Protest banners declaring “Protect our water,” “Water for people, not for AI,” and “Do not touch our water” can be seen from Texas to New Mexico and Arizona. Residents are increasingly voicing their concerns regarding water consumption, energy expenditure, and pollution caused by the rapid expansion of data processing centres across the United States, as technology companies strive to enhance their AI capabilities.
Data centres often rely on water to dissipate heat generated by their operations, using the process of evaporation to cool down, similar to how sweat regulates body temperature. The processors within these centres can reach internal temperatures of up to 80 degrees Celsius (176 degrees Fahrenheit), mirroring the overheating of an overloaded laptop.
While it is evident that data centres consume substantial amounts of energy, their water usage presents a more intricate picture. Online claims vary widely, with some suggesting that chatbots can use up to 500 millilitres of water per query, while others assert that the water consumption issue linked to AI is negligible. Furthermore, the water usage data provided by large tech firms is often incomplete and inconsistent.
Experts highlight that, overall, the water consumption of data centres is minimal when compared to sectors such as agriculture and certain manufacturing industries. Research estimates that the cooling systems of data centres consumed approximately 66 billion litres of water in 2023, representing less than 1% of the total water usage in the United States.
However, this figure could significantly increase with the rapid growth of AI data centres, which may not pose a substantial problem in regions with abundant water resources but could exacerbate water scarcity in drought-stricken areas like New Mexico and Arizona.
Innovative Solutions for Water Conservation
Fortunately, engineers assert that numerous strategies can be implemented—many of which are already in progress—to alleviate the pressure on local water resources. Fengqi You, an energy systems expert at Cornell University, states that “there are many good ideas.” Coupled with efforts to replenish and restore natural water resources, there is a strong possibility that net-zero water consumption for on-site cooling can be achieved over time.
Recent advancements in AI technology have already led to a reduction in water demand for many data processing centres. Shaolei Ren, an electrical and computer engineer at the University of California, Riverside, noted that an estimate he made based on GPT-4—a predecessor of the model powering ChatGPT—suggested that drafting a short email would consume 500 millilitres of water. However, this estimate has quickly become outdated as AI models have become more efficient. By 2025, Google estimated that only five drops of water are used to process an average-length query with its chatbot, Gemini.
Even the smallest quantities can add up, especially with the rising use of AI. Fengqi You’s team recently predicted that by 2030, data centres in the United States could consume between 731 billion and 1.12 trillion litres of water annually—equivalent to the annual drinking water supply of New York City.
Nevertheless, these estimates account for not just cooling but also the increased water used to generate the electricity that powers data centres. Much of this energy comes from burning coal or gas, which requires water to cool steam and convert it back into liquid after it has been used to turn electricity-generating turbines. In other words, the water consumption associated with AI could be drastically reduced by transitioning to solar and wind energy, which require little to no water for operation, according to You.
Strategic Location Choices and Energy Sources
You further suggests that companies should establish data centres in areas that are less prone to drought and have abundant renewable energy sources, such as regions in Montana, Nebraska, certain parts of Texas, and South Dakota, instead of constructing new centres in water-scarce locations like Arizona, New Mexico, and Southern California. This strategic location choice, combined with other measures, could potentially reduce the future water footprint of AI by as much as 86%.
Advancements in cooling technologies are also contributing to this endeavour. Modern data centres specialising in AI are already more efficient in their water usage than their predecessors. Traditional data centres employ fans to expel heat from processor-laden racks, followed by cooling systems that channel cold water throughout the facility to cool down hot air. The warmed water is then sent to a cooling tower that releases heat while also losing some water to the atmosphere. As Eric Masanet from the University of California, Santa Barbara, points out, this evaporative cooling method is quite water-intensive.
However, AI data centre processors generate so much heat that it is challenging to dissipate solely through air cooling. Consequently, many tech companies—including Amazon, Microsoft, and Google—have adopted a more efficient method known as “liquid cooling.” In this approach, a network of pipes filled with water, sometimes mixed with temperature-regulating chemicals, runs above the hardware, directly cooling the processors rather than the entire data centre. Because this is a closed-loop system, water is not lost; the heated fluid can be cooled by exposing the pipes to outside air, provided conditions are sufficiently cool, or through an air conditioning-like system that consumes energy.
Emerging Cooling Technologies and Sustainability Goals
When temperatures soar or humidity levels are high, centres must resort to additional, more effective methods involving water, such as spraying water into the air surrounding the pipes. Mechanical engineer Vaibhav Bahadur from the University of Texas at Austin notes that most AI data centres in Texas only utilise this water-based cooling during the hottest periods of summer.
Nevertheless, he and his colleagues have recently estimated that the growing number of data centres in the state—currently consuming less than 1% of Texas’s water demand—could rise to between 3% and 9% by 2040. Nonetheless, he anticipates that this figure will decrease significantly. “Data centres are becoming much more efficient in their water usage,” he asserts.
Indeed, recent technological advancements have facilitated the avoidance of water cooling even in warm climates, explains Josh Parker, sustainability lead at Nvidia, the company behind many processors used in AI data centres. Their latest generation of processors operates at such high temperatures that they can be cooled with water at approximately 45 degrees Celsius (113 degrees Fahrenheit), significantly warmer than traditional methods. This reduces the reliance on water cooling. Unless local temperatures regularly exceed 113 degrees Fahrenheit—a threshold that is typically surpassed only during extreme heatwaves—”efficient large fans are generally sufficient to cool the infrastructure, with ambient air being adequate to dissipate heat,” Parker elaborates.
Shaolei Ren warns that some companies may hesitate to allow their processors to function at such high temperatures, as this could impact computational performance, while others may also house non-AI servers within the same space that cannot tolerate the heat. However, alternative cooling technologies are in development.
Some of these innovations involve sophisticated pipe systems that extract heat from processors even more effectively. Several American companies are exploring immersion cooling, where the data centre hardware is submerged in a tank filled with a cooling liquid. In China, some companies have adopted a similar approach by installing data processing centres in the ocean, although this complicates access should hardware replacement or updates be necessary, Bahadur notes.
Commitment to Sustainable Water Management
Advancements in cooling technology will be essential for tech companies to meet the sustainability goals they have set for their operations. According to a statement from Amazon, its Web Services division aims to achieve a positive water balance by 2030—returning more water to communities than its direct operations consume—and by 2024, they expect to be over halfway there.
A spokesperson for OpenAI directs attention to a webpage outlining the company’s plans to minimise water consumption, while Google aims to replenish 120% of the freshwater consumed by its data centres by 2030 through various responsible water management projects.
Some companies, such as Google, are also discovering clever ways to harness the residual heat from data centres to address other sustainability challenges, such as reducing reliance on fossil fuels for building heating. Across Europe, some data processing centres are redirecting the heat from their cooling circuits to urban heating systems used for warming buildings.
The future of water consumption for AI will depend on the decisions made by tech firms, engineers, and policymakers in the coming years, Masanet asserts: the location of data centres, their local energy infrastructure and climate, as well as their design will all play crucial roles. “If you choose one set of options, your figure will skyrocket,” he explains. “If you choose another set, it will be significantly reduced.”