Training an AI model can leave a water footprint.

Why AI servers need so much cooling

Training a large language model means running thousands of processors at high power for long periods. Nearly all the electricity they use becomes heat. If that heat builds up, processors must slow down or shut off to protect themselves, so data centers move it away continuously.

One common method uses cooling towers. Warm water absorbs heat from the building, then some of it evaporates as air passes through the tower. Evaporation carries heat away efficiently, but the water that leaves as vapor must be replaced. Some water is also drained to prevent minerals from concentrating in the system.

The water cost depends on where the heat goes

Not every data center uses the same cooling design. Some rely more on outside air or closed-loop systems; others use evaporative cooling, especially when it can reduce electricity demand. The tradeoff is that a system can save power while consuming water. The source matters, too: facilities may use drinking-quality freshwater, reclaimed wastewater, or other supplies, depending on location and infrastructure.

Water can also be used indirectly to generate the electricity that powers a facility. That means the total water footprint depends not only on the cooling equipment, but also on the local power grid and weather. A hot, dry day can increase cooling demand precisely when water may be scarce.

Why a model’s footprint is hard to pin down

A model’s training run is only one part of the calculation. Estimates change with the length and intensity of training, the data center’s cooling system, its location, and how researchers allocate shared facility and electricity use to a particular job. Once a model is deployed, answering user requests adds another, ongoing demand for computing and cooling.

The practical consequence is local: a data center’s water use can matter most where it draws on a strained watershed. Choosing a site and cooling method is therefore not just an engineering decision; it can determine whether the heat from AI computing is carried away with abundant reclaimed water—or with water a community also needs.