How much water does artificial intelligence consume – and how to interpret the numbers?

AI consumes water both in the cooling of data centers and, indirectly, in the generation of the electricity that powers them. However, this doesn't mean that each question sent to a chatbot corresponds to a fixed amount of water. The footprint depends on the processing performed and the infrastructure that supports it, as well as what each study decides to include in the calculation.

Understanding this difference is important for scaling the problem. A high number, presented without context, might suggest that every interaction with AI has an extraordinary impact. A small number, on the other hand, might hide the cumulative effect of a large-scale operation. Neither of these readings, in isolation, helps decide where to reduce consumption, which facilities require more attention, or which improvements actually work.

The question is not whether to consider AI harmless or environmentally unacceptable. It is to identify how much it consumes, under what conditions, and with what consequences. To do this, we need to separate concepts that often appear mixed together.

Why do data centers use water?

Data center equipment transforms electricity into processing power and heat. To maintain operation within appropriate temperatures, this heat must be removed. Some facilities use cooling towers, in which some of the water evaporates during heat exchange; others combine mechanical cooling, use of outside air, and different fluid circulation circuits. The U.S. Department of Energy describes these alternatives and emphasizes that their performance depends, among other factors, on the climate and the facility's configuration.

Therefore, the presence of water in a system does not, by itself, indicate how much it consumes. One circuit may recirculate the same fluid, while another stage of the installation uses evaporation to reject heat. Similarly, "liquid cooling" should not be automatically interpreted as high consumption or zero consumption: it is necessary to understand the complete system.

There is also water associated with electricity generation. Depending on the power plants that supply the grid and their technologies, providing electricity may involve water consumption outside the data center. A more comprehensive assessment may also consider equipment manufacturing and other processes in the production chain. These analytical limits need to be explicit, as an operational estimate is not equivalent to a full lifecycle assessment.

Removing water is not the same as consuming water.

THE withdrawal, Removing water, also called intake water, is the water removed from a source for a specific activity. consumption This corresponds to the portion that is no longer available for immediate reuse, for example because it has evaporated or been incorporated into a product. USGS, The United States Geological Survey distinguishes these concepts precisely because they describe different relationships with water resources.

This means that two facilities can collect similar volumes and consume different amounts. It also means that a comparison between the collection of one sector and the consumption of another does not measure the same thing, even if both numbers are expressed in liters.

Evaporated water doesn't disappear from the planet. The problem is its availability in the place and time when other people, activities, and ecosystems need it. This is one reason not to treat the water footprint simply as a global volume calculation.

What do the data center numbers show?

THE 2024 report from Lawrence Berkeley National Laboratory It was estimated that data centers in the United States directly consumed approximately 66 billion liters of water in 2023. The document also calculated an indirect water footprint close to 800 billion liters, associated with the electricity used. Indirect values were estimated from regional power grids, not measured individually at each power plant for each installation.

Indicator	I estimated	Scope
Direct consumption of water	66 billion liters	US data centers, in 2023
Indirect consumption associated with electricity	Approximately 800 billion liters	US data centers, in 2023
Electricity used	176 TWh	US data centers, in 2023

This data refers to to the data centers as a whole, not exclusively to AI.. The infrastructure also supports storage, enterprise applications, and other digital services. Therefore, presenting the totals as "artificial intelligence consumption" would attribute activities to the technology that the survey does not separate in this way.

The numbers are useful for understanding the order of magnitude of the infrastructure and the importance of electricity in the water footprint. However, they do not provide a direct answer as to how much a person consumes when using a particular chatbot.

Demand is growing, but projections are not facts.

In an update published in April 2026, the International Energy Agency estimated that data centers consumed 485 TWh of electricity in 2025. Its central projection points approximately 950 TWh in 2030, close to 3% of global electricity demand that year. The growth of AI-focused installations is faster than that of the sector as a whole.

This justifies monitoring the expansion, efficiency, and availability of infrastructure. However, the projection should not be interpreted as an inevitable outcome: it depends on the adoption of applications, investments, technological improvements, and obstacles to the construction of new projects.

It is also incorrect to automatically convert electrical growth into hydrological growth in the same proportion. The amount of water depends on the combination of electricity sources, cooling, and location. An expansion may be accompanied by a reduction in hydrological intensity; the absolute result needs to be observed, not assumed.

Why isn't there a universal value for water per question?

A short question, an extensive document analysis, and video generation represent different services. Even within text generation, response length, template, equipment usage, and infrastructure alter the amount of resources required.

A study published in 2025 by researchers affiliated with Berkeley Lab found variations of over 10,000 times in water consumption per workload among the configurations analyzed. The result combines differences in computational efficiency and water intensity of the infrastructure. This doesn't mean that switching chatbots will produce savings of that magnitude., This does not constitute a direct comparison of all commercial AI services.

The main contribution of this study is to show why the location and method of performing a task matter so much. The number per interaction can help compare configurations when the methodology is consistent, but it loses meaning when it becomes a supposedly valid constant for any model and any use.

What does the estimate released by Google mean?

In an analysis from May 2025, Google estimated 0.26 milliliters of water for the median text prompt in Gemini Apps, According to their methodology, the water calculation considers the water consumed to cool equipment and infrastructure in data centers; it does not represent all the water associated with electricity generation and the manufacturing chain.

The value was derived from energy consumption and water efficiency factors of the infrastructure. Therefore, it does not correspond to a water meter registering a specific amount with each message. The company also stated that the data and claims had not been verified by an independent third party when they were released.

There is another distinction: median is not mean.. A representative measure of a typical interaction does not, by itself, allow for the calculation of total consumption by multiplying it by the number of messages. Highly intensive interactions may have a significant weight in the total without appearing in the same way in the median.

None of this makes the study useless. It makes it necessary to interpret it within its scope. The same care must be applied to larger estimates: comparing results requires verifying what was included, what tasks were performed, and how resources were allocated to them.

Training and everyday use are different stages.

Training is the stage in which the model learns patterns; inference is its execution to respond or perform a task. Studies on the water footprint of AI analyze these stages separately because their demands and frequency are distinct. An estimate for training an entire model cannot be directly compared to the consumption of a single response.

It's also not enough to simply choose the step that produces the most convenient number. Evaluating a service throughout its lifecycle requires considering its development and cumulative usage, with explicit allocation rules. The greater the adoption, the greater the importance of understanding the operation, and not just the initial training cost.

This distinction avoids a misleading comparison: placing large-scale training on one side and individual interaction on the other to conclude that the technology has become hundreds of times cleaner. In this case, the difference may lie primarily in the units being compared.

Comparing it to agriculture helps – provided the comparison is accurate.

According to the FAO, agriculture accounted for approximately 72% of global freshwater withdrawals in 2023. This is an important point: it places the discussion about AI within a much broader water problem and prevents the attention given to a recent technology from causing sectors with a larger share in water use to disappear.

Proportionality matters when setting priorities. Resources for monitoring, research, and environmental improvement are limited, and their allocation should not depend solely on the visibility of a particular issue. At the same time, lower overall participation does not resolve the analysis of a specific facility.

The methodological approach is simple: the FAO percentage deals with global withdrawals, while the American survey of data centers includes direct and indirect domestic consumption. These data do not allow us to calculate how many times one sector would be "worse" than the other.

It is also necessary to consider the purpose and the alternatives. Producing food, manufacturing textiles, and performing computing deliver different services. A liter provides information about volume; it does not, by itself, answer how much value was produced, what waste could be avoided, or which option would meet the same need with less impact.

And what about the textile industry?

The textile industry is relevant for comparison, but its supply chain encompasses very different activities. A UNEP analysis identified the cultivation of natural fibers, such as cotton, and the consumer use phase as key points for the impacts of water scarcity. Other impacts are concentrated in stages such as fiber production and fabric processing.

This shows why it's not appropriate to treat an estimate of the entire supply chain for a garment as equivalent to the water used solely for cooling a server. In one case, it may include cultivation, manufacturing, and use; in the other, only an operational step.

There is also overlap: cotton cultivation is part of agriculture and can be included in the footprint of a textile product. Adding the two categories together as if they were entirely independent would result in double counting.

Sectoral comparisons remain useful, but they must disclose their limitations. A realistic view does not need to deny that other activities use much larger volumes; it needs to avoid transforming this finding into a numerical equivalence that the data do not support.

The impact depends on the location, not just the volume.

Water is a locally shared resource. GRI 303, a reference for disseminating information on water and wastewater, guides the analysis of availability, quality, access, and basin context, including relationships with suppliers and communities. These factors help explain why the same volume can have different consequences in different regions.

This doesn't mean we should assume that every data center directly competes with domestic water supply. This relationship needs to be investigated: where does the water come from, what percentage is potable, what other uses exist, and how does demand behave during dry periods?

Similarly, establishing an operation in a seemingly abundant region does not eliminate the need to assess its capacity to meet demand. What matters is the combination of demand, infrastructure, environmental conditions, and management rules, not just a label of "region with water.".

Which solutions make sense?

There are improvements in both processing and infrastructure. Reducing unnecessary computing can decrease resource demand; improving thermal management, taking advantage of favorable weather conditions, and adjusting system operation can reduce energy and water consumption. The Department of Energy's technical guidance presents such alternatives, but emphasizes the dependence on the conditions of each facility.

There are also projects that prevent water evaporation for cooling purposes. However, this does not authorize them to be called facilities with a completely zero water footprint. Microsoft, when presenting one of these projects, He even acknowledged the possibility of increased energy consumption compared to certain evaporative systems.

The appropriate choice may vary. In a pressured basin, saving water may justify a certain increase in electricity; in another situation, a different combination may produce a better result. The goal is to evaluate the combined consequences, not to maximize one indicator while ignoring the others.

How to evaluate a claim about AI's water consumption

Before accepting a number, it's worth identifying the activity, the unit, the year, the geography, and the account limits. Is it consumption or withdrawal? Only refrigeration or also electricity? A specific task or all the services of an installation? An operational measurement, an estimate, or a project not yet executed?

It is also useful to ask what the comparison is intended to demonstrate. A figure may be appropriate for discussing efficiency but inadequate for assessing pressure on a particular basin. An overall total may help to size a sector but say very little about the conditions of a city.

Good environmental information is not about having the highest or lowest numbers. It's about what allows you to understand a decision.

Frequently Asked Questions

Does the water consumed by AI disappear?

No. “Consumption,” in this context, does not mean the destruction of water, but rather its unavailability for immediate reuse, as occurs with evaporation. The concern lies in local and temporal availability, not in the disappearance of water from the planet.

Is it correct to say that one question uses up a bottle of water?

Not as a universal rule. Any equivalence of this type needs to specify the model, the task, the period, and the components considered. Without these conditions, a specific estimate ends up inaccurately representing the entire technology.

Does the fact that agriculture uses more water make the impact of AI irrelevant?

No. Comparison helps to prioritize, but it doesn't replace the analysis of each facility. An activity may have a smaller overall impact and still require attention in a specific location.

Artificial intelligence's water consumption

AI uses physical resources, but its water footprint is not a fixed characteristic for every question. It depends on the task, the infrastructure, and the limits of the measurement. Recognizing this variability allows us to question both alarmist generalizations and overly reassuring claims.

The most useful response is to track absolute volumes, improve efficiency, and examine local consequences. This doesn't require assigning blame to every interaction or accepting any expansion as inevitable. It requires treating computing with the same criteria applied to other activities: understanding the benefit, recognizing the costs, and seeking technically feasible alternatives to reduce impacts.

To delve deeper into the relationship between water use, energy, and model choice, also read the article about sustainable AI and efficiency.

Sources and limitations of the analysis

This text uses the Berkeley Lab 2024 report, the research on water consumption by workload, definitions of USGS, projections of IEA and technical information from suppliers. Historical estimates, projections, and corporate data were identified according to their nature. No proprietary measurements were performed, and the results cited do not allow for establishing a universal value for water through interaction, nor a homogeneous comparison between all AI, agriculture, and the textile industry.