When you ask an artificial intelligence a question, the response feels completely weightless. It looks like simple text floating down from a digital cloud. But behind every prompt sits an enormous physical computer that consumes large amounts of electricity and gets extremely hot. To keep those computer chips from overheating, data centers consume millions of liters of freshwater.

The heat behind the screen

Computers process calculations using electrical current, and that electricity turns directly into heat. Massive data centers house thousands of servers packed with high-performance graphics processing units (GPUs) and specialized processors working together day and night.

When millions of people chat with an AI at the same time, the server racks generate blistering heat. If computer chips get too hot, they slow down or suffer permanent physical damage. To keep these facilities running safely, data centers must remove that heat around the clock.

Cooling towers and evaporation

While personal computers use small fans to push heat into the surrounding room, giant AI server farms often rely on evaporative cooling systems. Water absorbs and transfers heat far more effectively than ambient air.

In evaporative cooling towers, warm water carrying heat away from the server racks is pumped to an outdoor tower and sprayed into a stream of moving air. As some of the water evaporates into steam and drifts into the sky, it carries the thermal energy away. However, that evaporated freshwater does not return to the local reservoir; it is consumed. Furthermore, cooling systems require clean freshwater because mineral deposits in recycled or impure water would quickly clog the cooling pipes.

Measuring the water in a prompt

In 2023, researchers from the University of California, Riverside and the University of Texas at Arlington published a study titled Making AI Less "Thirsty". Led by researchers Pengfei Li, Jianyi Yang, Mohammad Islam, and Shaolei Ren, the team quantified the hidden freshwater footprint of modern AI models.

The researchers estimated that training OpenAI's GPT-3 model in Microsoft's United States data centers directly evaporated approximately 700,000 liters (about 185,000 gallons) of clean freshwater before the model ever answered a question from the public.

Daily conversations also add up. The researchers estimated that a standard exchange of 10 to 50 medium-length prompts consumes roughly 500 milliliters of freshwater—the equivalent of one standard plastic water bottle. While one bottle feels small on its own, billions of prompts per month across global users add up to enormous water demand.

Making AI less thirsty

Researchers point out that water consumption varies depending on where and when data centers operate. Running servers during cooler nights or in temperate climates reduces the need for evaporative cooling. Companies are also experimenting with closed-loop liquid systems that recycle coolant without evaporating it into the air.

Until these solutions become standard across the industry, every conversation with an AI model leaves behind a very real physical footprint in the world's freshwater supply.

Sources

This article accompanies episode 1 of our series Tech Explained.