Understanding AI’s Thirst for Water: An Explainer

Digital Colonialism, a series co-produced by NPQ and MediaJustice, explores how the rapid expansion of artificial intelligence (AI) data centers is reshaping communities across the United States.

In October 2025, Food & Water Watch, a national environmental organization, called for a national moratorium on “the approval and construction of new large-scale data centers,” until enough information about their effects has been collected, analyzed, and disseminated. It was driven in part by the Trump Administration’s AI Action Plan and by big tech companies’ lack of transparency about data centers’ true energy and water consumption.

Although data centers have been around for decades, built originally to support general computational tasks such as hosting websites, database management, and cloud services, generative AI requires even more massive, resource-intensive infrastructure that Big Tech is rushing to build. These hyperscale centers, used for modeling, training, and running AI, often require almost double the power of traditional data centers and pull more resources than are available.

Big Tech executives claim that these hyperscale data centers are the future of America. But advocates against the centers, like AI ethicist Masheika Allgood—founder of AllAI (pronounced “ally”)—believe the current race to build data centers is driven solely by monetary gain.

How Data Centers Are Cooled 

Data centers have traditionally used air-conditioning to cool their clusters of interconnected server racks, which house the processors and generate large amounts of heat. These racks are organized into hot and cold aisles. The servers face the cold aisle where cool air is circulated, and the heat pulled by the cool air is vented into the hot aisle that faces the air conditioner’s return ducts.

Big Tech executives claim that these hyperscale data centers are the future of America. 

Traditional data centers use central processing units (CPUs), which can handle general business applications and data storage. AI data centers use AI accelerators or AI chips, including graphics processing units (GPUs). A GPU is the hardware that powers software and accelerates its operations. In comparison to CPUs, GPUs and accelerators handle much larger and more complex workloads, requiring them to run longer, be more energy-intensive, and use more water for cooling.

In more “energy-efficient” centers, server racks are cooled by cold air, and there’s no direct water usage. These systems are connected to a chilled-water unit, which companies like Google promote as more efficient, sustainable, and water-friendly, because water-cooled data centers consume around 10 percent less energy than air-cooled centers. But in new data centers powering larger AI and generative AI workloads, air cooling alone can’t support the growing capacity.

These newer AI data centers use two main forms of liquid cooling: evaporative cooling and closed-loop system cooling. In evaporative cooling, cold water is poured on a hot surface, and heat is extracted through the steam, leaving behind cooled pipes of liquid. In this system, the evaporated water can’t touch the servers directly, as the moisture can corrode the equipment.

In a closed-loop system, two pipe loops are connected by a heat exchange unit, which allows cold water to constantly be running through the server racks. One loop runs through the racks, carrying cool water that extracts the heat from them. A second loop pulls heat out via the heat exchange unit, carries it to the evaporative cooling system, and then recirculates the cold water back to the servers through the first loop.

In some systems, that first loop may also carry a liquid coolant consisting of PFAS, also known as “forever chemicals,” which could have adverse health effects for surrounding communities, Allgood explained in an interview with NPQ. Additionally, some centers use a method of immersive liquid cooling where fresh cool water is sprayed directly onto hot spots in the chips and the hot water is then taken out. It’s unclear if that hot water is contaminated with any chemicals from the chips. It’s also unclear where that hot water goes—whether it’s recirculated into a cooling system and reused or dumped out.

Many environmental advocates oppose these evaporative cooling systems because evaporated water can’t be recovered and reused. And while some have characterized closed-loop systems as a more environmentally friendly alternative, Allgood explained that even though some of that water is being reused, the two loops do not actually minimize water loss.

“Closed loop just means that one of those loops never leaves the system,” Allgood said. The second heat exchange loop isn’t actually a closed-off loop, as water does eventually evaporate.

Allgood views this as a greenwashed solution—this type of system can’t function without a heat exchange unit because there wouldn’t be another way to extract the heat without the second loop. And that requires evaporative cooling.

Air cooling could be an alternative, but that wouldn’t meet the needs of hyperscale data centers, Allgood explains. The AI chips in those server racks can reach temperatures up to 194 degrees Fahrenheit per chip. When hundreds of thousands of chips are packed into hundreds of server racks, liquid cooling is the only solution.

The Cost of Data Center Cooling

A 2024 report from the Lawrence Berkeley National Laboratory found that in 2023, data centers in the United States consumed over 17 billion gallons of water for cooling. Of these, 84 percent—almost 15 billion gallons—was consumed by hyperscale data centers. Those figures are expected to increase; by 2028 data centers could consume up to 33 billion gallons of water for cooling.

And these estimates don’t account for the more than 211 billion gallons of water US data centers consumed indirectly through electricity use.

Currently, 40 percent of the proposed and standing data centers are located in places that are already water-strapped and at high risk, mostly because of the ample land, electricity, and flexible incentives.

In 2024, Microsoft announced it had figured out a “zero-waste solution” for cooling data centers through closed-loop systems that would reduce their water usage by 39 percent, saving about 33 million gallons of water annually. This was based on data Microsoft compiled about itself.

In the company’s 2024 Environmental Sustainability Report, Microsoft’s annual water consumption in 2023 was listed at over two billion gallons. Google, in its 2024 Environmental Report, said that its data centers consumed over six billion gallons in 2023.

Both Microsoft and Google use closed-loop systems, but their water consumption remains immense. Allgood pointed out that these companies don’t mention the heat-exchange loop publicly. “That’s just a lie by omission, and it’s intentional, [and] it’s deceiving,” she said.

So, whether the cooling system is air- or loop-based, a massive amount of water is constantly required to keep an AI data center functioning smoothly and safely.

“At the end of the day, there are no good solutions for cooling chips that run that hot at that scale of deployment,” Allgood emphasized. Unsurprisingly, tech companies are targeting areas that might be more business-friendly and have lower electricity costs, regardless of the water impacts.

Where Data Centers Are Turning Their Sights

Currently, 40 percent of the proposed and standing data centers are located in places that are already water-strapped and at high risk, mostly because of the ample land, electricity, and flexible incentives. However, Big Tech is turning its sights on regions such as the Great Lakes Basin, thanks to its cooler climate and proximity to large freshwater sources. This could affect the availability of surface water, which is already in high demand: The Great Lakes provide drinking water to more than 40 million people in the United States and Canada, in addition to many other uses, from agriculture to thermoelectric power production. And data centers competing for these resources will likely strain groundwater from local aquifers as well.

A report commissioned by the Joyce Foundation in partnership with the Freshwater Society, Water365, the University of Minnesota, and the Great Lakes Indian Fish and Wildlife Commission found that nearly every state in the Great Lakes has one or more locations facing a groundwater shortage. Climate change has affected those aquifers’ ability to recharge, and with the warming temperature and increasing severe weather events, “these trends are likely to accelerate.” And the region’s Indigenous communities will bear the brunt of these water shortages in part due to the rapid expansion of data centers, as well as existing water withdrawal policies.

Michiganders are already paying among the highest water bills in the country. In February 2025, the Great Lakes Water Authority, one of two primary agencies that manage water services in southeast Michigan, announced the largest water rate hike for fiscal year 2026, which residents later challenged.

Although people have been vocal, the conversations generally aren’t focused on community impacts. “When officials talk about water, it’s from a business standpoint and how the data center will be able to access water. Not from a community standpoint,” Allgood said.

One of Allgood’s recommendations is pursuing legal action, as many communities and organizations—including the NAACP and —are doing. “There’s a mountain of evidence and research that can be used that shows that these data centers are highly consumptive of water and electricity,” Allgood said.

Whether or not communities understand the science of data centers, Allgood said, she encourages taking action: “People think you go to court because you have everything. You go to court to get everything.”