How Much Water Do Data Centers Use?

Data center water usage is the part of AI's footprint people notice last, because water never shows up on an electricity bill. Yet cooling servers and generating the power to run them both consume water, and at scale those liters add up in places that can least afford them.

This article separates the two sources of data center water usage, explains the metric used to track it, gives a grounded per-kilowatt-hour estimate, and shows why the same workload can be water-cheap in one region and water-expensive in another.

The two sources of data center water usage

Data center water usage comes from two places. The first is on-site cooling: many facilities use evaporative cooling towers that consume water directly to shed the heat produced by dense racks of accelerators. The second is off-site, at the power plants that supply electricity, many of which use water for steam cycles and cooling.

Adding both together, a reasonable estimate is around 3.4 liters per kWh consumed. Per request that is tiny, a fraction of a small cup for a text answer, but multiplied across billions of requests it becomes a figure worth managing. Our companion piece how much water does ChatGPT use works through the per-query math.

• On-site: evaporative cooling that removes heat from servers
• Off-site: water used at power plants generating the electricity
• Combined estimate: about 3.4 liters per kWh consumed

WUE, the water version of PUE

Just as PUE tracks energy overhead, water usage effectiveness, or WUE, tracks how many liters a facility uses per unit of computing energy. A lower WUE means a more water-efficient site. The figure depends heavily on cooling design, with air-cooled and closed-loop liquid systems using far less water than open evaporative towers.

There is often a trade-off: evaporative cooling saves energy but spends water, while some drier-running designs save water but use more electricity. Understanding that tension is key to judging a facility fairly. For the energy side of the same coin, see what is PUE in data centers.

Why region changes everything

The same server can have very different water impact depending on where it lives. A data center in a cool, wet climate may barely need evaporative cooling, while one in a hot, dry region may lean on it heavily, and that region may also be water-stressed to begin with. The off-site component varies too, because a grid dominated by hydro or thermal plants uses more water than one built on wind and solar.

This is why a single global number can mislead. The honest framing is that data center water usage averages around 3.4 liters per kWh but ranges widely with climate, cooling design and grid mix. The broader context sits in our environmental impact of AI guide.

Reducing and accounting for water

Water impact shrinks the same way carbon does. Using efficient models cuts the energy behind every request, which cuts both the cooling water and the generation water at once. Siting compute in cooler or water-abundant regions, and choosing air or closed-loop cooling, reduces the on-site share directly.

Beyond reduction, water can be measured and offset. Ecoia estimates the water behind every request and retires verified water offsets past 200 percent of measured impact, alongside carbon. That measurement-first approach is described on our how it works page, and reduction habits are collected in how to reduce your AI carbon footprint.