The rush to fill data centers with graphics processors for generative AI is turning a once gamer-coded chip into a local infrastructure problem. Researchers and community groups say the costs show up as higher electricity demand, water stress, air pollution, mining waste and e-waste, often near people who get little say in the buildout.
GPUs sit inside phones, cars, consoles and gaming PCs, but AI has made them the object of a different fight. Nvidia, now the world’s most valuable company, says it introduced “the world’s first GPU” in 1999, although graphics hardware used in arcade games dates back decades earlier. Catherine Flick, a professor of ethics and games technology at the University of Staffordshire, said games have long served as a testbed for hardware and software that later raise broader ethical problems, including virtual reality and AI.
Flick said many consumers do not yet see the daily benefits of AI that venture capitalists and chief executives keep promising. Ashley Striblet, who works in product strategy and consumer AI, has argued that investors blaming weak consumer adoption on environmental criticism are dodging a product problem: people tolerate environmental tradeoffs when they believe a product gives them value.
Data centers make the chip boom visible
The United States has more data centers than any other country and plans for many more, according to the reporting. As companies expand hyperscale facilities for AI, local fights have followed over power use, utility bills, noise, pollution and water consumption.
The NAACP has sued xAI, now doing business as SpaceXAI, over air pollution from gas generators installed to power its data centers. The group has also warned technology companies that it is helping local organizations campaign against data center harms. SpaceXAI did not respond to emailed requests for comment, according to The Verge.
Shaolei Ren, an associate professor of electrical and computer engineering at the University of California, Riverside, studies data centers’ effects on nearby air quality and water resources. Ren said AI’s potential benefits, including scientific discovery, should not come at local communities’ expense. He argues that operators could adopt a “community-integrated data center approach” that reduces resource use and pollution.
Ren and colleagues at UCR and Caltech estimated in a 2024 preprint that training a model the size of Meta’s Llama 3.1 could produce air pollution comparable to 10,000 round-trip car journeys between Los Angeles and New York City. The study estimated public health costs from AI adoption could exceed $20 billion by 2028 and reach 1,300 premature deaths annually from air pollution by 2030. Meta declined on-record comment and pointed The Verge to its sustainability report and data center webpage.
Power and water demand are climbing
A 2024 Lawrence Berkeley National Laboratory study estimated that electricity use by GPU-accelerated AI servers in U.S. data centers rose from 2 terawatt-hours in 2017 to more than 40 terawatt-hours in 2023. The same study projected annual use could reach 165 to 326 terawatt-hours by 2028.
Alex de Vries-Gao, a PhD candidate at Vrije Universiteit Amsterdam, estimated that AI likely exceeded Bitcoin mining’s electricity consumption in 2025 and accounted for nearly half of global data center electricity use. He estimated AI-related carbon emissions at 32.6 million to 79.7 million tons a year.
Water is the other constraint. De Vries-Gao estimated AI could have used 312.5 billion to 764.6 billion liters of water in 2025, including water for electricity generation and data center cooling. Ren said annual totals can hide peak demand, because data centers use much more cooling water during hot periods. A preprint coauthored by Ren estimated U.S. data centers could need up to 1,451 million gallons per day of new peak water capacity through 2030, at a cost of as much as $10 billion.
The footprint starts before the server rack
Sophia Falk, a PhD candidate at Bonn University, and David Ekchajzer, a PhD student at the Université Paris-Saclay, have studied GPUs by physically breaking down donated cards to measure their material footprint. Falk’s research on Nvidia’s A100 found the chip was 90 percent heavy metals and silicon, led by copper, iron, tin and nickel.
Falk’s work estimated that one A100 GPU contains about 1.4 kilograms of copper. Training GPT-4 could have required between 1,174 and 8,800 A100 GPUs, which Falk and other researchers estimated could translate into up to 7 tons of toxic elements extracted and eventually discarded.
Another study coauthored by Falk and Ekchajzer assessed 16 environmental impacts from training GPT-4 with A100s. It found the GPU itself dominated 10 categories, with manufacturing responsible for most public health risks from toxic chemicals and nearly all of the cancer risk.
The pattern is familiar to anyone who has read semiconductor history instead of a launch deck. Santa Clara County, the center of Silicon Valley, has more Superfund sites than any other U.S. county because chip manufacturing chemicals contaminated sites there. AI did not create that legacy, but the GPU buildout is adding fresh demand to a supply chain that already had a dirty bill of materials.
This story draws on original reporting from The Verge.
