done
- More efficient chips and models. Newer hardware delivers more computation per watt, and smaller models handle many routine tasks, as we noted in our 2026 trends piece.
- Clean energy deals. Large technology companies have signed major renewable energy contracts and have turned to nuclear power. In September 2024, for example, Microsoft signed a 20-year agreement with Constellation to buy power from a restarted reactor at the Three Mile Island site in Pennsylvania.
- Better cooling and siting data centers where power and water are more sustainable.
- Transparency pressure. Researchers and regulators are pushing for standardized reporting of AI energy use. The EU AI Act requires providers of general-purpose models to document the known or estimated energy consumption of their models.
What it means for individuals
Your personal use of a chatbot is a small part of your overall footprint compared with transport, heating and diet. The larger questions are systemic: how fast demand grows, how it is powered and whether efficiency gains keep pace. Choosing smaller, faster models for simple tasks, where you have the option, is a reasonable habit, but guilt about asking a question is not a useful response.
What we still need to know
AI’s energy demand is real, growing and unevenly distributed. It is also poorly measured in public, although disclosures like Google’s are a start. Standardized, independently checked reporting from AI companies would do more to improve the debate than any single statistic currently in circulation.
Sources
- Energy and AI: Executive summary, International Energy Agency, 2025
- Measuring the environmental impact of AI inference, Google Cloud blog, August 2025
- Constellation to Launch Crane Clean Energy Center, Constellation Energy, September 2024
- Regulation (EU) 2024/1689 (Artificial Intelligence Act), Annex XI, EUR-Lex
