How Much of a Problem is AI’s Water Use?

The extent to which artificial intelligence (AI) impacts water consumption remains a contentious issue. While data centers' water use pales in comparison to that of other industries, it's a significant concern in water-stressed regions. Local communities in Texas, New Mexico, and Arizona are protesting the expansion of data centers due to their high water, energy, and pollution demands.

Data centers often employ water to absorb heat from air or surfaces, causing cooling as it evaporates, similar to how sweating cools our bodies. Inside data centers, processors can reach temperatures of up to 176 degrees Fahrenheit, much like an overheated laptop. However, water consumption in data centers is complex. Claims online vary widely, from the half-liter-per-query figure for chatbots to the assertion that AI's water issue is nonexistent.

Data center water usage is often reported incompletely or inconsistently by tech companies themselves. Experts emphasize that data center water consumption is dwarfed by that of agriculture and certain manufacturing sectors. Estimates suggest data center cooling systems consumed 66 billion liters of water in 2023, accounting for less than 1% of the nation's total consumption.

This figure could surge with the rapid construction of AI data centers, creating more strain in drought-prone areas like New Mexico and Arizona. However, engineers believe there are multiple ways to mitigate this strain. Fengqi You, an energy systems expert at Cornell University, points out numerous promising ideas, including replenishing and restoring natural water resources, which could potentially lead to net-zero water usage for on-site cooling.

Recent advancements in AI technology have already reduced water demand for many data centers. For instance, a 2024 estimate suggested that drafting a short email would consume 500 milliliters of water, a claim now outdated due to AI model efficiency improvements. Google's 2025 estimate indicates that processing a median-length query with its chatbot Gemini requires just five drops of water.

Nevertheless, these minute amounts accumulate given the surge in AI use. Fengqi You's team projected that by 2030, US data centers could consume between 731 billion to 1,125 billion liters annually, equivalent to New York City's annual drinking water supply. Yet, these figures include not just cooling but also the water used to generate electricity for data centers, which often involves cooling steam through water.

By transitioning to solar and wind energy sources, AI's water consumption could be substantially reduced, says Fengqi You. Companies should also consider site selection, favoring areas less prone to drought and with abundant renewable energy sources, such as Montana, Nebraska, parts of Texas, and South Dakota, instead of constructing new data centers in water-stressed regions like Arizona, New Mexico, and Southern California.

Strategic siting, coupled with other measures, could potentially cut AI's future water footprint by up to 86%, according to Fengqi You. Additionally, newer AI data centers are already more water-efficient than their predecessors. Pre-AI data centers relied on fans to dissipate heat from processor-filled racks, followed by chilling systems that circulated cool water through the building to cool the warm air again.

This evaporative cooling process uses substantial water, with heated water eventually escaping into the atmosphere through cooling towers. To enhance efficiency, many tech companies now employ liquid cooling, a method involving pipes filled with water (sometimes mixed with temperature-regulating chemicals) running atop hardware to directly cool processors.

This closed-loop system minimizes water loss, as heated fluid can be cooled through exposure to outdoor air or air-conditioning-like systems. However, during extreme heat or humidity, additional cooling methods involving water may be employed, such as misting the air around pipes. Mechanical engineer Vaibhav Bahadur at the University of Texas at Austin notes that most AI data centers in Texas currently use water-based cooling primarily during the hottest parts of summer.

Nonetheless, he and his colleagues estimate that by 2040, the state's burgeoning data centers — currently accounting for less than 1% of the state's water demand — could be utilizing 3 to 9% of that water.