No, there is not plenty of water for data centers
And, yes, we should worry about it, along with the facilities' power use
📈 Data CENTER Dump 📊
When I first read a recent headline in Matthew Yglesias’s Slow Boring newsletter, I assumed it was a sort of joke to rope me into reading. “There’s plenty of water for data centers,” it said, reassuringly. “Probably the last worry you should have about either water or AI.”
Unfortunately, he wasn’t joking. But he opened his piece with a line that should have warned his readers to take everything else he said with a grain of salt:
“I’ve spent my whole life in the urban northeast, which is both soggy and light on water-intensive economic activity, so I haven’t spent a lot of time thinking about water use or water scarcity.”
Before I continue with my rant, I’d just like to encourage Yglesias to do a little more thinking about water scarcity before writing about it. Oh, and also, maybe consider spending a little bit of time in the water-starved West before committing punditry about it. (This is the same guy who tweeted that Sen. Mike Lee’s proposal to sell off public land was “pretty reasonable” and an “okay idea on the merits”).
Yglesias acknowledges that data centers use water, and that more data centers will lead to more water consumption. But it’s okay, he says, because “We’re not living on Arrakis, and rich countries are not, in general, abstemious in their water usage.”
No, we are not on Arrakis, but have you seen the lower reaches of the Colorado River or even the mid-reaches of the Rio Grande lately? It’s looking pretty Dune-like if you ask me.
Well, sure, Yglesias argues, but even in those places, people are doing frivolous things with water, like filling up their Super Soakers or using it to make ice cubes for their cocktails. Yes, he used those actual examples. Never mind that the potable water used each day by a single Microsoft data center in Goodyear, Arizona, could yield more than 35 million ice cubes or fill about 223,000 Super Soakers. That would be one big, drunken water fight.
Yglesias also notes that agriculture, especially growing alfalfa and other feed crops for cattle, is an even larger water consumer than Big Tech. True, for now. And he writes:
“But I think the bottom line remains that water is not generally scarce and that coming up with new economically useful things to do with water is good. … Water is sufficiently plentiful in this country that we can use lots of it for lawns or data centers or alfalfa or golf courses or whatever else we want, while still giving people plenty to drink.”
His logic appears to be: People are currently using a lot of water for all sorts of things — frivolous or otherwise. So, it should be fine to use a lot more water for data centers in perpetuity, since water is “sufficiently plentiful.” This is the sort of thinking that got the Colorado River Basin into its current mess, in which there actually may not be enough water to drink very soon if its collective users don’t change their ways. Adding a fleet of water-guzzling hyperscale data centers to places like Phoenix, Las Vegas, and Tucson, where water is anything but “sufficiently plentiful,” will only exacerbate the crisis.
Researchers have tried various methods to determine how much water a single ChatGPT query or AI-assisted Google search uses as compared to, say, streaming a Netflix video or writing a standard e-mail. So far the estimates diverge wildly. An early calculation came up with a whopping 500 ml for each AI query, but the estimates have since gone down. The difficulty is due in part to the fact that water use data isn’t always publicly available, and also because data centers’ water use can vary depending on location, as do their carbon footprints.
What is clear is this: Data centers use large quantities of both energy and water, no matter where they are. The massive server banks churning away in warehouse-like buildings on the fringes of Phoenix and Las Vegas, and even in rural Washington and Wyoming, each gobble as much electricity as a small city to process AI queries, cryptocurrency extraction, and other aspects of our increasingly cloud-based society. The harder they work, the hotter they get, and the more power and water they need to cool off to the optimum operating temperature of between 70° to 80° F.
Evaporative or adiabatic cooling, where air is cooled by blowing it through moistened pads (i.e. high-tech swamp coolers), works well in arid areas like Phoenix, Tucson, or Las Vegas. They use less energy than refrigerated cooling, but also use far more water.
Data centers can also indirectly consume water through their energy use, depending on the power source. Thermal coal, nuclear, or natural gas plants need water for cooling and steam-production (some of this water may be returned to the source after use, except with zero-discharge facilities); natural gas extraction uses water for hydraulic fracturing; and solar installations can require large amounts of water for dust-suppression and cleaning. This explains how Google’s data centers withdrew 8.65 billion gallons of water globally in 20231.
A 2023 study found that a single Chat GPT-3 request processed at an Arizona data center uses about 30 milliliters of water, compared to 12 ml per request in Wyoming. That doesn’t seem like much (it’s less than a shot-glass) until you consider that there are at least 1 billion ChatGPT queries worldwide per day and growing, using a total of some 8 million gallons of water daily, worldwide. And, training the AI at an Arizona data center would use about 9.6 million liters — or 2.5 million gallons — of additional water.
Another estimate finds the average data center uses between 1 million and 5 million gallons of water per day, onsite, which would be far more than the aforementioned Goodyear center (56 million gallons/year), but in line with a planned Google data center in Mesa, Arizona. When Google was first planning the facility back in 2019, the city of Mesa guaranteed delivery of nearly 1 million gallons of water per day. If they reach certain milestones they can use up to 4 million gallons daily, or about 4,480 acre-feet per year.
Now multiply those numbers by the more than 90 data centers of various sizes and water and energy intensity in the Phoenix area, alone, which would amount to somewhere between 14 million to 450 million gallons per day. No matter how you add it up, they collectively are sucking up a huge amount of water and power, and enough to strain even Yglesias’s purported “sufficiently plentiful” supplies (which do not exist in Arizona, by the way).
The average Phoenix-area household uses about 338 gallons of water per day, or almost 123,000 gallons per year. One of these big data centers, then, could guzzle as much water as some 10,000 homes. And yet housing developments in groundwater-dependent areas on Phoenix’s fringe must obtain 100-year assured water supply certification before they can begin building. The same is not the case for data centers.
According to Open ET maps, a 75-acre alfalfa field in Buckeye (western Phoenix metro area), uses about 156 acre-feet — or 50.8 million gallons — per year. That’s far less than the 28-acre Apple Data Center in Mesa consumes. Of course, there are the equivalent of about 3,470 alfalfa fields of that same size in Arizona (260,000 acres), meaning the total water consumption of hay and alfalfa is still greater than that of data centers. But it shows that while replacing an alfalfa field with houses would result in a net decrease in water consumption, replacing those same fields with data centers would substantially increase consumption.
And don’t forget that the 75-acre alfalfa field produces about 690 tons of alfalfa per year, which could feed quite a few dairy cows, which in turn would produce a bunch of milk for making cheese and ice cream. Just saying. Maybe it’s time to update the old saying: “I’d rather see a cow than a data center.”
Data centers aren’t going away. After all, they are the hearts and brains of the Internet Age. Many of us may wish that AI (not to mention cryptocurrency), which are more water- and energy-intensive than other applications, would just up and vanish. But that’s probably too much to ask for. Besides, AI, at least, does have real value.
So what can be done to keep the data center boom from devouring the West’s water and driving its power grid to the snapping point? Here’s where Yglesias had a good point: Policymakers and utilities should adjust water and power pricing for large industrial users, i.e. data centers, to discourage waste, incentivize efficiency and recycling, and push tech firms to develop their own clean energy sources to power their facilities.
It’s imperative that utilities force data centers to pay their fair share for infrastructure upgrades made necessary by added water or power demand, rather than shifting those costs to other ratepayers, as is usually the case. Arizona should make data centers prove out their water supply, just like they do with housing developments. Plus, states should stop trying to lure data centers with big tax breaks, which ultimately are paid for by the other taxpayers. And local governments and planners should subject proposed data centers to the highest level of scrutiny, and not give in to promises of jobs and economic development if it means sacrificing the community’s water supply or the reliability of the power grid.
Proper policy isn’t a cure all, by any means. But it could mitigate the impacts of the imminent data center boom. Meanwhile, Mr. Yglesias, I will reiterate that the West, at least, does not have plenty of water for data centers, and I will continue to worry about them guzzling up what little water remains.
📖 Reading Room 🧐
The Land Desk is reading all of y’all’s great responses to last week’s open thread about forms of resistance. Check it out and weigh in if you haven’t already.
Len Necefer has had some really strong pieces on his All At Once by Dr. Len newsletter recently, including this one musing about the opportunities for the Navajo Nation to build a recreation economy on the San Juan River (great idea!). He writes about how strange it is that he, a Navajo Nation citizen, must get a permit from the BLM to raft the river, when it borders his homeland (and is at the heart of Diné Bikeyah). I also like that he sees boating/recreational opportunities along the entirety of the river, not just from Sand Island to Clay Hills Crossing. I’ve always thought it would be super cool to boat the reaches between Farmington and Bluff (actually, I’ve always wanted to boat from Durango to Farmington to Bluff).
Another Substack that’s been getting my attention is Time Zero, a podcast and Substack on “the nuclearized world.” The Wastelanding series is about the legacy of uranium mining and milling on the Colorado Plateau, the Navajo Nation, and on Pueblo lands. Very powerful stuff.
The Colorado Sun’s Shannon Mullane has a good story about the Southern Ute Tribe finally getting some of its Animas-La Plata water, which was the whole reason the last big Western water project, as it’s known, was finally built.
This is not the same as consumption, which is the amount of water withdrawn minus the amount returned to the source.
Yep. Dumb AI.
Anyway, for years, working in energy efficiency we promoted evaporative heat rejection instead of, or to reduce, electric compressor cooling. But it is possible to cool anything, anywhere w/o evaporation. Thermal energy storage also works to level out cooling demand, and was common in big campus systems a few decades ago. Vegas & Henderson banned cooling towers and evap cooling two years ago. IMHO, there is a lot more of an "abundance" of land for solar electricity than an "abundance" of water for cooling in the SW.
P.S. As far as I know, in the SW US, most cleaning of PV panels is now waterless. (Brushing and sweeping instead.) In some locations cleaning is just not needed if a small output decrease is OK. The Solar Industry Assoc lists 0.02 gal/kWh as an average, 1/25th of coal/nuke 0.5 gal/kWh.
Welp, there are always coal-fueled power plants🙄