Updated on July 16, 2026 at 3:41 PM ET
There are two competing schools of thought regarding how water-intensive AI will consume. For one thing, technology horribly I’m thirsty. Data centers will worsen droughts across the country and “drain the Great Lakes.” Popular threads on Reddit Put it down. Former Congresswoman Marjorie Taylor Greene took up the issue, post sarcastically At one point, he says, “How often do farmers complain that data centers are stealing their water and raising their electricity bills?”
The other school, mostly Silicon Valley types, has an article in a right-wing publication claiming that “the data center water crisis is not real.” pirate wire claimed In December. OpenAI CEO Sam Altman says the water crisis is ‘totally fake’, ‘insane’ and ‘has nothing to do with reality’ said A few months ago.
As with many discussions about AI, the arguments are not always presented honestly. But unlike, for example, ambiguity about how much work chatbots can or should be trusted to do, it seems to me that water usage should be measurable with clear facts supporting either side. That’s not the case. Does AI waste a lot of water? Not necessarily. sometimes. It really depends. “We don’t have real, truthful numbers to get everyone on the same page about water usage in data centers,” said Eric Massanet, a sustainability researcher who studies data centers at the University of California, Santa Barbara. Neither data center operators nor power companies provide much insight. Any public information that exists can be contextualized, recontextualized, and distorted to suggest basically anything.
But water issues have very significant implications not only for the well-being of residents in every county where new data centers are being built, but also for the future development and regulation of AI itself. Just this week, New York Governor Kathy Hochul signed During the nation’s first state-wide data center moratorium, her office noted that ultra-large data centers can require “large amounts” of water. In drought-stricken counties or during heat waves, data centers can place a significant strain on local water supplies. If water is plentiful and utility companies can provide new infrastructure, data centers may be of little concern.
Outside of water-stressed regions, “it’s not really a big concern right now,” said Fengqi Yu, an energy systems expert at Cornell University. “But as this field grows, you never know.” Turning a particular case into a blanket judgment creates little emotion and confusion, but heightens it. As ever, the truth is far more nuanced and conditional than activists and tech executives would like to admit.
The problem begins with frame composition. How much water do data centers use? Not a particularly useful question. On a national scale, these buildings use an astonishing amount of water. In 2023, data centers used just over 17 billion gallons of water for cooling. This is equivalent to the amount used by approximately 500,000 Americans in a year. This may seem far out of context, but this is less than one-tenth of a percent of the total amount of water the world uses. usa farms That year.
The situation may be different at the county level. The Meta data center campus under construction in Lebanon, Indiana, could require 8 million gallons of water per day at peak times. That’s a drop in the bucket for New York City, but more than double the town’s peak demand. Every data center requires a clear assessment, and conditions can change rapidly. Water usage depends not only on the actual data center design, but also on the local climate, local water supply, and local power grid configuration. Regarding Meta’s Indiana data center, city officials said: said It said the planned upgrades meant there was “no indication” the water would run out. Meanwhile, in Newton County, Georgia, local officials said He argued that all of the data centers that were going to be moved in “simply don’t have water,” and that the county must “compete” to improve local water recycling facilities.
Rather than worrying about the amount of water, it’s best to ask how your particular data center operates. Purpose water. In general, water is always used for the same purpose: to keep computer chips in buildings cool, but there are many ways to do it. Broadly speaking, data centers can exhaust the heat produced by chips using cooling towers that evaporate water, or air-cooled chillers or similar technologies that use electricity to expel hot air. Think of cooling towers as passively cooling us as our sweat evaporates, and chillers as car radiators that use chemical coolants and physically push heat out by pushing air through them.
The AI industry has settled on a car radiator approach for many of the largest and most controversial data centers. It runs water directly through the server to remove heat, circulates that water through an air cooler, and then, very importantly, reuses it. OpenAI’s Stargate data center in Texas, Meta’s Hyperion data center in Louisiana, Microsoft’s Fairwater data centers across the country, and many other large facilities operate this way most of the year. In such a closed-loop system, no water is lost through evaporation. This is likely what Altman was referring to when he called concerns about water “insane.” Once these data centers have water, they theoretically don’t need any more. That’s because the water never runs out.
The real motivation behind closed-loop systems is convenience. Building regional water infrastructure or drawing water to reservoirs can take years, the use of groundwater is controversial (groundwater doesn’t replenish quickly and can contaminate the environment when used), and drawing water from Lake Michigan to data centers is practically impossible. dusty arizona. Water availability can be a critical bottleneck for new data center construction. A closed-loop system with an air cooler avoids these problems.
So why don’t all data centers use chillers? First, air coolers require significantly more power to pump out the same amount of heat compared to cooling towers. “Can a data center use zero water? Yes, it’s very easy,” Shaolei Ren, an AI and sustainability researcher at the University of California, Riverside, told me. “But on the downside, the cost you have to pay” is that you use more power in different ways. That means 10 to 65 percent more. estimate. 19 gigawatts facility If we consider Utah, it would require about the same amount of electricity as New York City.
The almost unrealistic power demands of data centers are primarily met by combustion turbines, many of which are combustion turbines. Built by the tech companies themselves (To avoid waiting for upgrade to local grid). And water is also needed to generate electricity, including from private power plants that burn fossil fuels. in 2024Meta’s “indirect” water consumption (referring to water used during power generation) was 19 billion gallons. This was 23 times more than direct water consumption, most of which went to the data center.
Few industries consistently report indirect water use. One reason is that many of the estimates are flawed. Indirect utilization is obtained by averaging the water usage of all power sources on the regional grid, including hydropower plants that consume large amounts of water through evaporation from dams and lakes. In other words, indirect water consumption figures may underestimate or overestimate the contribution of individual plants. Paradoxically, the AI industry’s decision not to use water to dissipate heat within data centers may increase overall water usage for some facilities. “If you use water on-site to increase cooling efficiency, you will use less electricity,” said Jonathan Coomey, a data center and sustainability researcher. “It’s not a simple problem Poor use of water on site”
It is possible to design data centers that reduce both power and water usage. For example, in cooler climates, data centers can bring in outdoor air. This is what Amazon is doing at its massive data center in northern Indiana. In the summer, spraying a small amount of water into the outside air can reduce the temperature to a similar degree. Data centers are also beginning to experiment with running chips at higher temperatures, reducing overall cooling demands. And of course, a fairly obvious way to build a data center with less environmental impact is to choose renewable energy sources that consume very little water instead of natural gas turbines.
Until this approach becomes the norm, people will understandably worry about the impact these sprawling facilities will have on their communities. Assessing how data centers will impact nearby reservoirs ultimately requires the most scarce resource about these facilities: accurate, accurate information. Technology companies want to get their data centers up and running quickly, so they contract with: dummy company Negotiate behind closed doors and under non-disclosure agreements. Residents may only learn of new construction when shovels dig into the ground.
Already, rising temperatures and prolonged droughts are beyond anyone’s control. Now, large industrial facilities are arriving under cover of night, which may or may not exacerbate many of the problems caused by the climate crisis. This sense of confusion goes a long way toward explaining why the environmental impact of data centers has become one of the key battlegrounds in the AI backlash. Often the main problem is not a lack of water, but a lack of control.
This article originally mentioned a facility under construction in Utah. In fact, the proposal is still under consideration and construction has not begun.
