After concerns about growing electricity demand, attention is now turning to artificial intelligence's water consumption. But how significant is the issue really?
Data centers must continuously cool their servers, which operate 24 hours a day and generate substantial amounts of heat. Without effective cooling systems, equipment can malfunction or become damaged.
The rise of AI is intensifying this challenge. AI servers concentrate unprecedented computing power within increasingly dense racks, generating far more heat than traditional IT infrastructure.
Historically, cooling relied primarily on air, combined with air conditioning systems. However, as computing power continues to increase, water-based cooling solutions are becoming more attractive. Water transfers heat much more efficiently than air, making it particularly suitable for high-density computing equipment.
This is why water has become a central topic in discussions about AI's environmental impact.
Water Use in Perspective
Globally, agriculture accounts for approximately 70% of freshwater withdrawals, compared with 20% for industry and 10% for domestic use.
Within the industrial sector, energy production is by far one of the largest users of water. Thermal power plants, whether fueled by fossil fuels or nuclear energy, require significant amounts of water to cool their facilities.
For comparison, one study estimates that in the United States, more than 85% of industrial water use is linked to energy production, compared with roughly 0.1% for data centers.
Why Are Data Centers Still a Concern?
The issue is not only how much water is used, but also where it is consumed.
Because water is relatively inexpensive, it is often not a major factor when selecting a location for a data center. Operators typically prioritize access to affordable electricity and robust network infrastructure. As a result, some data centers are built in regions already facing significant water stress.
Texas is one example. The state hosts more than 500 data centers while also experiencing growing pressure on its water resources.
A medium-sized data center can consume up to 110 million gallons of water per year for cooling, equivalent to the annual water use of approximately 1,000 households. The largest facilities may use up to 5 million gallons per day, a volume comparable to the water consumption of a town of 10,000 to 50,000 residents.
The impact of data centers on water availability should therefore not be underestimated. However, it is also important to remember that other sectors consume significantly larger volumes of water and are often responsible for greater water pollution.
Withdrawing Water Is Not the Same as Consuming Water
To understand the issue, it is essential to distinguish between two concepts: water withdrawal and water consumption.
Water Withdrawal
Water withdrawal refers to taking water from a river, lake, or groundwater source and later returning it to the environment.
This is common in many power plants, which use water to cool their equipment before discharging it back into the environment. Although the water is generally returned, it is often released at a higher temperature or with altered quality, potentially disrupting local ecosystems.
In addition, certain activities significantly degrade water quality:
Agriculture: fertilizers, pesticides, and livestock waste
Mining: heavy metals and acid mine drainage
Textile industry: dyes and chemical effluents
Oil and petrochemical industry: industrial discharge and spills
Water Consumption
Water consumption means that water does not return directly to its original source after use.
This is the case in agriculture, where water is absorbed by soil and crops. It is also true for data centers that rely on evaporative cooling, where a significant portion of the water evaporates into the atmosphere.
It is estimated that around 80% of the water used in some data center cooling systems is lost through evaporation. As a result, it is no longer available in the short term for other uses, including household consumption.
AI's Water Footprint Extends Beyond Data Center Cooling
When discussing water consumption related to AI, data centers often receive most of the attention. Yet another stage of the value chain is even more water-intensive: semiconductor manufacturing.
The production of computer chips requires large quantities of ultrapure water. Thousands of times cleaner than drinking water, ultrapure water is essential for cleaning silicon wafers and removing microscopic particles that could damage semiconductor components.
Producing 1,000 gallons of ultrapure water typically requires between 1,400 and 1,600 gallons of municipal water.
Today, a semiconductor fabrication plant can consume up to 10 million gallons of ultrapure water per day, equivalent to the daily water use of approximately 33,000 U.S. households.
In other words, AI's water footprint extends far beyond data centers. It also includes the electricity generation needed to power them and the manufacturing of the chips that make this technological revolution possible.
Solutions Are Emerging
In response to these challenges, the industry is developing technologies aimed at reducing water consumption.
Direct Liquid Cooling
Direct Liquid Cooling circulates liquid directly over the hottest components.
This technology improves energy efficiency and can significantly reduce water requirements when combined with heat-rejection systems that do not rely on evaporation. Its main drawbacks are higher installation costs and more complex planning requirements.
Immersion Cooling
This approach involves fully immersing servers in a non-conductive liquid.
Heat transfer is extremely efficient, reducing both energy and water requirements. However, the technology requires substantial upfront investment and can make maintenance more complex.
Free Cooling
Free cooling uses outside air directly whenever climatic conditions allow.
This solution reduces reliance on mechanical air conditioning and limits water use. It is particularly well suited to colder regions, such as the Nordic countries. However, the heat generated by servers is often released into the atmosphere rather than reused.
Heat Recovery
The example of Infomaniak in Geneva illustrates an especially promising approach.
Commissioned in 2024, the company's data center captures the heat generated by its servers and feeds it into the district heating network. Ultimately, this recovered heat could meet the heating needs of approximately 6,000 homes.
The facility uses outside air for cooling and requires neither traditional air conditioning nor water-intensive cooling towers. Heat thus becomes a resource rather than a waste product.
Conclusion
Water is an essential resource, and concerns about its use by AI are entirely legitimate.
Today, water consumption by data centers remains modest compared with other sectors such as agriculture and energy production. Nevertheless, the rapid growth of AI warrants careful consideration of how this resource is managed sustainably.
The key issue is not so much the total volume of water consumed, but rather where that water is used, how it competes with local needs, and the efficiency of the cooling technologies involved. Over time, access to water could become as strategic as access to electricity for the development of artificial intelligence.
History shows that companies capable of using resources more efficiently often gain a lasting competitive advantage. In the AI sector, access to affordable energy and water may become almost as important as computing power itself. Companies that succeed in reducing energy and water consumption, increasing resource recycling, and recovering waste heat from their infrastructure are likely to be best positioned to support the sector's growth while meeting rising environmental expectations.
For investors, these innovations could become a key indicator of long-term value creation. More than a purely environmental issue, efficiency in the use of energy and water may prove to be a significant competitive advantage in the AI economy.
All companies mentioned in this article are not recommendations and sometimes not in our investment universe. If you need further in depth analytics please contact your Mirabaud Relationship Manager.
