A.I. Is on the Rise, and So Is the Environmental Impact of the Data Centers That Drive It
The demand for data centers is growing faster than our ability to mitigate their skyrocketing economic and environmental costs
A.I. Is on the Rise, and So Is the Environmental Impact of the Data Centers That Drive It
The demand for data centers is growing faster than our ability to mitigate their skyrocketing economic and environmental costs
Amber X. Chen
- AAAS Mass Media Fellow
September 29, 2025 8:00 a.m.
Amazon data centers sit next to houses in Loudoun County.
Jahi Chikwendiu / The Washington Post via Getty Images
Key takeaways: A.I. and data centers
As the demand for A.I. increases, companies are building more data centers to handle a growing workload.
Many of these data centers are more than 30,000 square feet in size and use a lot of power and water.
Gregory Pirio says he never would have moved to his townhome in Northern Virginia’s Loudoun County had he known that the area would soon be at the epicenter of a data center boom.
Pirio—who works as the director of the Extractive Industry and Human Development Center at the Institute of World Affairs—moved to the county, just about an hour’s drive outside of Washington, D.C. 14 years ago. Back then, he recalls the place being filled with forested areas and farmland, with the occasional sounds of planes flying in from Dulles.
“It was just really beautiful, and now it has this very industrial feel across it,” he says, adding that one can now drive for miles and just see data centers.
Data centers are buildings that house the infrastructure needed to run computers, including servers, network equipment and data storage drives. Though they’ve been around since 1945 with the invention of the first general-purpose digital computer, in the past few years there has been an explosion in data center development to match the rapid rise of artificial intelligence.
Over the past year, the environmental consequences of A.I.—specifically its most popular generative platforms like ChatGPT—have been under intense scrutiny. Last July, NPR reported that each ChatGPT search uses ten times more electricity than a Google search. In March 2024, Forbes reported that the water consumption associated with a single conversation with ChatGPT was comparable to that of a standard plastic water bottle.
The emissions of data centers are only projected to go up, especially as companies look to employ A.I. on users’ behalf. For example, in May, Google announced A.I. overviews, a new user enhancement strategy that uses A.I. to create succinct summaries based on websites associated with a Google search query. Those queries and others like it on different platforms increase the need for additional data centers, which will require more and more energy.
What are data centers?
Data centers come in a variety of sizes. According to a 2024 report by researchers at Lawrence Berkeley National Laboratory, they can range from smaller centers—integrated into larger buildings for internal use by companies—that are on average less than 150 square feet, to hyperscale centers which are operated off-site by large tech companies to facilitate large-scale internet services.
On average, hyperscale data centers are 30,000 square feet, although the largest of these data centers can reach sizes of well over one million square feet. As of 2024, more than half of the world’s hyperscale data centers were owned by tech giants Amazon, Microsoft and Google. Large data centers, particularly hyperscalers, are the data center of choice for companies looking to operate A.I. platforms, due to their high computing power.
Clusters of large data centers are strategically chosen based on proximity to clients, electricity costs and available infrastructure. For example, data centers have been running through Northern Virginia since the advent of the internet in the mid-1990s because of the area’s cheap energy, a favorable regulatory system and proximity to Washington. Northern Virginia holds the highest concentration of data centers in the world at over 250 facilities. Across the state, data centers are now near schools, residential neighborhoods and retirement communities.
According to Ann Bennett, data center issues chair at the Sierra Club’s Virginia Chapter, new data centers that have been popping up across the area are of an entirely different scale and era. “These are bigger, taller,” Bennett says. “They’re pretty much only building hyperscalers.”
How do data centers consume energy?
To power the digital world—from day-to-day digital communications, websites and data storage—data centers require energy to power the hundreds of servers within them. With the advent of more hyperscale data centers being built to support A.I., data center energy use has increased.
Benjamin Lee, a computer scientist at the University of Pennsylvania, breaks the high energy consumption of A.I. into two categories.
First, there is the training that A.I. models undergo, in which tens of thousands of graphics processing units, or GPUs, within a data center must consume large datasets to train the parameters of more powerful A.I. models. Second, once an A.I. model is trained, it performs inference—or the process of responding to user requests based on its training.
According to Lee, every word that a user provides to an A.I. model is processed to figure out not only what the word means but the extent to which that word relates to all other words that have been fed into the model. Thus, as more words increase processing time, more energy is consumed.
“Fundamentally, A.I. uses energy, and it doesn’t care where that energy is coming from,” Lee says.
Data centers mostly get their energy from whatever local grid is available to them. Globally, because most electric grids still rely heavily on fossil fuels, A.I. increases greenhouse gas emissions, says Shaolei Ren, a computer engineer at the University of California, Riverside. Virginia, for example, is part of PJM grid, for which the primary fuel source is natural gas.
According to Noman Bashir, a computer engineer at MIT, because data centers are huge power consumers they often disrupt electric grid infrastructure, which can decrease the lifespan of household appliances, for example.
In addition, Bashir notes that grid infrastructure must be updated when each new data center comes in—a cost that residents are subsidizing. In a 2025 report, the Dominion Energy found that that residential electric bills are projected to more than double by 2039, primarily due to data center growth.
Already, the technology industry has seen a growth in emissions, mostly fueled by data centers. In July, Amazon reported that its emissions rose from 64.38 million metric tons in 2023 to 68.25 million metric tons in 2024—the company’s first emissions increase since 2021, primarily due to data centers and the delivery fleet it uses. Google, too, reported that its 2023 greenhouse gas emissions marked a 48 percent increase since 2019, mostly due to data center development and the production of goods and services for company operations.
How else does A.I. impact the environment?
Another dimension of A.I.’s environmental footprint is its water consumption. To put it simply, Ren explains that these powerful computers that run A.I. also get extremely hot. So, to keep them from overheating, data centers cool them with power air conditioning systems that are run by water.
Water that is heated by computers is moved to massive cooling towers on top of a data center, and then is circulated back in. A data center’s direct water consumption is attributed to the water that evaporates during this process. This water loss is then left to the whims of the water cycle.
“You don’t know how long [the water] will take to return or whether it will return to a specific geographic location,” Lee explains. “So where water is scarce, it’s a concern.” In 2023, data centers in the U.S. directly consumed about 66 billion liters of water.
Bashir adds that the industry’s environmental impacts can also be seen farther up the supply chain. The GPUs that power A.I. data centers are made with rare earth elements, the extraction of which Bashir notes is resource intensive and can cause environmental degradation.
How will data centers affect power consumption in the future?
In order to meet A.I.’s hunger for power, companies are looking to expand fossil fuel energy projects: In July, developers of the Mountain Valley Pipeline—a natural gas system that spans about 303 miles across Virginia—announced that they were considering a plan to boost the pipeline’s natural gas capacity by 25 percent. Earlier this year, the Atlanta-based electric utility Southern Company announced that it would backtrack on its previous announcement to retire a majority of its coal-fired power plants, citing growing demand from data centers.
And when the grid can’t satisfy their needs, Lee says that data centers are now increasingly developing their own power sources—whether from renewable energy sources like nuclear or fossil fuel-based power plants.
Pirio lives about 150 yards away from a data center that is not connected to the local grid. Instead, it’s powered by natural gas turbines with back-up diesel generators.
He says that the noise pollution associated with the data center’s gas turbines is a huge problem for him and his neighbors, describing the din as a constant, humming sound. “Many of the neighbors, we got decimal reader apps, and it was off the charts. … They were like 90 decibels near our house,” he says.
Pirio explains that he can no longer open the windows of his house on cool evenings because of the noise. He says another neighbor put mattresses against their window to block the noise.
Pirio says he and his neighbors have no way of assessing what the emissions coming from the gas turbines are. “There’s just not structure for us to know, and they’re pretty much invisible,” he says. The Environmental Protection Energy notes that the presence of a fossil fuel-based power plant can significantly degrade air quality and emit toxic heavy metals like mercury into the atmosphere, harming local populations’ health.
Vantage Data Centers, the company which runs the data center near Pirio, says it has installed Selective Catalytic Reductions (SCRs) which, according to its website, can reduce nitrogen oxide emissions from diesel generators by up to 90 percent.
Resident health and quality of life are not the only factors associated with data centers developing their own power sources. Even when data centers produce their own energy, Lee says the grid still provides them with significant backup infrastructure—which as Bashir explains, can still overwhelm the grid, causing it to become more unreliable for residents.
How can A.I.’s data centers be made more sustainable?
According to Lee, the renewable energy sector is simply not growing fast enough to meet the needs of A.I. While some analyses position data centers to grow at a rate of as much as 33 percent a year, the World Economic Forum says that global renewable energy capacity grew by 15.1 percent in 2024.
Bashir and Lee both emphasize that much of the data center growth we are seeing is not being built on actual need, but speculation. According to Bashir, because tech companies are building data centers at such a rapid pace, these new centers will inevitably be powered by gas generators or other forms of fossil fuel, simply because infrastructure for widespread renewable energy does not yet exist.
Beyond improving investments into renewable energy, Lee says that working toward algorithmic optimization is another way for A.I.’s data centers to lessen their carbon footprint. In a 2022 article, Lee—in collaboration with researchers at Meta—identified ways in which optimizing A.I. models can also improve sustainability.
For example, researchers identified “data scaling”—in which a model is fed more data sets, resulting in a larger carbon footprint—as the current standard method to improve model accuracy. With a more efficient algorithm, energy costs could be significantly reduced.
Lee emphasizes that those working toward creating more efficient A.I. must also focus on achieving a lower carbon footprint.
Bashir adds that education remains an important tool to cutting back on A.I.’s emissions.
“People can be educated on what are the A.I. tools available at their disposal,” he says. “How can they optimize their use? And [we need to tell] them of all the negative impacts of their use, so that they can decide if a particular use is worth this impact.”
Get the latest Science stories in your inbox.
