The Hidden Environmental Impact of AI

While the mass adoption of AI has transformed digital life seemingly overnight, regulators have fallen asleep on the job in curtailing AI data centers’ drain on energy and water resources.

A power substation near the LC1 CloudHQ data center in Ashburn, Virginia, US, March 27, 2024. (Nathan Howard / Bloomberg via Getty Images)

In early May, Google announced it would be adding artificial intelligence (AI) to its search engine. When the new feature rolled out, AI Overviews began offering summaries to the top of queries, whether you wanted them or not — and they came at an invisible cost.

Each time you search for something like “how many rocks should I eat” and Google’s AI “snapshot” tells you “at least one small rock per day,” you’re consuming approximately three watt-hours of electricity, according to Alex de Vries, the founder of Digiconomist, a research company exploring the unintended consequences of digital trends. That’s ten times the power consumption of a traditional Google search, and roughly equivalent to the amount of power used when talking for an hour on a home phone. (Remember those?)

Collectively, De Vries calculates that adding AI-generated answers to all Google searches could easily consume as much electricity as the country of Ireland.

Unlike their chatbots, the companies behind these advances are far less willing to share information. Though researchers like De Vries can make educated estimates, because of a lack of industry transparency it remains surprisingly difficult to put an exact number on just how much power and water AI might use. Yet that demand is soaring as the technology is tacked on to everything from your iPhone’s operating system to how your car insurance company calculates your rates.

While the mass adoption of AI has transformed digital life seemingly overnight, regulation of its very physical impacts has not kept pace. Federal agencies like the US Energy Information Administration, which collects information about industries’ energy use, aren’t tracking the demand of the data centers that enable AI — even as their footprint skyrockets.

“We do not have any mandated disclosures on the amount of energy or resources that general AI systems use,” says Merve Hickok, president and research director of the Center for AI and Digital Policy, a nonprofit research organization. When journalists file record requests to get this information, it’s usually redacted. This secrecy limits the ability of utilities and regulators to know how these needs are changing.

That’s a problem, because data centers are rapidly outgrowing the electric grid while keeping dirty sources of power, like coal plants, operating. Tech companies also have a long track record of arranging for special, discounted rates for their massive power consumption — which means in many cases, ratepayers like you are subsidizing data centers’ undisclosed energy use.

In addition to power, these facilities suck up substantial amounts of water to cool their servers, and are often located in places where land is cheap — like deserts. Only a few operators report their water usage, even though a fifth of servers draw water “from moderately to highly stressed watersheds.” One paper estimates that globally, the demand for water for data centers could be half that of the United Kingdom within the next several years.

Yet even as questions about data centers’ impact on the public grow, companies are expanding restrictions on what they share about their operations. In a written response to us, a Google spokesperson says that since introducing generative AI to its search, associated machine costs have decreased by 80 percent. They say that based on internal data, De Vries’s analysis is “an overestimation and our systems are far more efficient.”

But they declined to provide any further specifics about their energy use, other than noting that predicting the future growth of energy consumption and emissions from AI data centers is challenging.

“We’re actually seeing less and less disclosure,” De Vries says, as companies claim information about models harms their competitive advantage. “In terms of transparency, we’re actually going backward.”

Everyone’s Paying AI’s Secret Costs

AI answers a question about itself. (ChatGPT)

Last December, Stephen Ward stepped up to a podium at a public hearing in Prince William County, Virginia, as the Board of Supervisors considered approving one of the world’s largest data center projects. After getting in line before dawn for a numbered spot, he spent five hours waiting in the lobby before finally getting a chance to speak.

In the 1970s, Ward was an economist at the Environmental Protection Agency, where he advised policymakers on hazardous waste regulations, before becoming chief investment officer of Charles Schwab Investment Management in the 1990s. “I’m used to making multibillion investment decisions,” he told the board. “Scale should not grant an exemption for mandatory details; it should increase concern.”

The PWC Digital Gateway development proposed bringing as many as thirty-seven new data centers to the rural area, building on more than two thousand acres of land. These massive warehouses hold tens of thousands of servers, which are usually stacked in towering vertical racks. When you chat with an AI-generated bot about a mislaid package, your query is sent to one of these servers. Their high-performance hardware runs your question through a computer program that makes decisions similar to the human brain. This so-called neural network is trained on a large language model before sending the generated answer back to your device.

This process often requires consistent, round-the-clock power, making it more difficult for utilities to manage the load. It also means that these facilities need high-capacity electrical equipment like transformers, circuit breakers, and often new substations to connect to transmission lines. All this power generates a lot of heat, so interspersed throughout the center’s corridors are a variety of complex heat exchangers and cooling systems.

Northern Virginia already has the world’s largest market of data centers. The state is home to companies like Amazon Web Services, Google Cloud, and Microsoft Azure. It’s the latest chapter in the region’s long history with the technology industry; the internet itself was born in Arlington in 1969, when a military project originally intended to connect universities was switched on.

A map of data centers in Virginia. (The Virginia Economic Development Partnership)

In the following decades, Virginia laid a profusion of fiber-optic cables, providing attractive high-speed connectivity. In the 1990s, this attracted early dot-com companies like AOL — whose purring dial-up noise was once ubiquitous with being online. Last year, the tech company’s abandoned campus headquarters was torn down to make way for more data centers.

But as more and more server farms spring up, the state’s largest electric utility, Dominion Energy, has scrambled to keep pace. The industry’s peak energy usage in 2022 was almost 2.8 gigawatts, or about a fifth of the utility’s total statewide sales. That same year, Dominion told its customers in Loudoun County it could no longer guarantee it would deliver as much power as they needed, stalling the breakneck development.

Instead, data companies began to eye places like Iowa, Georgia, and nearby Prince William County — where residents like Ward warn that similar problems are on the horizon. Critics say the Digital Gateway proposal alone will require at least three gigawatts of electricity, or the equivalent of the power demand of 750,000 homes. “Where is that power coming from?” Ward asked the Board of Supervisors.

It’s a question that PJM Interconnection, the regional transmission organization that spans thirteen states and the District of Columbia, is also asking. The organization, which coordinates the movement of wholesale electricity in parts of the eastern United States, recently approved a set of $5.1 billion transmission projects, primarily to deliver more power to Virginia’s data centers.

The problem is that these costs will be distributed across the various states within the network, says David Lapp, Maryland People’s Counsel, an independent Maryland state position that advocates for Maryland’s residential utility consumers. Though the upgrades primarily benefit private companies in Virginia, they will result in rate hikes for ordinary Maryland customers, a move Lapp calls “fundamentally unfair.”

The added transmission capacity is more than what Maryland’s largest utility itself currently uses at peak times. “The scale, scope, and cost of the [Digital Gateway] projects are unprecedented,” Lapp wrote to the PJM Board of Managers before arguing to federal energy regulators that PJM had unfairly allocated costs to Maryland. The Federal Energy Regulatory Commission denied his request in May, leaving Maryland on the hook for $551 million.

This is a common scenario. In Indiana, for instance, utility regulators recently approved a new $800 million data center campus with Meta Platforms, Inc., which owns Facebook and other social media services. The secret negotiated rate for the facility’s power has been redacted from public filings, but what was included was that the infrastructure required to connect the facility to the grid will cost $82 million.

Duke Energy Indiana redacted details about the special electricity rates it provided to Meta Platforms for its new data center campus.

The Office of the Utility Consumer Counselor allowed the facility to shift those costs onto ratepayers, arguing it would bring capital investment to the area. Data centers, however, don’t create many local jobs. Nevertheless, the Indiana facility will be doubly subsidized, as it is also receiving a thirty-five-year sales tax exemption from the state.

Just as mortgage companies make money on interest, incentivizing them to sell more mortgages, utilities make money by spending on infrastructure. That’s because regulations allow these companies a return on their investments in upgrades like new transmission lines. Utilities, in other words, also profit from outsourcing data center costs to the public. In fact, Dominion’s most recent investor presentation proudly claimed “robust rate base growth,” and forecast a staggering 8,500 megawatt spike in demand.

“It’s very counterintuitive,” Lapp says, that utilities “make money by spending other people’s money.”

The arrangement sends the wrong price signal to the industry, Lapp argues. If tech companies paid full freight for their energy infrastructure, they would be incentivized to find ways to use less power. Instead, many are going in the wrong direction: despite its goals to become carbon-free by 2030, Microsoft’s emissions jumped by 30 percent in 2023, thanks to its recent investments in AI.

Altogether, a new report by the Electric Power Research Institute found that AI could comprise roughly 9 percent of the country’s total energy demand by the end of the decade. Other estimates suggest global data center energy demand could double by 2026, while some utilities, like those in Arizona and Washington, may see as much as 10 percent load growth.

This insatiable hunger for power is slowing the transition to green energy. When the owner of two coal-fired power plants in Maryland filed plans to close last year, PJM asked them to keep running till at least 2028 to ensure grid reliability. Meanwhile, AI is also being used to actively increase fossil fuel production. Shell, for example, has aggressively deployed AI to find and produce deep-sea oil.

“The truth is that these AI models are contributing in a significant way to climate change, in both direct and indirect ways,” says Tom McBrien, counsel for the Electronic Privacy Information Center, a digital policy watchdog.

Even before Google’s AI integration this spring, the average internet user’s digital activity generated 229 kilograms of carbon dioxide a year. That means the world’s current internet use already accounts for about 40 percent of the per capita carbon budget needed to keep global warming under 1.5 degrees Celsius.

But in the absence of government tracking and regulation, the industry continues to surge unchecked. Back in Prince William County, the public hearing stretched on as more than 350 people testified. Outside, the stars twinkled and then faded into a muted dawn. Finally, twenty-seven hours after the meeting began, the board held a final vote, deciding to allow the Digital Gateway project to move forward.

“The grid is now in trouble,” Ward said. “Only when it’s too late will we begin to measure and regret our lack of foresight.”

When Regulators Can’t Keep Up

AI answers a question about itself. (ChatGPT)

There’s a lot of urgency to address AI’s rocketing growth, but government moves notoriously slowly. In September, Senate Majority Leader Chuck Schumer, a New York Democrat, held the first event in a series of closed-door meetings with AI leaders to discuss the industry’s future.

“We have no choice but to acknowledge that AI’s changes are coming,” he told tech titans like Elon Musk, Bill Gates, and Mark Zuckerberg. “What role does Congress and the federal government have in this new revolution?”

Advocates like Grant Fergusson, an equal justice works fellow at the Electronic Privacy Information Center, say the meetings were a clear example of how much power industry voices have in crafting AI policies.

“The whole process didn’t discuss the environmental impacts in any meaningful way,” says Fergusson, adding that the Senate discussions only belatedly included significant civil rights issues after pressure from advocacy groups. “This bespoke, industry-driven process unsurprisingly led the Senate to an industry-friendly destination,” summarized a coalition of fifty-two advocacy and research groups.

That’s not a coincidence: A report by nonprofit watchdog Public Citizen found that the number of lobbyists on AI issues mushroomed in 2023, increasing by 120 percent. The vast majority worked for corporate interests, including sixty lobbyists employed by Microsoft, while Amazon hired an additional thirty-five. Last year, Amazon, Google’s parent company Alphabet, Meta, and Microsoft each spent more than $10 million on lobbying for various interests.

(Public Citizen)

“As federal agencies move forward with developing guardrails for AI technologies, stakeholders will likely rely even more on their lobbyists to shape how AI policy is formed,” Public Citizen’s authors wrote.

So far, much of the discourse around artificial intelligence’s risks has centered around hyperbolic scenarios straight out of science fiction, like chatbots developing sentience or artificial general intelligence. The industry likes to center those concerns, McBrien says, because “dramatic future harms distract from their actual current business practices.”

Last fall, President Joe Biden issued an executive order instructing the National Institute of Standards and Technology, which develops measurements and guidelines for various fields, to develop additional AI standards. To inform their efforts, the agency created the Artificial Intelligence Safety Institute Consortium, which recently held private meetings to discuss the technologies’ risks.

At one of its recent events, according to an attendee, who did not want to be named because they were not speaking with their organization’s permission, industry representatives objected to including energy use in the agency’s draft framework. The framework currently calls for measuring environmental impacts and addressing greenwashing concerns.

Despite deliberations behind closed doors about AI’s ecological footprint, tech companies’ lobbying has kept these concerns from being addressed in state or federal legislation. When asked for an interview about the industry’s sustainability, a National Institute of Standards and Technology representative said, “Our AI experts have been quite overtaxed,” and then that they didn’t have “anyone with that expertise.”

Fergusson thinks the opacity of these conversations is part of the problem. “The fact that environmental impacts are being debated so intensely within these [AI] standards, and the fact that they are almost never brought up within any of the legislative solutions we’ve seen at the state or federal level, is in part due to some pretty heavy lobbying by tech companies.”

The Energy Information Administration, which analyzes energy-related information and would likely be the agency that carries out any AI monitoring regulators ultimately recommend, also does not currently calculate data centers’ energy use. Its press officer explained that the administration’s last survey on commercial buildings’ energy use, which was conducted in 2018, couldn’t separate out data centers, in part because of “low cooperation rates.”

This persistent gap hinders the public’s understanding and regulators’ ability to mitigate the problem, says Senator Edward Markey, a Massachusetts Democrat. He introduced legislation in February to at least create voluntary reporting guidelines on how AI is affecting the environment. In an email to us, he wrote, “We must do everything we can to protect communities and our planet from the threats of climate change and enable a livable future for everyone.”

But De Vries, the researcher calculating Google’s AI search power consumption, says that relying on companies’ voluntary disclosures for this data will never be sufficient.

“Big tech companies are obviously not going to provide that information voluntarily,” he says. “We would really need a push from regulators.”

A Perilous Equation

AI answers a question about itself. (ChatGPT)

In the meantime, the backlash to data centers is growing, and not only from progressive advocates. Historically prodevelopment states like South Carolina are now considering legislation that would prevent data centers from receiving sweetheart utility deals.

Companies like Amazon have for years used their influence to demand special treatment from utilities, like shunting the $170 million cost of burying power lines associated with a Virginia data center to ratepayers.

The company also recently negotiated an arrangement with Ohio’s public utilities commission for a discounted rate on the enormous amount of power their data centers will use for the next ten years. The exact terms are secret, because of a 2017 agreement to hide a former Amazon discount from the public, but it’s estimated to shift $135 million a year onto the utility’s customers.

Overall, Goldman Sachs analysts found that US utilities will need to invest $50 billion to support new power generation for data centers — potentially resulting in big rate hikes for consumers.

(Goldman Sachs)

This spring, the US Department of Energy tried to force at least crypto mining companies to report their energy use, but the effort was kneecapped by a Texas judge, who issued a temporary restraining order to stop it.

The Lone Star State’s stance may be changing, however: In June, the CEO of the Electric Reliability Council of Texas, the organization that runs the state’s grid, announced that the state’s power infrastructure would need to double in the next decade to keep up with data centers and crypto mining. In response, the state’s lieutenant governor signaled he would be taking a closer look at these “niche industries that have massive power demands and produce few jobs.”

Even industry insiders acknowledge that something needs to change.

“AI companies have strong financial incentives to avoid effective oversight,” a group of former OpenAI employees wrote in a recent open letter, “and we do not believe bespoke structures of corporate governance are sufficient to change this.” They also say these companies can’t be relied on to share information voluntarily, and that they require independent regulation.

Of course, like any technology, AI can be used in both good and bad ways. Microsoft, for example, claims that AI can help discover and develop climate solutions, like more efficient renewable energy production and developing new materials for carbon capture. Microsoft is also working with Google to aggregate their demand for renewable energy, while Google is committed to net-zero emissions across its operations by 2030.

Yet the sheer volume of energy needed for the tech industry is simultaneously delaying similar state commitments to transition to green energy. In Michigan, for instance, which passed groundbreaking climate legislation last year, experts warn that data centers will  prevent the state from achieving its goal of carbon-free energy by 2040.

Some experts, like Rob Gramlich, the principal of Grid Strategies, a power-sector consulting firm, say that it’s not a utility’s role to determine if energy is being used responsibly.

“Most state laws and electricity policies are written around the obligation to serve,” he says, “so the utility serves all customers for all uses.”

This is, in fact, one of the core arguments behind instituting something like a carbon tax, which would put a price on all carbon emissions — providing a process to at least price the true consequences of how we use energy.

As states begin to rethink their appetite for data centers, AI’s negative consequences are increasingly being pushed to countries in the Global South, where companies can exploit lower electricity and water prices.

“It’s difficult to reckon with the physical harms of AI, says Brian Chen, policy director at the nonprofit Data & Society, because “tech companies invisibilize the consequences of these systems, most people don’t have to think about it.”

Latin America, for example, is now seeing a surge in data center development, including near drought-stricken Mexico City, which is hurtling toward a day in the near future when its taps run dry.

“We also need to question who disproportionately suffers,” said Boxi Wu, a graduate research student at the Oxford Internet Institute. They say that scrutiny needs to include an analysis of AI’s entire supply chain, including the rare earth minerals its infrastructure requires and the electronic waste being generated by rapidly advancing chip technology.

Wu recently published a paper highlighting how global economic and political power balances in AI production are linked to past colonial dynamics — such as how exploitative and pollution-intensive mineral mining tends to occur in places like China, Africa, and Latin America, while the end products are enjoyed in the United States and Europe.

No one wants to think of their internet habits as stealing someone else’s drinking water. Looking at your monthly power bill, the link between Apple’s latest AI announcement and your rate increases isn’t necessarily obvious. And technology companies are currently spending a lot of money to make it harder to say that’s happening. But it’s difficult to escape the sinking sense that the benefits of AI are being accrued by a small number of powerful companies, while the physical harms are borne by people out of sight.

“I wouldn’t say it is unfair to look at a technology and ask these questions,” says De Vries. “We’re in a time where we have to make choices about how we use our resources.”

Back in Virginia, Ward sees land being cleared as he drives through his neighborhood, trees cut down to stumps. He recalls visiting the nearby Civil War–era Manassas National Battlefield as a kid, picturing what it might have been like for the people who had once fought across the fields. Soon, enormous, four-story buildings will loom over the historic park.

But it’s the consequences for the area’s water that concerns him most. The Digital Gateway project, Ward says, has been given “a blank check.” It won’t have to report how much groundwater it’s using, and there’s no limit to how much it may consume. He says the vast expansion of paved surfaces will cause widespread runoff and prevent recharging the aquifers on which the county depends.

People don’t realize that when the groundwater is exhausted, there’s no alternative.

“They don’t live in the natural world,” Ward says. “They live in a world where there’s a pipe in the wall, and you get water, and they have no idea where it comes from.”

He adds, “But when it’s gone, it’s gone.”