A River Runs Through the Cloud: The Hidden Water Consumption of Swiss Data Centers

This text was translated from German by the human translators of Toptranslation.

Right behind the shimmering screens of our digital devices lie data centers of enormous scale. In their endless corridors, energy vibrates, data traffic hums, and cooling systems steam. This is where the daily flood of emails, WhatsApp messages, Instagram videos, cat memes, and deepfake videos has found its physical home. Here, in the starkly lit, high-security confines of data centers, thousands of terabytes are stored, processed, and transmitted by an army of servers running hot. Packed tightly together and stacked in racks weighing tons, these machines keep our digital lives running around the clock.

Data centers have experienced unprecedented growth in recent years. As the infrastructural backbone of the ever-swelling floods of data generated by social media platforms, streaming services, office applications, and AI tools, the industry is expanding rapidly. Already more than 11,000 data centers have been set up worldwide, mostly in the United States, China, Germany, and the United Kingdom. In recent years, however, Switzerland has also emerged as an attractive location. While no official figures reveal the precise number and location of domestic data centers, analyses indicate that there must be at least 120. This gives Switzerland one of the highest numbers of data centers per capita in the world.

Image credit: Christoph Miller (CC BY 4.0)

Driven by the global race for powerful digital services, new facilities have been springing up across Switzerland at high speed for around a decade, while existing data centers have been expanded. Many are now multi-story buildings the size of soccer fields, with some grouped together to form entire “campus complexes.” They benefit from Switzerland’s central location in Europe, proximity to the financial centers of Zurich, Geneva, and Frankfurt, and the country’s well-developed fiber-optic and energy infrastructure. The operators are primarily large American companies that build what are known as colocation data centers. Rather than using the buildings themselves, the owners lease space equipped with the necessary power and cooling infrastructure to third parties—increasingly including major cloud providers such as Microsoft, Google, and Amazon Web Services, which operate their powerful swarms of servers in large-scale rooms and cages. Exactly who rents space in individual buildings, however, remains a closely guarded industry secret. One thing, however, is undisputed: operating the oversized gray blocks consumes a formidable amount of energy.

Image credit: Christoph Miller (CC BY 4.0)

According to a recent study by the Swiss Federal Office of Energy (SFOE), electricity consumption by Swiss data centers accounted for 3.6 percent of overall national consumption in 2024. By 2030, it could rise to as much as six percent. That would be more electricity than Europe’s largest dam—the 285-meter-high Grande Dixence Dam in Valais—generates in a year. However, the reliability of the energy consumption figures published by the SFOE are difficult to judge since they are based on a survey in which many operators refused to provide data or were uncertain about their future clients and capacities. Moreover, considerably higher figures emerged last year, when it was reported that data centers could account for as much as 15 percent of Switzerland’s electricity consumption by 2030. It remains to be seen, if the country's electricity infrastructure can be expanded fast enough to keep pace with such growth rates, and if the expansion will affect domestic electricity prices.

The Silent Thirst of Servers

Although the rising electricity consumption of Swiss data centers is attracting increasing public attention, it tells only half the story of their resource needs. What still often goes unnoticed is the considerable thirst of servers running hot. Detailed data on the water use of specific data centers, however, is difficult to find—both in Switzerland and internationally. Only rarely does the consumption of individual facilities become public, usually as a result of investigative reporting or court proceedings. As is the case, for example, at the Fayetteville Campus—a data center in Georgia that drew over 130 million liters of water from the public water supply in just a few months. That is enough water to fill 44 Olympic-size swimming pools. The facility’s high consumption came to light after an investigation was launched when residents noticed unusually low water pressure at their taps.

Such high water consumption rates are not unusual for data centers. The precious resource is essential not only for cooling servers, but also for generating the enormous amounts of electricity they consume. A study by the University of California, Riverside concludes that by 2027 data centers worldwide could require an almost unimaginable 6.6 trillion liters of water. That is four to six times Denmark’s annual water demand. When it comes to visualizing such immense quantities of water, even a conversion into Olympic-size swimming pools is of little help—there would be more than 2.5 million of them.

The high water consumption of data centers is particularly problematic in the context of the growing water crisis. In the future, large parts of the planet will suffer prolonged droughts and water shortages caused by the climate crisis. Even Switzerland, commonly regarded as Europe’s “water tower” and seemingly endowed with almost limitless water reserves, will not be spared from water stress. According to Paul Steffen, deputy director of the Federal Office for the Environment, “the risk of drought in Switzerland will increase in every season.” In some parts of the country, such as Ticino and Aargau, water for households and agriculture already has to be rationed at times because local water supplies reach their limits, especially in summer. For this reason, a nationwide “early detection and warning system for drought” has been under development since 2022, under the responsibility of the federal offices for the environment, meteorology, and topography. The data collected is intended to help authorities take measures “to protect our water resources and reduce water scarcity,” says Steffen.

Against this backdrop, it seems questionable that Switzerland’s booming data center industry is subject to no requirements whatsoever regarding water use. On the contrary, operators are neither obliged to disclose their water use nor subject to any specific rules governing their approach to water utilization. Therefore, they have no obligation to comprehensively track their consumption, use water efficiently, or make their consumption public. The industry itself also remains silent on this sensitive issue. All the more reason, then, to attempt an analysis of water use by Swiss data centers. For, as we will see, this “blue gold” is essential fuel for Switzerland’s data traffic—fuel, which is becoming scarcer in an age of shrinking glaciers, drying riverbeds, and falling groundwater levels.

But first things first. Anyone trying to determine how much water a data center consume needs to track down both its “direct” and “indirect” consumption. Direct consumption refers to the amount of water consumed within the data center itself, primarily for cooling the servers. Indirect consumption, by contrast, seeks to capture the amount of water consumed outside the facility’s security perimeter, primarily to generate electricity.

Image credit: Christoph Miller (CC BY 4.0)

Indirect water consumption by data centers: Invisible Flows

The hidden relationship between flows of electricity and water is particularly pronounced in the Zurich region, where the processed amount of data is growing rapidly. The most visible result of this growth is the massive expansion of data centers and their rising demand for electricity. According to analyses by Tamara Keller, the region’s five largest facilities will soon consume electricity equivalent to 70 percent of the city of Zurich’s total electricity consumption. The American company Vantage, for example, is currently building a 100-megawatt data center in Volketswil that will consume as much electricity as the entire city of Winterthur. Once expanded, Green’s campus in Dielsdorf will consume seven times as much electricity as the municipality itself, while Digital Realty’s facility in Glattbrugg will soon consume the equivalent of 15 percent of the city of Zurich’s electricity demand.

What is often overlooked is that generating such vast amounts of electricity depends on the availability of enormous quantities of water, especially for cooling and energy conversion processes in power plants. Different forms of electricity generation consume different amounts of water. And although the indirect water consumption associated with electricity generation is extremely relevant to a data center’s total water consumption, it rarely features prominently in the debate. One reason is that it cannot simply be measured by a water meter that turns the flow into numbers as water rushes through a pipeline into the building.

Image credit: Christoph Miller (CC BY 4.0)

Reconstructing the indirect water consumption of Zurich’s data centers is therefore notoriously difficult. After all, doing so would require knowing exactly where the electricity for individual data centers comes from within the kaleidoscopically pulsating web of the power grid. Thus, it would be necessary to determine not only whether individual data centers obtain their electricity from nuclear, hydroelectric, or wind power, but also which specific power plants they are connected to, or to what extent they draw electricity from multiple power plants, or even whether they combine different forms of energy. Unsurprisingly, such information is nowhere to be found.

What can be found, however, are the operators’ websites. And there is plenty to see there: PR-ready images of weightless data clouds floating above glaciers, rivers, and lakes, for example. The sites tout “the highest sustainability standards” and “green energy.” Images of the buildings show glistening solar panels, and crystal-clear water streams. But data showing the water consumption of individual data centers is nowhere to be found.

An analysis of the data centers’ technical specifications, however, shows that nearly half of the providers in the Zurich area use “renewable energy.” Which forms of renewable energy these are, however, is not specified. What is clear, nonetheless, is that hydropower, as a cornerstone of European and Swiss electricity generation, must play a key role. It accounts for nearly 60 percent of electricity production in Switzerland, nearly 90 percent in the canton of Zurich, and in the city of Zurich, it exceeds 80 percent. Hydropower generation, however, is less “renewable” than it is often portrayed. Apart from its negative effects on aquatic ecosystems, significant amounts of water also evaporate at hydropower facilities. Even a reservoir just 100 meters long and 100 meters wide can lose up to 40,000 liters a day, depending on elevation and ambient temperature. In Switzerland, 200 reservoirs exceed this size. Many of them are monumental structures that stretch across the horizon, some extending over several square kilometers. In addition, there are hundreds of smaller reservoirs, dams, and balancing reservoirs. The amount of water that evaporates at Switzerland’s hydropower facilities is therefore likely to be considerable—even though it has never been systematically measured.

Image credit: Christoph Miller (CC BY 4.0)

Image credit: Christoph Miller (CC BY 4.0)

Image credit: Christoph Miller (CC BY 4.0)

Ask around in the industry and you will soon hear that some of the larger Swiss data centers are powered entirely by nuclear energy—which also requires immense quantities of water. At the recently reopened Gösgen nuclear power plant, for example, water from the Aare River is fed into the sky-high cooling tower, where up to 700 liters of water evaporate into the atmosphere every second. A little math shows what this means: if the water consumption of the nuclear power plant (which generates 8–9 TWh of electricity annually) is scaled down to a single midsize 20-megawatt data center (of which Zurich alone will soon have eleven), the result is water consumption of around 1.2 million liters a day just to supply that one data center with nuclear power. That is equivalent to the water used every day by more than 8,000 people in Switzerland.

As a preliminary conclusion, then, one thing is clear: Powering Switzerland’s expanding data centers undoubtedly requires large quantities of water.

The substantial quantities of water that quite literally vanish into thin air at nuclear and hydropower plants, however, are not included in data centers’ overall consumption figures. Many operators are not even aware of these quantities and do not ask the electricity producers about them either. And although estimates from the International Energy Agency suggest that electricity generation can account for up to two-thirds of a data center’s water consumption, and the latest U.S Data Center Energy Report describes indirect water consumption as the “true water cost of data centers,” it is not taken into account in the analyses of the federal government, the cantons, or recognized certification organizations such as the Swiss Data Center Efficiency Association. Tracking indirect water consumption is simply too great a hurdle for many operators, according to the association. Meanwhile, data centers’ electricity consumption could rise by more than half again by 2030—and with it, their hidden indirect water consumption.

Direct Water Consumption by Data Centers: The Anachronism of Air Conditioning

Water is lost not only “indirectly” in generating electricity for data centers, but also “directly” inside operators’ buildings—primarily to cool servers. Data centers are, after all, veritable strongholds of heat. Inside, hundreds of processors work ceaselessly on energy-intensive computations. The sensitive components run hot, and temperatures in the cramped server racks can quickly reach up to 75 degrees Celsius. Without cooling, the risk of overheating rises rapidly; if it gets too hot for even a moment, the entire system can collapse. That is exactly what happened during the scorching summer of 2022, when two London data centers reported “thermal events". Cooling issues and the resulting power outages left Google and Oracle cloud services in the United Kingdom unavailable for several hours. A similar incident occurred in May of this year at an Amazon Web Services data center in Northern Virginia. There, a cooling system failure caused the online services of Signal, Slack, and Snapchat to go down for hours.

Image credit: Christoph Miller (CC BY 4.0)

To prevent such costly disruptions, data centers invest in reliable cooling systems with built-in redundancy. In technical jargon, this is known as “N+1”. According to the Swiss Federal Office of Energy, many of Switzerland’s older data centers cool their servers using what are known as “chillers.” Put simply, these systems are colossal air conditioners. They blow cool air into the racks of servers running hot, carrying the heat away to the outside. This ensures stable operation at room temperatures ranging from 18 to 27 degrees Celsius. The chillers consume considerable amounts of electricity, but no water, and are also deafeningly loud.

For several years now, however, air cooling on its own has been considered largely outdated. The reason is simple: As more and more companies stream, share, post and enter AI prompts, ever-larger volumes of data are being processed in the Swiss cloud. This calls for more power in the racks. More power, however, also means more heat—which in turn drives up cooling costs. To bring those costs back down, there has been what energy expert Adrian Altenburger calls a veritable “arms race” in cooling technology in recent years.

On the one hand, operators are increasingly turning to “free cooling” to protect servers from heat stress. This approach takes advantage of Switzerland’s favorable climate and uses the cool outside air to lower the temperature in server rooms. On warmer days when outdoor temperatures climb above 20 degrees Celsius, however, this doesn’t work. For this reason, large data centers in particular now rely almost exclusively on so-called “adiabatic cooling systems.” These systems evaporate water to extract heat from the outside air, then blow this cooled air into the server rooms. The advantage: Lower electricity costs and no need for power-hungry chillers. The disadvantage: High water consumption.

Image credit: Christoph Miller (CC BY 4.0)

Standing in front of an average Swiss data center, you would not notice anything about the cooling infrastructure it requires. Only when viewed from above—from an airplane or a satellite image—does the massive cooling system reveal itself to the eye. Here, endless rows of cooling units, pipes, and fans rotate, pump, and pulse. Gray boxes, dripping mist nozzles, and chrome-plated valves, all packed tightly together—an endless cooling landscape across the rooftop: the characteristic image of a contemporary data center.

Altenburger estimates that by now half of all Swiss data centers have installed some form of adiabatic cooling system. From an economic standpoint, this makes sense because water is relatively inexpensive in Switzerland, where 1,000 liters cost around two Swiss francs. As a result, cooling a server rack using an adiabatic system costs a fraction of what it does using power-intensive chillers. Another factor is that both the Federal Office of Energy and the media and scientific community have for some time—and with good reason—been scrutinizing the rising electricity consumption of data centers. And one effective strategy for reducing a data center’s electricity consumption is to use water rather than electricity for cooling.

Image credit: Christoph Miller (CC BY 4.0)

Ultimately, a data center’s actual cooling-water consumption depends on a wide range of factors, including the local climate, cooling architecture, and server utilization. According to Energie Schweiz, however, a midsize data center in Switzerland can easily consume 100,000 liters of water per day. Yet no specific figures exist for individual facilities. Anyone optimistic enough to ask the operators directly will receive variations on the same answer: No information available, trade secret, not public. Instead, the companies assure us that they use cooling technologies that are as “smart,” “efficient,” and “advanced” as possible. As a result, only moderate quantities of drinking water are said to be consumed (yes, drinking water—it is preferred for cooling because the water needs to be pure enough to prevent hoses and pipes from becoming clogged). In some cases, that may well be true—for example, at Stollen Luzern and the Swiss National Supercomputing Centre in Lugano, where lake water is used for cooling. Or at Digital Realty in Glattbrugg and EKT in Frauenfeld, where treated rainwater is used in part as cooling water.

Nevertheless, some cases show that the water demand of individual data centers can place a heavy strain on local resources. One such case occurred in the canton of Vaud. According to unofficial reports from an industry insider, a municipality’s drinking water supply was at risk there after a data center consumed more water than expected. Operators and local authorities resolved the problem between themselves. It is not known what was agreed upon behind the scenes. The incident never became public knowledge.

The case points to a deeper problem: In Switzerland, responsibility for supplying cooling water lies with the municipalities where the data centers are located. As a rule, local utilities are required to connect them to the local water network. This means that municipal water utilities find it difficult to refuse to supply these valuable water resources to data centers. Only the actual amount consumed is open to negotiation. This negotiating process often ultimately results in supply contracts guaranteeing that specific quantities of drinking water will be provided to the international tech companies. The details, however, are not made public, and those involved generally disclose no figures. Recent reporting on the lobbying activities of major data center operators at EU level suggests that non-disclosure agreements are common. And because, according to the Zurich Department of Construction, the planned consumption volumes do not have to be specified in publicly accessible building permit applications either, residents are generally left completely in the dark about how much water is being pumped through the pipes beneath their feet toward the data center.

For similar reasons, there is some uncertainty among the approximately 5,000 residents of the municipality of Beringen in the canton of Schaffhausen, where a high-capacity data center operated by Stack Infrastructure is scheduled to go live later this year. However, the amount of cooling water originally requested was made public only after a formal inquiry by the local Social Democratic Party (SP) group: It amounted to over one million liters of drinking water a day—enough to meet the daily needs of as many as 8,000 people. Following closed-door negotiations, however, consumption is now said to have been reduced to one-eighth of the original amount. The actual figures remain highly disputed. One reason is that the response to the formal inquiry itself points out that the amount of water consumed by the data center may vary “depending on the actual local climatic conditions and any changes in operations.”

The case of Beringen in particular illustrates the urgent need for greater transparency and binding rules regarding data centers’ water consumption. The municipality is dealing not only with uncertain consumption forecasts, but, according to Switzerland’s National Centre for Climate Services, it also falls within a water scarcity zone. Here, water consumption in the summer already exceeds water supplies. Without reliable consumption figures, planning for periods of drought is likely to prove difficult.

Image credit: Christoph Miller (CC BY 4.0)

The Zurich municipality of Rafz is also located in the water-scarcity zone. Data from the cantonal statistics office shows that drinking water consumption there has increased by a quarter since 2020. Significantly, the 10-megawatt data center in Rafz went online that same year.

The Aargau municipality of Lupfig also lies in the drought zone, where a large data center operated by Green is currently being expanded. According to the technical specifications, the facility is equipped with “cooling units” and “cooling-water piping.” At the same time, the data center is just ten kilometers from the municipality of Fischbach-Göslikon, which had to restrict its water use in the summer of 2023 because groundwater reserves were running low. How much water does Green consume in Lupfig? Unknown. The company did not respond to inquiries.

Image credit: Christoph Miller (CC BY 4.0)

On the horizon: A River Runs Through the Rack

Atienza Alonso, an EPFL professor and expert in AI computing architecture, recently estimated that Swiss data centers already consume an amount of water equivalent to the needs of 400,000 to 800,000 people. The growing adoption of cloud and AI technologies, he believes, will make these facilities even thirstier.

To assess the impact of AI technologies on water consumption in Switzerland, it is important to distinguish between training AI models and using them. Today, large language models are generally trained in gigantic data centers, where tens of thousands of chips spend weeks searching for discernible patterns in staggeringly large datasets. This requires enormous quantities of electricity and water, often placing severe strain on local resources. For instance, in the U.S. state of Iowa, over 50 million liters of water were consumed in a single month during the training of OpenAI’s GPT-4 model.

A look at the footnotes of some major AI tools suggests that such training remains relatively rare in Switzerland. However, so-called inference—that is, the use of trained AI models—is gaining traction. Microsoft alone is currently investing more than 400 million Swiss francs to operate its AI services in Swiss data centers. Amazon Web Services and Google are also making similar investments. In particular, an increasing number of high-capacity server racks are being installed in the Zurich region, enabling generative AI applications such as Claude, ChatGPT, DALL-E, and Gemini to be used without the processed data ever leaving Switzerland. These AI tools are utilized for risk analysis in the financial sector, simulations in the pharmaceutical industry, or data analysis in the postal and insurance sectors, yet also to generate marketing campaigns and commercials.

Processing a single prompt in a Swiss data center requires relatively little electricity and water. Estimates suggest that ChatGPT consumes 10 to 20 milliliters of water to generate a medium-length response in cool temperate climates such as Switzerland’s. At first glance, that seems modest. Multiplied by millions of queries every day and several similar AI providers, however, this quickly adds up to water consumption on the scale of Swiss towns with populations of 10,000.

Next to high usage rates, there is another reason for AI tools’ high water consumption. To handle the flood of data associated with large language models, AI servers pack computing hardware many times more densely than conventional server systems. This means that more chips are running at full tilt in a smaller space. While racks in conventional data centers previously had power ratings of about five to ten kilowatts, those figures are now increasing by a factor of ten to thirty. For example, at a breakfast session hosted by the Swiss Data Center Association, NorthC reported on new 100-kW racks, while Digital Realty cited loads of 300 kW per AI rack, and Flexbase, an “AI data center” scheduled to go into operation in 2028, cited as much as 1,200 kW per rack. The amount of heat generated can be imagined as roughly equivalent to piling 1,000 ordinary toasters on top of one another and running them at their highest browning setting.

Conventional air-cooling systems—whether electricty-based or adiabatic—reach their limits with heat output on this scale. The most advanced data centers are therefore increasingly turning to “liquid cooling.” This involves circulating water through tubes as close to the chips as possible to carry away the heat. Veritable rivers thus flow through the racks. The data center industry often markets this cooling method as a new wonder weapon. The arguments: Water can absorb server heat a thousand times more efficiently than air, and within the closed-loop system it does not evaporate.

That, however, is only one side of the coin. While very little water is actually lost in the tubes that run around the chips, the heated water must itself be cooled. To do this, it is usually routed outdoors, where the hot water pipes may, for example, be sprayed with cold water or cooled adiabatically—a process in which water once again evaporates.

Image credit: Christoph Miller (CC BY 4.0)

Although “liquid cooling” consumes substantially less water than water-assisted air-cooling systems, estimates suggest that these water savings will be offset by the steadily increasing performance of AI chips and the rapidly expanding use of AI applications. The International Energy Agency and cooling-system experts such as Geoff Gage therefore predict that the direct and indirect water consumption of data centers could more than double in the coming years.

Image credit: Christoph Miller (CC BY 4.0)

Thus, technical solutions will not eliminate the need for a political debate on data centers’ water consumption. Worse, conflicts over water use could intensify, especially during the increasingly dry summer months. When outdoor temperatures are high, data centers require dramatically more water to prevent their servers from overheating. A recent study by the University of California, Riverside, concludes that data centers’ peak demand on hot days can be three to ten times higher than the annual average. At the same time, agriculture and households also need more water for livestock, fields, and gardens then—water that is already in short supply in some parts of Switzerland during the height of summer. As a result, shortages and conflicts over water use will become more likely if these thirsty monoliths are built in the wrong places and equipped with the wrong cooling technology—or, worse, if they are allowed to quench their thirst as they please, free from regulation and requirements governing consumption and efficiency.

From this perspective, it seems all the more urgent that operators provide specific and reliable figures on their water consumption. How much water do they consume exactly? How efficiently do they use it? How high are their peak loads? Only with such site-specific water data can authorities plan effectively and manage these valuable water resources sustainably.

At present, however, there is little sign of greater transparency. While the federal government is currently trying to collect more comprehensive data on local water consumption by households, industry, and agriculture in order to develop its drought early-warning system, data centers remain exempt from disclosing their water consumption. That fits the pattern: according to the Center for Secure Water, data centers are among the least transparent industrial water users of all. In light of the current water crisis, it is therefore all the more important to step up the pressure on the operators. Only with data on both direct and indirect water consumption will there ultimately be a public debate about how local water resources should be allocated in the future, and where it makes sense to build new data centers at all. Time is running short: even in Switzerland, the water tower of Europe, inexhaustible supplies of “blue gold” can no longer be taken for granted.