Public policy debates often revolve around questions that frequently lack all the necessary information needed. We fight over the impact of illegal immigration without knowing how many undocumented people are actually in the country. We alternately praise or condemn the police without knowing how much crime law enforcement prevents.
Concerns over data centers are not like this. Their actual power and water usage is knowable down to the last megawatt and gallon. But debate rages nationwide about the burden these facilities place on local infrastructure because the most basic and definitive measure of their impact — their utility billing records — frequently are treated like state secrets.
“Utility data is considered to be private,” said Richard Michelfelder, a finance professor at Rutgers University’s Camden, N.J. campus and a former utility executive and consultant. Michelfelder said patterns in power and water usage can reveal when someone is home or what’s happening inside a commercial building. “It’s competitive information,” he said.
Knowing it would be a challenge, Collier’s attempted to obtain the electricity and water billing records, sampling 12 data centers throughout the nation.
The centers are in Arizona, Georgia, Indiana, Iowa, Nebraska, Nevada, North Carolina, Ohio, Oregon, Tennessee, Texas and Virginia. Three of them belong to Meta, four to Google and one each to Microsoft and Elon Musk’s xAI. The rest are associated with lesser-known data center operators, EdgeCore, IREN and Digital Realty.
Fifteen of the power and water utilities queried refused to release any information, citing customer privacy. Critics find the secrecy frustrating.
“Data centers aren’t [just] any customer,” said Clara Vondrich, senior policy counsel for the climate program at Public Citizen. “They’re not a customer like you or I whose records and privacy might arguably be protected. These are behemoths that are destroying communities, and in that light, they should be treated in a separate class, and that class needs to disclose its energy and water use.”
Ironically, their counterparts on the other side also would like to see greater transparency from data centers. They want it to prove the opposite point: that the controversial facilities aren’t an excessive or unreasonable drain on resources.
“I think a lot of people are really being blinded by just the raw scale of the numbers involved. They’ll be like, ‘Millions of gallons of water!’” said Andy Masley, an independent researcher who argues that the panic over the local impacts of data centers is overblown. He said if the billing data were released and compared to that of other industries, “that would kind of seal the deal for my case.”
Faced with growing tensions, state and federal policymakers have recognized the need for more clarity, especially given data centers’ frequent use of non-disclosure agreements with city officials.
Numerous proposals this year sought to require existing or prospective data centers to report their water and/or power usage, but ultimately only three were enacted: an executive order by Texas Gov. Greg Abbott and bills approved by the Nebraska and Utah legislatures, according to research by Alex McWard, senior policy specialist for the National Conference of State Legislatures.
That might not be enough to satisfy a growing chorus of angry citizens who feel data centers have been forced onto their communities. Abre’ Conner, director of the NAACP’s Center for Environmental and Climate Justice, said the demand for greater transparency around data centers is so intense right now that elected officials could start getting voted out of office, beginning with the midterms this November.
“I think that is a real possibility,” she said. “Where there have been special elections, that is literally what is happening. People are literally being voted out of office when they are not listening to people.”
Data centers are essential to the artificial intelligence revolution just beginning to thoroughly remake modern life, and that evolution is likely to take some time. Meanwhile, there appears to be little debate that the typically hulking, windowless buildings generally use a lot of electricity to power the computers they house as well as a lot of water to cool themselves when those computers get hot.
Where things start to become unclear is when you try to unpack what “a lot” might mean in this context. There’s a debatable gap between consuming resources on a large scale and consuming resources so much that it puts pressure on the overall supply.
Both sides come equipped with studies and statistics to bolster their case.
Critics, for example, point to academic research from 2021 that found roughly 20% of the nation’s data centers draw water from watersheds already moderately to highly stressed by droughts or other factors, suggesting that in an era of climate change data centers are consuming a potentially endangered critical resource.
Supporters, on the other hand, refer to a 2024 study by the state of Virginia, the data center capitol of the world, which reported that 83% of data centers use less than or the same amount of water as an average large office building — 6.7 million gallons per year.
Jen Springman, the coalition’s manager of the Goldwater Institute, a right-of-center think tank firmly in the data-centers-are-not-a-danger camp, told Collier’s that thanks to the strength of their side’s research there’s a general consensus amongst policy professionals and industry insiders that “the water issue” is “taken care of at this point.”
Whether critics agree is another question, and, even if they do, it only addresses half of the concerns. Springman conceded data centers “use a large amount of electricity. That’s not a lie.”
Indeed, that same Virginia study cited as proof that data centers aren’t so bad also reports that the facilities “consume substantially more energy than other types of commercial or industrial operations,” lending credence to critics’ contention that data centers drive up energy costs for individual consumers.
The study describes a “smaller” data center recently constructed in the state that can draw up to 18 megawatts of power. That’s roughly equal to one mid-sized automobile assembly plant, 60 large office buildings or 4,500 homes.
The largest data centers, known as hyperscale data centers, can draw 100 to more than 200 megawatts each. Hyperscale data centers are used to train and run AI models like the now-ubiquitous ChatGPT. Traditional data centers, which are typically smaller, support local businesses and private corporate networks.
Springman said that electricity concerns for both hyperscale and traditional data centers can be addressed by increasing the nation’s transmission capacity, an endeavor that she said has been burdened by needless red tape. The electricity demands, however, also reveal another subtle wrinkle in the raging public debate: not all data centers are the same.
“There’s at least two different types, the traditional data centers and the hyperscale data centers,” said Chris Brown, chief technical officer of the Uptime Institute, a global authority on data center performance. “They’re different because of the industries and function they serve, the workloads that they’re housing, the clients they serve and the purpose that they’re undergoing.”
Brown said data centers can be distinguished by their workload or rack density, which measures the amount of electrical power a single server within the center uses at any given moment.
A traditional data center, which may have 50 to 1,000 racks, may have a rack density of 10 to 20 kilowatts. A hyperscale data center, which will have hundreds or thousands of racks, may have a rack density of 60 to 300 kilowatts or more.
While industry insiders understand these differences, Brown said members of the public may be confused about proposed or existing data centers in their area.
“People don’t have a good idea what data centers are because we as an industry haven’t really been teaching them enough about what they do and how they work,” Brown said. “They hear about one style of data center and suddenly every data center is that way. It’s like if all you ever saw was a Cadillac, then every car is a Cadillac.”
Collier’s attempt to bring more clarity to the discussion was hindered by the 15 utilities that refused to release information on consumer privacy grounds. But that wasn’t the only obstacle.
Two utilities queried didn’t respond by press time. Memphis Light, Gas and Water, which uniquely provides both water and electricity to Musk’s Colossus 1 data center, also denied Collier’s request because the requester is not a Tennessee resident, which is allowed under the state law there.
The city of The Dalles, Oregon, home to a Google data center, required $750 to pull six and half years of monthly water billing records, then wanted $125 for just two months of information. Such fees, justifiable or not, create a sizable financial obstacle to citizens researching the issue.
The county of Fayetteville, Georgia, the site of a Microsoft Fairwater AI data center, provided roughly two and a half years of records for $46.12. Loudoun Water, the water utility for the Digital Realty IAD39 data center in Loudoun, Virginia, provided six and a half years of records free of charge. So did the city of Henderson, Nevada, which is home to a Google data center.
Columbus Water & Power in Ohio, which provides water to a Meta data center, also provided two years of water billing records for free.
In the end, no billing records from any power utilities were accessible and only those four water utilities complied with requests. Each complying utility, provided monthly water bills for multiple years. Google, Meta and Digital Realty didn’t respond to requests for comment. Microsoft only responded with links to three of its own stories about data centers.
The bills showed a wide disparity in the daily amount of water each of the four facilities used on average — compounded by the size of the facilities and how long they’ve been in operation.
The Microsoft data center operation in Georgia, which is only a couple of years old and encompasses four buildings, uses 126,451 gallons of water, according to an analysis of the billing records provided.
On the other hand, the Google data center in Nevada, Meta’s New Albany Data Center in Ohio and the Digital Realty data center in Virginia are older, but the Google operation compasses six buildings, Meta’s four and the Digital Realty center only one.
The Google campus uses on average 633,010 gallons daily, while the Meta campus uses 69,941 and the Digital Realty building uses 4,795, according to analyses of the provided billing records.
Considering these are daily averages, even Digital Realty’s usage appears like it could be excessive. But are they really? We asked Shaolei Ren, an electrical and computer engineering professor at the University of California, Riverside who studies data centers, to look at the numbers.
He was definitive in his assessment.
“The daily average number looks reasonable,” he wrote in an email. “The huge differences in the water usage numbers reflect different cooling systems, climate types and data center sizes.”
Ultimately, the AI data center power and water usage debate remains difficult to define. With inconsistent regulations and reporting requirements on cooling systems, data center types, power sources, size of centers and clear indications of the speed of implementation of emerging technology only regulations and transparency compliance will provide enough data to afford comprehensible and factual discussions. For now, AI data centers do not provide enough data on that.
