How much energy does AI use?

A headline compares an AI question with your kettle. What did it count? Every answer needs electricity, cooling, and chips that took energy to make. Estimates for 1 question, training 1 model, and every data center on Earth measure different things. Check what each includes before comparing them.

A data center is a building full of servers. For AI, many of those servers hold accelerators: chips designed to do the huge amounts of arithmetic a model needs. The building also runs cooling, power conversion, networking, and spare machines kept ready for busy moments. All of it runs on electricity, and many cooling systems use water.

AI work there has 2 main kinds. Training builds the model: thousands of accelerators run for weeks on a large collection of text and other data, once per model version.

Inference uses it. A server reads each question and writes an answer. Each answer is cheap, but billions a day add up throughout the model's service life. The how AI answers work guide shows what happens during one of those answers.

A third cost sits before both: making the chips and building the data centers. Researchers call this embodied energy or emissions. Many per-question figures leave it out.

The IEA, an international energy body, publishes the most cited global numbers. Its 2025 Energy and AI report estimated that data centers used 415 terawatt-hours (TWh) of electricity in 2024, about 1.5% of world electricity use. In its central case, that more than doubles to around 945 TWh by 2030. For 2035, its cases range from 700 to 1,700 TWh, depending on how fast AI spreads and how quickly power supply and efficiency keep up.

A year later, Key Questions on Energy and AI (April 16, 2026) put 2025 use at 485 TWh, up 17%, with AI-focused data centers growing 50%. Its central projection is about 950 TWh in 2030, around 3% of global electricity demand.

These totals cover all data centers, including streaming video, email, banking, and cloud storage. The 2030 and 2035 values are scenarios. The IEA calls efficiency and uptake trends "rapidly evolving and uncertain." A small global share can still strain a local grid.

The 2026 report also names physical limits. It describes a shortage of high-bandwidth memory, a chip component AI servers depend on, expected to last through at least the end of 2027. It reports slow grid connections and a 70% rise in gas turbine orders in 2025. Chip factories, grid lines, and power plants take years to build.

In August 2025, Google published a measurement of its Gemini Apps assistant. For the median text prompt, using May 2025 data, it reported 0.24 watt-hours of energy, 0.03 grams of carbon dioxide equivalent, and 0.26 milliliters of water. By the paper's comparison, that is about 9 seconds of television.

It counted accelerators, the host computer's processor and memory, idle machines, cooling and power overhead, and estimated chip-manufacturing emissions. Google's blog post on the study shows why that matters. Counting only the active chips gives 0.10 watt-hours, less than half the fuller figure.

It excluded training, outside networks, and your phone or laptop. The median is the middle text prompt; long, heavy requests sit above it. Google measured its own service, hardware, and electricity mix without independent verification. The blog also reports that energy per median prompt fell 33 times between May 2024 and May 2025.

The IEA's 2026 report says energy per AI task has dropped by at least an order of magnitude a year recently. It also says video generation, reasoning, and agent tasks can use hundreds or thousands of times more energy per request than a simple text answer. Stanford's 2025 AI Index reports hardware energy efficiency improving about 40% a year. Cheaper questions and more of them can still add up to more electricity.

Training figures are rarer. A 2021 study by Google and University of California, Berkeley researchers estimated that training GPT-3, a 2020 model, used about 1,287 megawatt-hours and emitted about 552 tonnes of carbon dioxide equivalent. That estimate used information OpenAI shared, plus assumptions about the data center.

As a rough illustration only: 1,287 megawatt-hours is the energy of about 5 billion prompts at 0.24 watt-hours each. That arithmetic mixes a 2020 model on one company's servers with a 2025 assistant on another's, so it shows scale, not a real exchange rate. Training is a large one-time cost; inference grows with every person who uses the model.

The same study found that model design, chips, data center, and location together could change a training footprint by up to roughly 100 to 1,000 times.

A study on AI's water footprint by University of California, Riverside researchers separates 2 questions:

• Withdrawal is water taken from rivers or the ground, even if it goes back. Consumption is water that does not return, mostly evaporated in cooling.
• On-site water cools the building. Off-site water is used by power plants to make the electricity the building runs on.

Google's 0.26 milliliters counts on-site cooling. The Riverside study's estimate of a 500 milliliter bottle for every 10 to 50 medium responses counts both on-site and off-site water, for GPT-3 under 2023 assumptions. They answer different questions. Newer systems have changed the older assumptions; climate and cooling design also affect the figure.

For an everyday user, a single short text question is small: the IEA says it typically uses less electricity than running a television for the same time. The larger effects come from what is used at scale and how: long documents pasted into every turn, video generation, and agents that run many steps. Pick the simpler request when it does the job.

Builders can save energy and money by using the smallest model that passes a real task test, sending less text, and reusing unchanged answers. The AI costs and tokens guide explains input costs. Location matters too: Google Cloud publishes carbon-free energy percentages by region for 2025 and suggests the cleanest region for flexible batch jobs.

An authored illustration: the organizer of a fictional neighborhood book swap asks an assistant for 20 invitations, one per street. Those short text requests are a tiny load. Asking for a video invitation per street could take hundreds of times the energy per request, by the IEA's range. The method changes the energy needed for the same task.

A fictional headline, written for this lesson:

"Asking AI a question uses as much power as charging your phone, and AI will soon use more electricity than whole countries."

Before reading the answer, list what you would need to know to judge it.

Answer: The headline gives no model, no product, and no year. It does not say whether "a question" is a short text prompt, a long document, or a video, or whether the figure is a median or a worst case. It does not say what was counted: chips only, or the whole building, cooling, and spare machines. Training may be mixed into a per-question cost. "Whole countries" could mean all data centers or AI alone, measured or projected. Ask who measured it and whether anyone checked. A claim that answers these questions might still be right. A claim that answers none of them is a mood, not a measurement.

For more practice reading claims like this, see understanding AI news.

• IEA, Energy and AI executive summary: the 2024 baseline and scenarios to 2035.
• IEA, Key Questions on Energy and AI executive summary: the April 2026 update.
• Google, Measuring the environmental impact of AI inference: what a per-prompt figure includes, in plain terms.
• MIT News, Explained: Generative AI's environmental impact: a January 2025 overview of electricity, water, and chip manufacturing.
• Li and colleagues, Making AI Less "Thirsty": the water definitions above, in a technical paper.