Is AI’s Water Use a Serious Problem? Location Matters Most

Introduction

The claim that an AI conversation consumes half a liter of water is memorable, but it is not a universal measure of AI’s water footprint. Data centers need energy and cooling, yet their water demand varies with model efficiency, hardware, climate, electricity sources, and cooling systems. The central issue is therefore not only the national total. It is also whether new facilities are being built in communities that already face water shortages.

Key takeaways

• The national share is modest, but demand is growing. Researchers estimated that US data center cooling systems consumed about 66 billion liters of water in 2023, less than 1 percent of total national consumption and far below agriculture or some manufacturing sectors. Rapid AI expansion could push annual US data center water use to 731 billion to 1.125 trillion liters by 2030. The upper estimate is roughly comparable to New York City’s annual drinking-water supply.
• The half-liter estimate is dated. It came from earlier work involving GPT-4. As models and hardware have become more efficient, Google estimated in 2025 that a median-length Gemini query used about five drops of water. Small per-query figures can still accumulate at massive scale, however, while company disclosures remain incomplete and inconsistent.
• The footprint is both direct and indirect. Conventional facilities often rely on evaporative cooling. AI processors produce more heat, increasing the value of liquid cooling. Water is also consumed upstream: coal- and gas-fired power plants use water to cool equipment, so a data center’s footprint cannot be measured only at the site.
• Location can matter more than averages. A facility in a water-rich region with abundant renewable power may create less local stress than one in a drought-prone area. Arizona, New Mexico, and Southern California illustrate why a small national percentage can still matter greatly to a particular community.
• Cooling technology is improving. Liquid-cooling loops can circulate fluid without the continual losses associated with evaporation. Newer processors can operate with warmer cooling water, reducing the need for additional water-based systems. During very hot or humid conditions, facilities may still turn to water-intensive measures such as misting.

Why it matters

AI’s water footprint cannot be responsibly summarized with one number. A meaningful assessment should distinguish on-site cooling water from water used to generate electricity, and should disclose local water stress and seasonal peaks. National averages may otherwise hide the effect on a specific watershed during the hottest months.

The response is not limited to conserving water inside the building. Experts and studies point to more efficient processors and models, greater use of wind and solar power, and siting new facilities where drought risk is lower and clean electricity is available. One study cited in the source estimated that combining strategic siting with renewable energy and efficiency measures could cut future water use by up to 86 percent. Replenishment and ecosystem restoration may also help support net-zero water for on-site cooling, although that depends on implementation and accounting boundaries.

AI water use is therefore neither a meaningless online scare nor automatically a nationwide water crisis. It is a location-sensitive infrastructure problem. In water-stressed regions, permitting, transparent reporting, and seasonal limits should be treated as part of responsible compute expansion.

Source: Ars Technica AI