How much energy does AI use? Nobody can give one clean answer, and anyone who does is probably leaving something out. A single chatbot reply costs a fraction of a watt-hour by the best public estimates, which is trivial. Billions of replies a day, plus the training runs that built the models and the data centers that house them, add up to something grid planners are taking very seriously. The real picture sits between the scary headlines and the breezy reassurances, and a good deal of it rests on numbers that companies report about themselves.

Where does AI’s electricity actually go?

Three places, roughly. First is training, the one-off (though repeated) process of building a model by running enormous amounts of data through it. Second is inference, which is the everyday use: every prompt you send, every image generated. Third is everything around the chips: cooling, networking, power conversion and the manufacturing of the hardware itself.

Most public attention goes to training, because it is easy to point at a single big run. Over a model’s working life, though, inference can come to dominate, since a popular service answers vastly more requests than it ever took to build. Anyone quoting only training numbers is telling part of the story.

Another complication is that AI does not have its own electricity meter. It runs inside data centers that also handle email, video streaming, online shopping and cloud storage. Estimates of “AI’s” share are therefore built from assumptions about chip shipments, utilization and typical power draw, not from a separate measurement.

How much electricity do data centers use overall?

Here the numbers are better, because data centers are at least a measurable category.

A report published by Lawrence Berkeley National Laboratory for the US Department of Energy in December 2024 found that data centers used about 4.4 percent of all US electricity in 2023. In absolute terms, consumption had risen from 58 terawatt-hours in 2014 to 176 terawatt-hours in 2023. The report projected that the share could reach roughly 6.7 to 12 percent by 2028, a wide range that tells you how uncertain the future path is.

The International Energy Agency, in April 2025, put the global picture this way: worldwide data-center electricity demand is projected to more than double by 2030, to about 945 terawatt-hours, roughly the total consumption of Japan today. The agency identified AI as the most significant driver, with demand from AI-optimized data centers projected to more than quadruple by 2030. It also expected data centers to account for almost half of the growth in US electricity demand over that period.

Two cautions apply. These are projections, and projections about fast-moving technology have a way of missing in both directions. And “more than double” is a large increase in a quantity that is still a modest slice of total world electricity use. Both things are true at once: the growth is real, and the starting point is small.

How much energy does one AI query use?

This is the number people most want, and it varies by an order of magnitude or more depending on the model and the task.

Researchers who measured open models in 2024 found that simple classification tasks used roughly 0.002 to 0.007 watt-hours per prompt, while text generation and summarization averaged around 0.05 watt-hours. Generating an image was far heavier, averaging about 2.91 watt-hours per image, with the least efficient model tested using 11.49. A University of Michigan study from 2025 measured the open Llama 3.1 family and found about 0.03 watt-hours per response for the 8-billion-parameter version and up to about 1.86 watt-hours for the 405-billion-parameter version. Bigger models cost more per answer, as you would expect.

The companies have started publishing their own figures. In June 2025, OpenAI’s chief executive Sam Altman said an average ChatGPT query used about 0.34 watt-hours and about 0.32 milliliters of water. In August 2025, Google reported that a median Gemini text prompt used about 0.24 watt-hours and 0.26 milliliters of water, and said the energy per prompt had fallen by a factor of 33 over a year.

Those are company-reported figures. They may well be accurate, but they have not been independently verified, and what counts as an “average” or “median” prompt is the company’s own definition. Long prompts, reasoning modes, image and video generation all use more.

To put a number on 0.3 watt-hours: that is about what a 10-watt LED bulb uses in under two minutes. One query is nothing. Volume is the issue.

Is a chatbot query worse than a web search?

Probably yes, though by how much depends on whom you ask. Estimates for the electricity cost of a chatbot response compared with a conventional search run from about five times to about ten times as much. The spread comes from different assumptions about the model, the hardware and what a search involves.

A 2024 estimate suggested that putting a ChatGPT-style response into every Google search would add on the order of 10 terawatt-hours a year. That is a back-of-the-envelope scenario, not a measurement, and efficiency gains since then may have changed the arithmetic.

Treat the comparison as directional. A language model does far more computation per answer than an index lookup, so costing more is expected. The useful question is whether the answer is worth the extra, and that is a judgment call, not a physics one.

How much energy does training a model take?

Training figures are scarcer, because companies seldom disclose them for current frontier models. What exists is mostly older or outside estimates.

For GPT-3, released in 2020, one researcher’s estimate put training emissions at about 552 metric tons of carbon dioxide. The open BLOOM model, trained in 2022, was reported at about 25 metric tons for the training run itself, and roughly double that once the manufacturing of the chips was counted. A 2025 lifecycle analysis of one language model estimated about 493 metric tons of carbon dioxide including hardware manufacturing.

These are small next to the emissions of, say, aviation, but they describe models that are now several generations old. The amount of computing used to train leading models has been growing quickly; one analysis estimated it doubled about every 3.4 months. Whether current frontier training runs are much larger than those cited here is very likely, but I have not seen published figures I can quote with confidence.

What about water?

Data centers use water in two ways. They use it directly, mostly for cooling, and indirectly, through the water consumed to generate the electricity they draw. Chip manufacturing also needs large amounts of ultrapure water.

A 2023 study by Shaolei Ren and colleagues estimated that training GPT-3 evaporated about 700,000 liters of freshwater, and that a conversation of 10 to 50 medium-length responses could consume around half a liter, depending on when and where the servers run. The same team projected that global AI-related water withdrawal might reach 4.2 to 6.6 billion cubic meters in 2027.

Company reports have pointed the same way. Microsoft’s and Google’s water use rose 34 percent and 20 percent respectively in 2023 reports, increases partly attributed to AI work. Against that, the company-reported per-prompt water figures of 0.26 to 0.32 milliliters are tiny, roughly five drops. The two sets of numbers differ mainly in what they count, and in whether indirect water is included.

Location matters more here than the total. Using water in a wet region is a different matter than drawing it in a drought-prone one, and annual national averages hide that.

Are the headline claims reliable?

Some are, some are not, and the difference usually comes down to what was measured and by whom.

Be wary of figures that lack a stated scope. “AI uses as much electricity as a country” might refer to a forecast, a current estimate or a scenario, and the country in the comparison is often chosen for effect. Be wary also of per-query numbers that do not say which model, what prompt length and whether cooling and other overhead are included.

Comparisons with humans are especially contested. A 2024 paper in Scientific Reports argued that AI systems emit far less carbon than humans for writing and illustration, by a factor of 130 to 2,900. A 2025 study criticized that kind of comparison, and for programming found that GPT-4 had a carbon impact of 5 to 19 times that of human programmers. Two studies, opposite conclusions: the answer depends heavily on how the human and machine tasks are defined.

Renewable energy claims need care too. A company buying clean power does not necessarily mean the data center runs on clean power every hour, and some experts argue that clean supply claimed by one customer can be replaced with fossil power for others. This is a debate about accounting methods, and it is not settled.

What is being done about it?

More efficient chips, smaller models distilled from larger ones, better cooling and software that skips unnecessary computation all lower the energy per answer. Google’s reported 33-fold efficiency gain over a year, if it holds up, shows how fast the per-query number can move.

There is an old economic wrinkle worth knowing, though. When something gets cheaper to use, people tend to use more of it, and total consumption can rise even as efficiency improves. This pattern is sometimes called the Jevons effect. Whether it applies to AI is a hypothesis, not an established result, but the projections above suggest that demand growth is outrunning efficiency so far.

Common questions

Does using ChatGPT or a similar chatbot hurt the environment?

A single query has a very small footprint by the best available estimates, so one person’s use is not a big lever. The concern is the total across hundreds of millions of users, and the data centers built to serve them. Your habits matter far less than how the electricity is generated.

Is an AI image more costly than a text reply?

Yes, by a wide margin. In a 2024 measurement, image generation averaged about 2.91 watt-hours against roughly 0.05 for text generation, more than fifty times as much. Exact multiples depend on the models compared.

Will AI cause blackouts?

There is no established finding that it will. Grid operators and the IEA describe a real planning challenge, since data centers cluster in particular regions and take years to connect, but whether supply keeps pace is a regional question and not one with a single answer.

Why don’t the companies just publish exact numbers?

Some now publish per-prompt averages, as noted above. Detailed training and facility figures are mostly withheld, and the reasons given or implied include competition and the difficulty of attributing shared infrastructure. I have no source stating a single reason, so that part is inference.

How big is AI compared with the rest of the data center load?

No clean split exists in the sources I read. Data centers together used about 4.4 percent of US electricity in 2023, and AI is one portion of that, a portion the IEA expects to grow fastest. Anyone quoting a precise AI-only share is working from assumptions.

What to remember

Per query, AI uses little energy; at scale, it uses a lot, and the trend points upward. The most solid numbers come from the US Department of Energy’s laboratory report and the International Energy Agency, and both frame the growth as large but uncertain. Per-prompt figures from the companies are useful but self-reported.

When you see a claim about AI’s footprint, ask four things: which model, which task, what is included, and who measured it. That filters out most of the noise.

Further reading

A widely read book that covers the material and environmental costs behind AI systems. See it on Amazon (affiliate link)

As an Amazon Associate I earn from qualifying purchases. A commission may be earned if you buy through this link, at no extra cost to you.

Sources