How much of a problem is AI’s water use?
Pre-AI data centers use fans to carry away the heat from processor-filled racks, then use chilling systems funneling cool water through the building to cool the warm air down again. The heated water is then fed into a cooling tower that removes the heat but also loses some of the water to the atmosphere. This evaporative cooling uses a lot of water, says Eric Masanet of the University of California, Santa Barbara, who researches data center sustainability.
But the energy-hungry processors inside AI data centers generate so much heat that it’s hard to remove it with air alone. That’s why many tech companies—including Amazon, Microsoft, and Google—use a more efficient technique called liquid cooling.
In this method, a system of pipes filled with water, sometimes mixed with temperature-regulating chemicals, runs on top of the hardware, cooling the processors directly rather than cooling the entire data center. And because this is a closed-loop system, no water is lost: The heated fluid can be cooled down by exposing the pipes to the outside air, if conditions are cool enough, or an air-conditioner-like system, which uses energy.
But when it’s extremely hot or humid, the centers need to bring online extra, more effective methods that involve water; one of these is to mist the air around the pipes. Mechanical engineer Vaibhav Bahadur at the University of Texas at Austin says that most AI data centers in Texas only use this kind of water-based cooling for the hottest parts of the summer.
Still, he and his colleagues recently estimated that the state’s growing number of data centers—which currently consume less than a percent of the state’s water demand — could be using 3 to 9 percent of that water by 2040. He expects that figure to reduce considerably, though. “Data centers are becoming much more efficient in their water usage,” he says.
Indeed, recent technological developments have made it easier to avoid water-based cooling, even in hot climates, says Josh Parker, head of sustainability for NVIDIA, the company that designs many of the processors that populate AI-specialized data centers. Their latest generation of processors run at such high temperatures that they can be cooled with water at around 113 degrees Fahrenheit which, Parker says, is significantly warmer than traditional approaches. That reduces the need to bring on water-based cooling.
Unless local temperatures regularly exceed 113 Fahrenheit, a threshold generally crossed only during severe heat waves, “typically we can just get away with large efficient fans cooling the infrastructure and the ambient air is sufficient to pull the heat away,” Parker says.
