The Coming Tsunami of AI Data Centers (businessinsider.com)
Chatbots may live in the cloud, but they’re powered by massive concrete boxes — and they’re coming to a town near you.
Pageland Lane is a roughly 5-mile-long, one-lane road in Prince William County, Virginia, an hour south of Washington, DC. To the east it hugs Manassas National Battlefield Park; to the west lie thousands of acres dotted sporadically with homes and miles of pristine farmland, with grazing horses and tight round hay bales lazing in the sun. It all looks plucked from a postcard. It’s also, Elena Schlossberg repeatedly reminded me while driving me on a tour of the area, very likely doomed.
For the past few years, this community next to one of America’s most historic battlegrounds has been the site of a new kind of civil war. Two deep-pocketed tech companies, QTS and Compass, are seeking to cover some 2,100 acres of rural land with 23 million square feet of data centers — one of the largest such developments in the world. Just what they’ll be used for — cloud storage, AI, or otherwise — is still an open question. But by some estimates, the Prince William County Digital Gateway, as the project is known, would suck up as much as 3 gigawatts of energy, enough to power millions of homes around the clock.
For Schlossberg, a self-proclaimed “big-mouth mom” who runs the grassroots group Coalition to Protect Prince William County, the Digital Gateway represents not just an invasive eyesore but also an existential threat. All over the world, data centers are multiplying at record rates — and growing bigger and more power-hungry by the year. Inside their boxy bellies is the lifeblood of the internet, miles of cables connected to racks of computers that fuel our telehealth appointments and family group chats, our Netflix binges and Google Docs. In the process, data centers gorge on electricity and guzzle exorbitant amounts of water to keep cool, often overburdening local power grids and water supplies.
The arms race over artificial intelligence, which runs on special chips that suck up even more energy, has accelerated the boom. In the first half of 2023, North America set a record for data-center construction, up 25% in the top eight markets. And that’s just for data centers that lease out server space to other businesses. Tech giants such as Google, Microsoft, Amazon, and Meta — so-called hyperscalers — are spending billions of dollars more to build out their own data centers. Jensen Huang, the CEO of the AI-chip giant Nvidia, predicts that in the next four years alone, companies will spend $1 trillion to add to their arsenal of data centers — creating what Blackstone, one of the world’s largest asset managers, calls a “once-in-a-generation engine for future growth in data centers.”
But that growth may come at a steep cost. At the heart of the data-center boom lies a strange paradox: The more the internet has consumed our lived reality, the easier it’s been to ignore the physical infrastructure required to power that reality. Today, the complexity of the large language models being developed by OpenAI, Google, and Meta — and the frenzy to bake those models into everything from Google Search to Facebook stickers to “Harry Potter” fan fiction — is forcing more and more Americans like Schlossberg to confront the high price of our digital addiction.
“People need to know that every picture, every TED Talk, every Instagram, everything they save is going into a concrete box that requires power,” Schlossberg said. “It’s not free, and it’s not ethereal, and it’s not fluffy.” For all of the fears about AI’s brain — that chatbots will steal our jobs or somehow destroy humanity as we know it — it’s AI’s body that could claim us first.
For most of human history, doing almost anything required having and storing a lot of stuff. Books to access information. A Rolodex to call clients. Vinyl, then tapes, then CDs to play music. Dusty photo albums to recall old memories, and bulging filing cabinets to retrieve old invoices. For the most part, we knew where our stuff lived because, for the most part, it lived with us.
When the internet came along, a lot of that stuff seemed to disappear. It was stored somewhere we called the cloud, an abstract, celestial space where we never had to clean out our closets and that miraculously delivered our stuff to us with a click or a swipe or no prompting at all. “You have a new memory,” our iPhones tell us. At first, the transition was disorienting. People grasped for some understanding of how the internet physically functioned. In a now classic segment of the “Today” show in 1994, the befuddled hosts puzzled over how email worked. “What, do you write to it, like mail?” Bryant Gumbel asked. “Allison, can you explain what internet is?” Katie Couric pleaded with an apparently tech-savvy off-camera producer.
But today, three decades later, how many of us could actually explain to Couric “what internet is”? As the web became the water we all swim in, the impulse to ask, “What the hell is water?” disappeared. At a certain point, it all just started to feel invisible. And perhaps no technological advancement of the past few decades has felt like more of a magic trick than generative AI. Type a prompt into ChatGPT and — alakazam! — it will just as easily spit out a five-year plan for your fintech startup as it will a bedtime story on Santa Claus’ first visit to the moon, written in the style of Upton Sinclair.
Like any good magic trick, the internet is based on an illusion. Our stuff hasn’t vanished — it’s just moved to data centers. Crammed with everything we once kept close at hand, these massive storage facilities fill campuses the size of small villages and are powered by squat substations that occupy even more land, which are connected to lengthy transmission lines that funnel power from faraway coal plants, nuclear facilities, solar farms, or whatever source it takes to satiate their increasingly voracious appetites. And AI data centers are the hungriest of all.
Training generative-AI models requires special chips called graphics processing units or tensor processing units, which are better at multitasking and pack significantly more computing power into a smaller footprint. But the data centers that rely on those chips consume massive amounts of energy, even by internet standards. According to the Department of Energy, traditional data centers already use up to 50 times as much energy as a typical office building, accounting for 2% of America’s electricity use. Server racks dedicated to training generative AI, by contrast, may require as much as seven times the energy as traditional server racks do. When it comes to power use, AI is the internet on hyperdrive.
Like any good magic trick, the internet is based on an illusion. Our stuff hasn’t vanished — it’s just moved to data centers.
And then there’s the demand from us users. Any time we ask a chatbot to plan our vacation, solve our calculus homework, or make us a meme, we’re asking a data center to tap more power than we’d use doing a web search for the same stuff. Research has found making even a single image with generative-AI tools can use as much energy as fully charging your cellphone.
All this demand is creating energy bottlenecks. “That power capacity isn’t just sitting out there available,” said Jon Lin, an executive vice president at Equinix, which operates data centers in 32 countries. Indeed, in a recent Cisco survey, 97% of business leaders say they’re feeling pressure to implement an AI strategy — but more than half don’t feel they’re fully prepared to meet the necessary power demands.
Faced with this pressure, local utility companies are making tough environmental trade-offs. Virginia’s Dominion Energy recently predicted electricity demands in its service area would nearly double in the next 15 years, an unprecedented growth rate that the utility attributed, at least in part, to data centers. Amazon alone plans to invest $35 billion in data centers in Virginia, adding to the $52 billion it’s already spent there. (Full disclosure: I’m married to an Amazon employee.) To keep up with power demands, Dominion says that while 95% of its new power plants will be carbon-free, it will also need to build new gas-powered plants and keep some fossil fuel plants that were set to be retired operating for far longer than expected. The utility now predicts that in 25 years, its carbon emissions will be 65% higher than it was expecting just two years ago, but Dominion says those estimates are subject to change.
Such constraints are likely to push data centers into new pastures — literally. Google plans to spend $1.2 billion to build its fleet of centers in Nebraska to capitalize on the state’s wind power. Microsoft similarly opened data centers in Sweden because of the country’s ability to provide plenty of clean energy. “We are spreading out,” said Noelle Walsh, the corporate vice president of cloud and innovation for Microsoft. “We’re going to where the customers are but also going to where the resources are as well.” Of course, the industry’s growth in new territories also means more communities are about to be exposed to data-center sprawl, and its consequences.
To show me what that conquest looks like, Schlossberg drove me 25 miles north of Pageland Lane to Data Center Alley in Loudoun County, Virginia. The area bursts with the largest concentration of data centers on earth, underpinning vast amounts of the world’s internet traffic. As we tooled around in her Ford Fusion, every corner revealed another imposing gray cube with a corporate logo emblazoned on the side, stretching across tens or hundreds of thousands of square feet. These beasts can be noisy neighbors, too. “DataCenter is generating noise like a freight train all night long,” says one of more than 40 complaints to Loudoun County that Business Insider obtained through public records. “Constant humming noise like a lawn mower outside 24/7,” says another.
To be sure, many residents in Prince William County favor the arrival of data centers. Proponents of the Digital Gateway development argue that it would generate an estimated $400 million in tax revenues a year for the county. In Loudoun County, data centers brought in $663 million in taxes last year, accounting for one-third of the county’s annual budget. All told, according to industry estimates, the construction and operation of data centers in the U.S. generated nearly $100 billion in federal, state, and local tax revenue in 2021 — helping to fund schools, parks, and other essential infrastructure.
“That’s a pretty good return,” says Terry Clower, a professor of public policy at George Mason University. “If you didn’t have the data centers, then you’d probably have a combination of both lower services being offered and higher tax rates.”
Besides, it’s not as if the area is untouched by data centers. It’s already home to dozens of them. In fact, it was local landowners who conceived of the Digital Gateway to begin with, after the industry’s growth throughout Northern Virginia began to creep into their backyards. “We decided we needed to create our destiny,” said Mary Ann Ghadban, a local resident and real-estate broker who handled a portion of the project.
On Ghadban’s 55 acres on Pageland Lane now stand two 115-foot-tall transmission towers, connected to 250-foot-long transmission lines, which were erected during an upgrade by Dominion Energy, in part to accommodate all the machines in Data Center Alley. For Ghadban, the data center takeover of Prince William County is already here. But with the Digital Gateway, she says the county could model a better way for this fast-expanding industry to build — one that takes up less land and is farther removed from things like housing and schools. According to QTS, Prince William County has set stricter standards for the Digital Gateway than any other jurisdiction where QTS has operated. “I think it’s going to set the precedent for, obviously, this development as well as many others around the country,” a QTS executive told Business Insider.
But while companies trumpet the benefits of their data centers, they sometimes go to great lengths to keep the downsides a secret. To keep their machines from overheating, for example, data centers often consume and evaporate millions of gallons of water a day. What makes the problem tough to track — and, therefore, invisible to local residents — is companies rarely disclose just how much water any single data center is consuming. In The Dalles, a city in Oregon, Google glamorizes the plumes of water vapor that rise above the cooling towers for its data centers and “create a quiet mist at dusk.” But when a local newspaper sought public records on how much water the data centers consume in the process of creating that mist, the city fought the request in court — with Google paying its legal bills. Last year, when the case finally settled, residents learned that Google is responsible for more than a quarter of the city’s water use. Worldwide, according to site-specific data Google finally disclosed, its data centers consumed 4.3 billion gallons of water in 2021. (Google didn’t respond to requests for comment.)
Then there’s the question of jobs. The industry says a single data center can create thousands of construction jobs over the course of the buildout, which in turn fuels more economic activity at local businesses and drives job growth. But opponents of data centers point out that for all the land they consume, they directly employ only a few dozen to a few hundred people after construction is complete. Take Meta, which is currently building its biggest data center complex ever, spanning more than 5 million square feet in Altoona, Iowa. When it’s finished, the campus will employ just over 400 employees. By contrast, the Mall of America, which is just slightly bigger than Meta’s data center complex, employs some 11,000 people. Even the average Amazon fulfillment center, which encompasses only 800,000 square feet, employs around 1,500 people — 23 times more jobs per square foot than Meta’s center in Altoona. If you’re looking for employment, you probably won’t find it in a data center.
Critics fear that turning your town into a corridor for data centers can also discourage the sort of business development that would directly employ a lot of people. When Mesa, Arizona invested in its power grid and broadband fiber, it hoped to attract advanced manufacturing facilities to the area. Instead, says Scott Somers, a city council member, the town has been “inundated” with what he calls “warehouses of ones and zeroes,” leaving few places for jobs-rich businesses to expand. “When data centers use up all the power or the water or the land in the community,” Somers says, “then where are we going to put other developments?”
To account for the swelling energy consumption demanded by their data centers, tech giants have also become the largest investors in new sources of renewable energy. Last year, according to BloombergNEF, Amazon, Meta, Google, and Microsoft were the top four global purchasers of clean energy. Microsoft has even signed an ambitious deal to purchase energy from a nuclear fusion facility by 2028.
Big Tech’s investments in clean energy aren’t enough to keep up with the power demands of their data centers.
But the big new investments in renewable energy aren’t enough to keep up with power demands. Data centers operate at full speed every day, around the clock, even when the sun isn’t shining and the wind isn’t blowing. That means they often have to rely on other, dirtier sources of energy. Meta, for example, says it offsets 100% of its annual energy usage with renewable energy. But Ben Lee, a professor at the University of Pennsylvania who specializes in computer architecture and energy efficiency, found that Meta’s energy consumption on an hourly basis in 2021 was carbon-free only half the time. Google has also reported that, hour-to-hour, it is able to use carbon-free energy only 64% of the time. While it’s undoubtedly a good thing that these companies are offsetting their emissions by adding new renewable energy to the grid, Lee says, “Your actual usage is only as green or brown as your local transmission lines.”
On December 12, hundreds of residents of Prince William County cycled in and out of the drab government hearing room where the local Board of Supervisors was meeting to decide the fate of the Digital Gateway project. One by one, over 27 tedious hours, they approached the hot pink microphone to plead with the weary board members.
Supporters accused opponents of being elitist, NIMBY obstructionists, hell bent on ruining a golden opportunity for the entire county. Opponents framed supporters as mercenary sellouts attempting to pad their own retirement funds at the expense of their neighbors. There were costumes and props and one particularly cringe adaptation of Joni Mitchell’s “Big Yellow Taxi.” (“They paved paradise, and put up a data center. Oooh. Stop. Stop. Stop.”)
By the time Schlossberg spoke, it was after 1:30 a.m. After handing out kitschy Hanukkah gifts as a peace offering to the board — an owl for wisdom, gelt to remind them “there are more important things than money” — she urged them to consider the Digital Gateway’s impact on their own legacy. “This is going to be your mistake, that you’re going to have to apologize for,” she warned them. “There is an off ramp. Take it.”
But the board didn’t take it. Instead, in a straight party-line vote, they approved the Digital Gateway. The Democratic majority voted in favor, the Republican minority against. One board member abstained to break the tie.
To be fair, there’s only so much that local opposition can do to stem the global tide of data centers. Unlike crypto mining, another energy-sucking industry that has inspired widespread resistance, data centers are essential to the way the modern world works. They power our hospitals and governments as much as our TikToks and Midjourney masterpieces. In 2019, Singapore imposed a three-year moratorium on new data centers, and the Dutch government recently instituted its own nine-month moratorium for hyperscalers. But countries that forbid data centers from opening are risking their own role in the future economy. “Growth eventually will leap frog you if you don’t allow it and manage it,” says Clower, the public policy expert.
If AI is, in fact, poised to make data centers more plentiful and powerful, what matters now is not wishing them away, but finding new, less destructive ways for them to grow. That starts with requiring the companies that run them to share more information about the resources their AI models are consuming. “The minimum we should ask for is transparency,” says Sasha Luccioni, a research scientist and climate lead at Hugging Face, a platform for AI developers.
An equally important part of the solution is asking ourselves a more fundamental question: Do we really need to use such immensely powerful AI models for every single daily chore? The corporate rush to pack these models into just about every product is the technological equivalent of hiring a team of bodybuilders to pick up your kids’ Legos. “You don’t need to ask your fridge for a recipe for chocolate chip cookies,” Luccioni says. “You can literally do a Google search. Switching out a model that will find you information on the internet to one that will generate your information is just using more energy for the same task.”
In the end, though, opponents of new technologies have seldom had much success at convincing consumers to forgo the Latest Cool Thing in the name of broader social concerns. When we’ve spent so long detached from the physical reality of the digital world, it makes it a lot easier to just ignore the environmental and human costs. Misbelieving that the internet simply exists somewhere out there in the cloud has rendered us unprepared to grapple with what happens when it all comes crashing down to earth.
Issie Lapowsky is a freelance journalist covering tech, politics, and society. Her work has been featured in The New York Times, Wired, and Fast Company.