AI uses water: Cooling and energy
The question of why Artificial Intelligence (AI) consumes water may be surprising at first glance, as AI is perceived as pure software. However, behind the seemingly intangible processes lies an extensive physical infrastructure. AI does not "drink" water in the direct sense. Water consumption mainly arises from cooling servers in data centers, electricity supply, and the manufacturing of hardware on which AI applications run.
Source: Own representation
AI & Water Consumption
When discussing AI and water consumption, data centers are almost always meant. These enormous halls full of servers consume a particularly large amount of power for AI workloads and consequently release a lot of heat. This waste heat must be reliably dissipated to prevent overheating of chips and hardware failures. (Source)
Water is used in two main forms: directly on-site as cooling water (Scope 1) and indirectly via the electricity mix (Scope 2). Many power plants require water for cooling and steam generation. A third, often overlooked, aspect is water consumption in the supply chain, especially in chip and server manufacturing (Scope 3). (Source)
Many data centers use cooling systems that evaporate water, as evaporation dissipates heat very efficiently. A common principle is the cooling tower, where a portion of the water evaporates and carries the heat away. The rest of the water remains in the cycle but must be regularly replenished. Here, the distinction between "Withdrawal" (total water extracted) and "Consumption" (evaporated portion) is important. In practice, with good water quality, a large part of the withdrawal evaporates, often around 80%. (Source)
The actual water consumption depends heavily on weather, location, and operating method. Estimates for evaporation in data centers range between 1 and 9 liters per kWh of server energy. On hot days, the demand increases because more water is needed for the same cooling capacity. (Source) Local examples show that the water consumption of individual large data centers can reach the scale of municipal consumers, such as the Council Bluffs, Iowa location with 1.3 billion gallons of drinking water consumption in 2024. (Source)
Industry and policy reports confirm that large facilities can require several million gallons of water per day in extreme cases. To make water consumption comparable, operators use key figures such as "Water Usage Effectiveness (WUE)", which relates water consumption to IT energy. (Source)
Source: bblloobb.com
The water cycle in data centers: water is used for cooling and is lost through evaporation.
Background & Context
Water becomes a point of contention when data centers grow in water-scarce regions or when new projects downplay the demand. Authorities are increasingly including the issue in approval processes, as the case in Chile shows, where a court demanded stricter environmental reviews for a planned Google data center, partly due to water issues in a drought-prone region. (Source) International analyses describe that the expansion of AI data centers often takes place in regions with already high competition for water. (Source)
Even if a data center evaporates little water on-site, the indirect proportion via the electricity mix remains relevant. Thermal power plants require water for cooling, and this demand depends on the type of power plant and cooling method. (Source) The U.S. Geological Survey publishes data and reports on water withdrawal and consumption from thermal power generation in the U.S. A key consequence is that cooling decisions can only "shift" water. Dry cooling systems often save water on-site but frequently increase electricity demand, which can enlarge the indirect water footprint in the grid. (Source)
Another aspect is water consumption in the hardware supply chain. Chip manufacturing is extremely water-intensive because wafers are rinsed in many process steps and treated with ultrapure water (UPW). Approximately 2,200 gallons of water are needed to manufacture integrated circuits on a 300mm wafer, about 1,500 gallons of which is ultrapure water. (Source) This part is easily overlooked, as it is not directly visible at the data center, but it is part of the complete water footprint of AI hardware. (Source)
Source: user-added
Projected global water consumption by AI in 2027, illustrating the scale of the problem.
Solutions & Challenges
Part of the answer lies in site selection and climate profile, as cooling requires more water or more electricity in hot weather. (Source) Technically, much is shifting towards liquid cooling and closed-loop systems, as AI racks achieve high power densities and air cooling reaches its limits. (Source) Water-based cooling can be energy-efficient, and alternatives like immersion cooling with non-aqueous liquids can avoid water evaporation. (Source)
Source: interestingengineering.com
AI can also be a valuable tool for optimizing and managing water resources.
Large technology companies are addressing the issue through goals and replenishment programs. Microsoft has set a goal to be "water positive" by 2030, meaning it will return more water than it consumes. (Source) Google pursues a similar strategy with a replenishment goal of 120% of consumption by 2030. (Source) AWS and Meta have also set the goal of "water positive by 2030" and publicly report on their progress. (Source) (Source)
However, these commitments do not automatically solve the core problem that communities feel: water becomes scarce where it is withdrawn and evaporated. Many experts therefore call for greater transparency on direct and indirect water use. (Source)
In summary, AI does not directly "drink" water, but water consumption arises from the heat generated by computing power, the cooling of infrastructure, and the water-intensive processes in electricity generation and the supply chain. Whether this becomes a problem depends heavily on the context: the cooling technology used, the electricity mix, and especially the local water stress at the location. The question ultimately is how much digital convenience fits into a physical system that cannot supply water faster on hot days than the environment allows. (Source)