AI is a thirsty customer.

We’re aware of that when it comes to energy and water, as most everyone who covers the rise of artificial intelligence-enabled data centers knows. It’s hard to stay cool while coming to terms with the hyper consumptive traits of supercomputing.

The alarms are sounding to residents everywhere who are concerned about jobs and natural resources. At the same time, keeping the U.S. atop the data game is ranked as a key national security issue by many.

So, are we invariably and interminably caught between a rock and dry place when it comes to AI of the future? Not necessarily, says one of the market leaders for a company which specializes in heat exchange solutions for data centers.

The key to managing water and energy for gigawatt-scale AI compute is recognizing the challenges and limitations while also mulling the efficiency of various cooling technologies, according to Henry Cubillan, who is senior vice president of commercial Americas for heat-exchange provider Kelvion. The answers are situational as well as locational.

“AI is putting new pressure on electricity and water resources because of the scale and density of the infrastructure required to support it,” Cubillan said in response to emailed questions from Water Technology Online. “But I would not describe AI itself as the threat. The bigger question is whether infrastructure development keeps pace with the resources available in a particular location.”

Rising rack density as AI processing becomes more powerful is considered as much a formidable threat for water resources as it is for the energy grid facing an imbalance of power generation capacity to energize those next-gen loads. Nearly half of the U.S. faces drought conditions, while many AI data centers are looking toward liquid cooling to deal with the heat generated by mega computations and large language models.

There are options out there to deal with that, Kelvion’s Cubillan pointed out. Hyperscalers and data center developers need to keep their minds open as they plot new digital infrastructure.

“Water consumption therefore depends on the cooling architecture, climate, operating conditions and other facility requirements, not simply the size of the data center,” he said. “The industry needs to assess these factors when designing new facilities rather than assume that greater computing capacity automatically means greater ongoing water consumption.”

Some reports estimate that AI-enabled data centers might consume millions of gallons of water per day. Cubillan cast doubt on those kinds of projections given that developers can build computing while utilizing one or more liquid cooling technologies tailored to the situation.

“Closed-loop liquid cooling paired with dry coolers can significantly reduce ongoing water consumption for cooling because these systems do not rely on evaporation to dissipate heat,” he noted. “By contrast, systems that use cooling towers or other evaporative methods consume water during normal operations.”

In a typical closed-loop system, the cooling circuit is filled with a mixture of water and propylene glycol continuously circulating, somewhat like a car radiator, to transfer heat away from the IT equipment. Evaporative cooling technologies such as towers are used in power generation plants and some data centers.

Continuous freshwater demand is taxing community systems and worrying residents about the availability and affordability of water and electricity services for themselves, their families and businesses. The threat is strong enough that hyperscalers such as Meta are working on circular approaches such as water recycling, outside-air cooling and the closed-loop liquid cooling with dry coolers.

“The goal should be system-level resource efficiency, considering water, energy, carbon, local resource availability and reliability together,” Kelvion’s Cubillan said. “The best solution is not necessarily the one that uses the least water in isolation, but the one that manages the facility's overall resource requirements responsibly.”

Yes, AI is putting pressure on electricity and water resources as server rack density and heat generation rise with supercomputing capacity. But, as Cubillan noted, AI also can help.

“AI can actually help optimize the infrastructure that supports AI,” he added. “Data centers are dynamic environments, and cooling systems do not need to operate at maximum capacity at all times. AI and advanced controls can analyze variables such as workload, temperature, humidity, airflow and equipment performance and adjust cooling systems accordingly.”

AI is an ultimate next-gen challenge for managing natural resources, perhaps, but building it does not change the planning dynamic builders have used for ages.

“Ultimately, the right cooling solution needs to balance performance, reliability, resource availability, and total cost for the specific project,” Cubillan said. “The earlier these considerations are addressed, the more flexibility developers have to choose a design that meets computing needs without creating unnecessary pressure on local resources.”

The future of cooling in AI compute

Front and center at Data Center Frontier Trends Summit

Co-located May 5-7 with Microgrid Knowledge Conference in Glendale, Arizona