Skip to content
Deerdate

2026-11-03 · 6 min read

How much energy does AI use? What we know so far

AI uses a real but still modest share of global electricity. Estimates differ a lot, so here is a calm look at what the best sources say and what stays unclear.

AI Energy Climate

AI uses a growing amount of electricity, but it is still a small part of the world total. Most of that power runs data centers, and the exact share for AI alone is hard to measure. Good sources give ranges, not one clean number.

This article shares what we know, what we do not know, and why the numbers keep changing.

Start with data centers

Most studies do not measure AI alone. They measure data centers. A data center is a building full of computers. These computers run search, video, email, cloud storage, and AI.

AI is one part of that load. It is a fast-growing part.

The global picture

The International Energy Agency (IEA) published a report called Energy and AI in 2025. It estimated that data centers used about 415 terawatt-hours (TWh) of electricity in 2024. That is about 1.5 percent of global electricity use.

The IEA also gave a base-case projection. It expects data center demand to reach about 945 TWh by 2030. That is more than double the 2024 level.

Two points matter here:

The United States picture

Lawrence Berkeley National Laboratory (LBNL) published a report in December 2024. It looked at US data centers only. It estimated that they used about 176 TWh in 2023. That was about 4.4 percent of US electricity.

For 2028, LBNL gave a range, not a point. The range was 325 to 580 TWh. That is roughly 6.7 to 12 percent of US electricity.

The LBNL authors built their historical numbers from earlier studies and from shipment data for servers. They did not measure each building directly. That is a common limit in this field.

What does one AI question cost?

People often ask about the cost of one chat message. The answer depends on the model, the length of the answer, and how the data center is built.

In August 2025, Google researchers published a paper called Measuring the environmental impact of delivering AI at Google Scale. They reported that the median text prompt in Gemini Apps used about 0.24 watt-hours (Wh). They also reported about 0.26 milliliters of water.

Some context on that result:

For scale, 0.24 Wh is a very small amount of energy. It is also one company, one model, and one year. Other services may differ a lot.

Why estimates disagree

You may see very different numbers in the news. There are clear reasons for this.

  1. Companies often do not publish full data. Many firms share totals for the whole company, not for AI alone.
  2. Methods differ. Some studies count only the chip. Others count the whole building and the power plant behind it.
  3. Hardware changes fast. New chips can do more work per watt. Models also grow bigger, so savings can disappear.
  4. Use changes fast. If more people use AI for more tasks, total demand rises even when each task gets cheaper.
  5. Forecasts depend on assumptions. Small changes in growth rates create large gaps by 2030.

Because of this, treat any single big number with care. Look for the source, the year, and the method.

Training and daily use

AI energy has two main parts.

Training

Training is the work of building a model. It runs for days or weeks on many chips. It happens once per model version, though companies train many versions.

Inference

Inference is the daily use of a model. Every question you ask is inference. A model that millions of people use all day can add up to more energy than its training did. Researchers still debate how the split looks across the industry.

Where the power comes from

Energy use is only half of the climate story. The other half is the type of power. A data center on a grid with much wind, solar, or hydro power causes fewer emissions than one on a coal-heavy grid. The IEA report discusses both sides: how much power data centers need and how the supply mix could change.

This is why two models with the same energy use can have different carbon footprints.

A calm way to think about it

Here is a fair summary:

You do not need to feel guilt about asking a question. It helps to know the topic, and to choose well when a simpler tool does the job.

Frequently asked questions

Is AI the main reason electricity demand is rising?

No. Data centers are one source of growth among several. The IEA reports that data centers make up a fairly small share of global demand growth. In some countries, such as the US, their share of growth is larger.

Does one chatbot question use more energy than a web search?

It can, but the gap depends on the system. Published numbers vary by company and method. We did not find one trusted, neutral comparison, so we do not give a ratio here.

Are the Google numbers typical for all AI tools?

Not necessarily. The 0.24 Wh figure is for the median text prompt in one Google product in 2025. Image, video, and long reasoning tasks can use more. Other companies have not all published similar data.

Will AI get more efficient?

Often yes. Chips and software improve each year, and Google reported large drops in energy per prompt over one year. But total use may still rise if people use AI more often. Both trends can be true at once.

Sources