layout-1

layout-1

‘Thirsty Machines: AI on Tap’

Background information

The hidden impact of AI on water consumption

Invisible background activity

At first glance, the connection between artificial intelligence and water consumption is not obvious. However, digital devices like smartphones and laptops run an increasing number of applications in the background that use artificial intelligence, and these continuously consume energy. The appearance of modern devices is so clean and efficient that we scarcely suspect the damage they are causing to the environment.

Sophisticated marketing strategies of large corporations lie behind the design of digital devices; these encourage us to use our devices as often as possible and to keep on buying the latest models. However, the production and use of assistive technology not only cause significant CO2 emissions; they also consume large quantities of water and other valuable resources. The elegant appearance of modern devices contradicts their ecological impact.

Credits: From the exhibition “Thirsty Machines: AI on Tap” – Mapping water scarcity and data centres in a 3D model, Carla Streckwall

How and where AI consumes water

How does AI use water?

In 2024, researchers from the University of Illinois Urbana-Champaign identified three key factors that result in AI’s water consumption:

1 Cooling (direct consumption):

Many AI models rely on data centres and these use water to cool their servers, which results in evaporation. Depending on outside temperatures and operating modes, between one and nine litres of water per kilowatt hour (kWh) evaporate due to power consumption. A medium-sized data centre that consumes around 24,000 kWh of electricity per day can lose between 24,000 and 216,000 litres of water per day in the form of steam emissions.

2 By generating electricity (indirect consumption):

Many AI models rely on data centres and these use water to cool their servers, which results in evaporation. Depending on outside temperatures and operating modes, between one and nine litres of water per kilowatt hour (kWh) evaporate due to power consumption. A medium-sized data centre that consumes around 24,000 kWh of electricity per day can lose between 24,000 and 216,000 litres of water per day in the form of steam emissions.

3 In the manufacture of AI technology:

The production of microchips requires water to cool machinery and clean components. Approximately 8 to 10 litres of water are required to produce a single chip. A smartphone contains around 10 to 20 microchips, which corresponds to a water consumption of around 90 to 180 litres of water for its chip production.

In which world regions does AI consume water?

Most data centres are currently located in regions with a stable technical infrastructure, such as North America (USA and Canada), Europe (Ireland, Germany and Scandinavia) and Asia (Singapore, Japan and China). However, with the growing demand for cloud services, new data centres are increasingly being built in emerging economic markets. These are primarily located in African countries (particularly South Africa), India and the Middle East (e.g. UAE, and Saudi Arabia).

In 2024, researchers from the University of Amsterdam noted that new data centres have recently been built even in areas with water shortages, such as the Sahel region and Uruguay. One example is Google’s planned data centre in Uruguay, which has a predicted water consumption of 7.6 million litres per day, even though the country has just experienced one of the worst droughts in 74 years.

The consequences of increasing water consumption due to AI

Water scarcity

Water is the source of all life forms and an indispensable resource. The increasing scarcity of water in many regions of the world has already led to social and political conflicts. In armed conflict, control over water resources is used as a strategic tool. The relocation of data centres to water-scarce regions intensifies these conflicts and presents the poorest countries in the world in particular with a serious dilemma: they must strike a balance between the economic benefits of international investment and increasing pressure on their already scarce water resources.

Water guzzlers

Large AI models such as GPT-3 and GPT-4 are real water guzzlers. For their training, such as improving data sets and pattern recognition, modern data centres are run at full speed and require constant cooling. Vast amounts of water are needed to do this – around 700,000 litres per day. This corresponds to the water requirements of around 10,000 people on a hot summer day in a temperate region. These huge quantities of water consumption often go unnoticed, even though they represent significant damage to the environment.

Zombie servers

Large AI models such as GPT-3 and GPT-4 are real water guzzlers. For their training, such as improving data sets and pattern recognition, modern data centres are run at full speed and require constant cooling. Vast amounts of water are needed to do this – around 700,000 litres per day. This corresponds to the water requirements of around 10,000 people on a hot summer day in a temperate region. These huge quantities of water consumption often go unnoticed, even though they represent significant damage to the environment.

Data centres worldwide

It is estimated that there are currently almost 11,000 data centres worldwide, with the majority of them located in the United States, Germany, Great Britain and China. The market for data centres is constantly growing. According to forecasts, revenue in the data centre market will reach around 412.3 billion euros by 2025, with an expected annual growth rate of 8.35 per cent by 2029.

The global dimension: predicted water demand

According to estimates, AI applications could use up to 6.6 billion cubic meters of water per year by 2027. If this development continues unabated, conflicts are likely to arise in water-scarce regions, as protests against new data centres in the USA have already shown.

Examples and everyday effects

The example of GPT: everyday water consumption

Scientists from the University of California have published a study showing that during training in Iowa, GPT-3 required around 4.88 million litres of water over a period of one to two months. By comparison, this volume covers the daily drinking water requirements of around 65,000 people. But using AI also consumes water: a single GPT-3 query uses around 15 millilitres. According to this calculation, 33 queries result in the consumption of a half-litre bottle of water.

Water consumption in Germany

As a simple example, the average German household uses around 120,000 litres of water annually for showering, cooking, washing and gardening. If we base calculations on the number of Google data centres worldwide (of which there are 23, according to the company’s statements), it consumed 21 billion litres of water in total in 2022: this corresponds to the annual water consumption of more than 166,000 German households.

Why we need to act

Fresh water is a limited resource that is becoming increasingly scarce worldwide. Two-thirds of the world’s population suffers from water shortages for at least one month a year. Increasing water consumption due to AI applications, ranging from large language models to other data-intensive technologies, is significantly exacerbating this problem. Given this development, it is clear that AI must not only reduce its CO2 emissions and energy consumption but also its water footprint.

Environmental consequences: more than just water shortages

Warming of lakes and rivers

When water heated during cooling processes in data centres is released back into rivers or oceans, water temperatures rise. This can have dramatic effects on ecosystems, particularly in sensitive habitats such as coral reefs. Corals are extremely sensitive to temperature: even slight increases can lead to coral bleaching, a condition in which corals lose their symbiotic algae and die. Other marine organisms, such as fish, algae and mussels, also depend on stable water temperatures. When water is warmed, this can jeopardise their reproduction, growth and survival, and lead to an imbalance in the marine ecosystem.

Destruction of habitats

Building data centres near the coast can have significant effects on natural habitats. Valuable land is often claimed for construction, which leads to the destruction of coastal and river landscapes. Especially in coastal regions, where wetlands and mangrove forests serve as protection for biodiversity, the construction of data centres can destroy these ecosystems. The conversion of land for data centres not only results in lost habitat; it can also make coastal regions vulnerable to extreme weather events. The effects on local flora and fauna are often irreversible and lead to a decline in biodiversity.

Sustainable solutions and responsibility

Sustainable AI — utopia or reality?

Tech giants such as Google and Microsoft are advertising ambitious goals, such as a ‘water positive’ future by 2030. Measures such as time-flexible training – i.e. training AI models during low network utilisation or widespread availability of renewable energy – and the use of more efficient data centres are intended to reduce the water consumption of AI models. But because of the enormous waste of resources, the question is whether this is enough – or whether operating large-scale AI models in itself presents an insoluble problem. Transparency is essential here. Companies must be required to disclose actual water consumption to take responsibility and facilitate credible steps toward sustainability.

Legal control – AI Act

The Heinrich Böll Foundation’s Wasseratlas 2025 (Water Atlas 2025) points out that the EU passed the AI Act in 2024 – the world’s first law to regulate AI. The law stipulates documentation requirements for energy consumption and computing resources when training AI models – but not for water consumption, as it only regulates AI products and not the technical infrastructure. The EU Energy Efficiency Directive imposes reporting requirements on data centres regarding water consumption, thus increasing transparency in Europe. However, higher investments are needed worldwide to reduce water consumption, for example through alternative cooling systems.

Water and AI in Germany

In Germany, too, the consequences of water shortages are increasingly being felt. Despite high (and increasing) rainfall, Germany has been losing an average of 2.5 billion cubic metres of water per year for around 20 years. The main reasons for this are, on the one hand, changes in precipitation patterns due to climate change, with more frequent heavy rainfall but also longer dry periods. On the other hand, roads and large sealed areas in cities prevent rain from seeping into and being stored in the ground; instead, it flows into the sewage system.

Water prices and local resources

The increasing impact of modern technologies on the environment may so far be barely noticeable in many regions of Germany; but as users and consumers, our decisions actively contribute to this impact. In cities such as Berlin, people are directly affected when this impact results in potential water scarcity and, as a result, rising water prices.

Individual contribution to the water footprint

Our personal energy consumption also plays a role in the overall analysis. Anyone who regularly uses AI and cloud services is thus directly contributing to increasing resource consumption. It makes sense to critically question your own digital footprint and consciously consider more sustainable alternatives or more efficient use.

Responsibility in the digital age

As consumers, we have the opportunity to make digital decisions on a daily basis that not only impact our own lives but also the world around us. By choosing sustainable products and taking a critical look at the companies whose services we use, we can reduce our environmental footprint and contribute to a more efficient use of water and energy resources.

A critical view

Although AI optimists believe that AI can help solve global water problems, a key question remains: are the environmental and social benefits of AI offset by its enormous water consumption? It must be made public that AI models use significantly more water than conventional digital processes. While a Google search requires 0.5 millilitres of water, ChatGPT uses around 15 millilitres. Or to put it another way: the water consumption of thirty Google queries is the same as a single ChatGPT query.

AI versus alternative resources

Many resources that use less water can be used to find answers to questions. It is important to place the use of AI in the context of existing analogue knowledge resources and alternative digital solutions. When does the use of AI really make sense, and how can it be used in a sustainable, resource-saving way?

Possible alternatives

Resource-saving knowledge bases include analogue sources such as books, libraries or direct exchange with experts. Energy-efficient solutions can be used digitally, such as targeted search engine enquiries, offline databases or optimised software. More sustainable use of AI can be achieved through time-flexible training, green data centres and conscious application. Decentralized approaches such as using local AI models or federated learning also reduce the water consumption of data centres.