The rapid advancement of artificial intelligence has a physical cost that is beginning to be felt outside of screens: electricity and water. More and more governments and cities are taking measures to curb or place conditions on the construction of data centers, while international organizations warn of a growth in energy consumption that will multiply in a few years.
The debate is no longer just technical. It affects electricity bills, water availability in some regions, and the urban planning of cities that compete to attract or, in some cases, reject new facilities.
How much electricity do AI-dedicated data centers consume
According to the Energy and AI report by the International Energy Agency (IEA), data centers consumed around 415 terawatt-hours (TWh) in 2024, which is equivalent to about 1.5% of all electricity generated in the world that year.
The geographical distribution is very uneven. The United States accounts for 45% of that consumption, China for 25%, and Europe for 15%, according to the same IEA data.
Growth is the most striking figure. Since 2017, the electricity consumption of data centers has increased by 12% each year, a rate more than four times higher than that of global electricity consumption as a whole.
With this trend, the IEA projects that consumption will reach about 945 TWh by 2030, a figure higher than the entire current electricity consumption of Japan. For 2035, the agency's base scenario places the figure at around 1,200 TWh, although it acknowledges a wide margin of uncertainty: between 700 and 1,700 TWh depending on the scenario that unfolds.
The IEA itself compares the consumption of a large AI-focused data center with that of an aluminum smelter, but with a key difference: these facilities are much more geographically concentrated, which makes their impact felt strongly in very specific areas.
How much a single AI query costs
Google has published its own figures on the cost of a text query in Gemini. According to the company, a "medium" query consumes 0.24 watt-hours of energy, generates 0.03 grams of CO₂ equivalent, and uses 0.26 milliliters of water, the equivalent of about five drops.
The company also claims that, in the last twelve months, the energy required per query has been reduced 33 times and the associated carbon footprint 44 times.
It is worth qualifying these data. They come from the company itself and have not been verified by an independent body. Furthermore, they refer to a medium text query and daily system use, not the model training process, which has a different energy cost. Efficiency is improving rapidly, but the total number of queries is growing even faster.
Cities and governments that are curbing new data centers
Several territories have begun to impose limits or conditions on the construction of new facilities, according to a Reuters review of restrictions in different countries.
United States
In New York, a one-year moratorium was approved for data centers of 50 megawatts or more. Pennsylvania, for its part, requires environmental safeguards, transparency, and local approval for these types of projects.
In Texas, Governor Greg Abbott requested an audit in August on the impact of data centers and ordered a halt to the connection of new facilities to the power grid. On September 21, he expanded the measure by also stopping environmental permits until that audit is concluded, as reported by Texas Standard. Medium-sized centers, between 25 and 75 megawatts, are excluded from the grid connection pause.
Europe
Amsterdam has banned the construction of new data centers or the expansion of existing ones until at least 2030.
Ireland lifted the blockade it maintained around Dublin in December 2025, but now requires that every new electrical connection be accompanied by its own generation capacity.
Ways to reduce consumption without slowing down AI
There are several ways to reduce the energy and water expenditure associated with artificial intelligence. Some are already applied in practice; others are proposals under study.
- Use the smallest model for each task. Not all queries require the most powerful model available. Directing each task to the simplest model capable of solving it saves energy and costs.
- Run AI directly on the device. Models that run on your own phone or laptop do not need to rely on a data center.
- Leverage excess capacity. The IEA points out that many data centers operate with underutilized servers and proposes that regulators study incentives to leverage that capacity with greater flexibility.
- Reuse generated heat. Heat from servers can be used for heating in nearby areas instead of being wasted.
- Schedule non-urgent tasks during cheap energy hours. Train models or process large volumes of data when there is a surplus of renewable energy available.
A proposal under debate: sharing idle computing power
There is an idea, still in the discussion phase and not a confirmed fact, that proposes taking advantage of the millions of computers and servers that remain inactive for a good part of the day. A network in which their owners would rent out that computing time for non-urgent AI tasks could reduce the need to build new infrastructure.
There are already platforms with a similar philosophy. Akash presents itself as a decentralized computing marketplace, Render offers capacity from idle graphics cards in different parts of the world, and io.net follows a similar approach. Their real reach is yet to be proven.
A variation of this idea proposes paying for that computing time with AI usage credits instead of money, as a form of barter. The advantage would be to avoid complex payments and facilitate the participation of those who only seek to use AI for free. Among the difficulties would be verifying that the work was performed correctly, protecting the data that travels to third-party machines, and preventing that credit from becoming a speculative asset.
What this implies for the user
- An individual query to an AI consumes relatively little, but the real impact arises from adding up billions of daily queries.
- Choosing the smallest model or a version that runs on the device itself, when sufficient, is the simplest way to reduce consumption.
- Those who live near a data center project can request detailed written information from the promoting company about its expected water and electricity consumption.
Frequently Asked Questions
How much electricity do AI data centers consume today?
According to the International Energy Agency, in 2024 they consumed about 415 terawatt-hours, about 1.5% of all electricity generated in the world.
How much will that consumption grow in the coming years?
The IEA projects about 945 TWh in 2030 and around 1,200 TWh in 2035 in its base scenario, although the uncertainty range for 2035 goes from 700 to 1,700 TWh.
How much water does an AI query spend?
According to Google data, a medium text query in Gemini consumes about 0.26 milliliters of water, roughly the equivalent of five drops.
Which cities have limited the construction of data centers?
New York, Pennsylvania, and Texas in the United States, as well as Amsterdam in Europe, have imposed moratoriums, audits, or specific environmental conditions. Ireland lifted restrictions in Dublin but requires its own electrical generation for new connections.
How can AI energy consumption be reduced without stopping its use?
Among the available ways are using smaller models for simple tasks, running AI on the device itself, leveraging underutilized servers, reusing generated heat, and scheduling non-urgent tasks during cheaper energy hours.
What is the so-called Airbnb of computing power?
It is a proposal under discussion that suggests that owners of inactive computers and servers rent out that computing time for non-urgent AI tasks, thus reducing the need to build new data centers. It is not a confirmed fact, but an idea under debate.
