Who’s afraid of the big, bad GPU?

How does AI make you feel? Are you excited to “vibe-code” your smart home? Or anxious about all the added pollution and billions of gallons of water used by data centers? Dig a little deeper and you’ll start to question the actual value of the GPUs that underpin all the leaps and promises of generative AI.   Right now, GPUs, hundreds of thousands of them, are being crammed into data centers around the world to power the AI boom. These chips are also found in everything from smartphones to cars to gaming PCs. Nvidia — once a niche chipmaker that has become the world’s most valuable company — still brags about releasing what it calls “the world’s first GPU” and a “gaming breakthrough” in 1999, although some trace the origins of the GPU back to at least the ’70s with graphics hardware used in arcade games.  “There are massive hardware developments that start because of games,” says Catherine Flick, a professor of ethics and games technology at University of Staffordshire. And whether it’s a testbed for new graphics processing units or the development of virtual reality and AI, “all these sorts of things that have quite significant ethical issues, a lot of these start with games,” Flick says. So since its inception, the GPU has been at the root of some of the biggest ethical questions new technologies pose. What impact do games, and now AI chatbots, have on how we interact with the world around us? Whether mining for raw materials or exposing workers to toxic chemicals in semiconductor factories, GPU manufacturing can leave behind a big mess. Collectively, GPUs warehoused in data centers burn through an enormous amount of water and energy, which can lead to more air pollution and greenhouse gas emissions causing climate change. And at the end of its life, a GPU can do even more damage in the form of e-waste.   What’s worth taking those risks? Do potential AI-driven advances in weather forecasting or wildlife conservation justify the environmental footprint of a data center? And what about the GPUs in a gaming PC or iPhone — do they deserve just as much scrutiny?  Hitting close to home  While GPUs have been around for a long time, AI has thrust them into the spotlight in a way that, quite literally, hits close to home for many Americans. The US has far more data centers than any other country and has plans to build many more. As tech companies race to expand a new generation of hyperscale data centers for AI, communities are grappling with the prospect of having these hulking warehouses full of servers as their neighbors.  “Isn’t it nice to have the environment as a scapegoat?” As an environmental journalist, I sometimes get comments on my stories about whether these data centers or AI models are being unfairly portrayed as the menacing new bogeyman, especially when other industries — like fast fashion, for example — exacerbate climate change and pollute the environment. And as for GPUs, they’re not just found in data centers, but also in consumer electronics. So why single out AI for its sustainability issues?  Some venture capitalists have blamed AI’s slow consumer uptake on the bad press surrounding its environmental impact. To that, Flick can’t help but chuckle: “Isn’t it nice to have the environment as a scapegoat?”  A similar sentiment is bubbling up on social media. Ashley Striblet, who works in product strategy and consumer AI, recently published a video on TikTok pushing back on those VCs for “blaming consumers” for their own failures to invest in products people actually want. Many people understand that fast fashion or consumer technology can be bad for the environment, Striblet says, but they still buy these things because they feel like they’re getting value from the products. “I think most people know that ordering clothes from Shein or Amazon is not the most environmentally conscious thing to do,” Striblet tells me. “We also know Gen Z at some point were big consumers of these brands that provide really cheap clothing.” The cost savings from buying cheap clothes turns out to be enough of a tradeoff to justify the purchase for many consumers, Flick adds.  On the other hand, many people just aren’t yet seeing the benefits to their lives that VCs and CEOs are promising with AI, Flick and Striblet each tell me. “It’s just rubbish what they’re saying that their technology can do,” Flick says. “It’d be nice if we all had personalized butlers or whatever it is. But how do you get from point A to point B? It’s not with what we’ve got.”  And while the average American has yet to see the kinds of big leaps forward that are being promised, they’re increasingly likely to see the costs of a new AI data center moving next door. For years, hubs for fast fashion or semiconductor manufacturing have been concentrated in Asia. Now, an explosion of data centers springing up in the US is making headlines for raising utility bills and creating more pollution. It’s not happening in some faraway country, but in the same places where big American tech companies are courting more affluent consumers.   To be sure, even in the US, data centers are landing in some places where low-income neighborhoods and communities of color have long had to fight against the environmental injustice of polluters setting up shop at their doorsteps. The NAACP, for example, has sued xAI (now doing business as SpaceXAI) over air pollution from gas generators the company installed on-site to power its massive data centers. (SpaceXAI didn’t respond to emails from The Verge.) The civil rights group had already warned tech companies to “be on alert” as it helps local groups across the nation mount their campaigns.  Growing globally, impacting locally Shaolei Ren has witnessed the costs communities often pay in the name of progress. He’s a researcher who has broadened the study of data centers’ environmental impact beyond climate change to focus on air pollution and water scarcity near the facilities. Having grown up in a coal-mining region of northern China in the 1980s, Ren remembers keeping the windows shut 24/7 to keep out the black carbon accumulating on the streets outside. And with limited water infrastructure in the province, his family stored water in tanks to ration throughout the day. More powerful GPUs are demanding ever more energy, creating air pollution and greenhouse gas emissions in the process. Ren, who is now an associate professor of electrical and computer engineering at the University of California, Riverside (UCR), leads research into the impact data centers have on air quality and water resources for nearby communities — costs that are often invisible each time a user enters a query into a chatbot.  Sitting in his office at UCR on a rare gloomy Southern California afternoon just before a rainstorm, Ren tells me he’s optimistic about the benefits AI can ultimately bring, including aiding scientific discovery. But those gains shouldn’t be made at the expense of local communities, he says. The whiteboard behind him is covered with equations scrawled in blue marker for his research exploring how to tweak data center operations in order to use fewer resources and minimize pollution. Ren has a vision of what he calls a “community-integrated data center approach” to prevent these facilities from harming nearby residents. “It’s definitely doable,” he says.  But for now, that’s not what’s typically happening. More powerful GPUs are demanding ever more energy, creating air pollution and greenhouse gas emissions in the process. Running them around the clock while keeping the hardware from overheating requires lots of water. It all adds up over the lifespan of a GPU, which might only be several years in a data center.  The energy needed to train a model as large as Meta’s Llama 3.1 can lead to as much air pollution as 10,000 round trips by car between Los Angeles and New York City, Ren and colleagues at UCR and Caltech estimated in a 2024 preprint study. (Meta declined to comment on the record and instead referred The Verge to its sustainability report and webpage on data centers.) Public health costs associated with growing adoption of AI could reach more than $20 billion by 2028 and 1,300 premature deaths annually from air pollution by 2030, the study found.  AI is quickly surpassing the energy use and climate impact of other applications for GPUs. Gaming used roughly 34 terawatt-hours (TWh) per year in the US and led to carbon dioxide emissions equivalent to about 5 million cars on the road (about 24 million tons of CO2), according to a comprehensive study published in 2019. Since then, newer consoles have become more energy-intensive, although their climate impact depends a lot on user behavior and how dirty the electricity grid is wherever they’re playing. For a rough comparison, the energy use of GPU-accelerated AI servers in data centers grew from 2TWh in 2017 to more than 40TWh in 2023 in the US, according to a 2024 study by the Lawrence Berkeley National Laboratory. Now, with all the hype, AI servers’ annual power consumption could grow to between 165 and 326TWh by 2028, the study predicted. The lower estimate would be roughly equivalent to the energy more than 8.7 million homes in the US might use in a year. In 2025, AI likely exceeded the power consumption of Bitcoin mining, accounting for nearly half of all the electricity data centers used around the world, according to a study by Alex de Vries-Gao, a PhD candidate at Vrije Universiteit Amsterdam Institute for Environmental Studies. The resulting carbon emissions likely reached between 32.6 million and 79.7 million tons annually, he estimates. For comparison, New York City’s climate pollution reaches around 50 million tons of CO2 annually. In Southern California, where both Ren and I live and work, massive warehouses storing and sorting Americans’ e-commerce orders increasingly define the region. Goods from Asia arrive by ship in the nearby Port of Los Angeles before making their way inland to the “dry ports” made up of Amazon fulfillment centers and similar facilities buzzing with big rigs around the clock.  “When we use AI we tend to forget that this has an impact on the real world around us. This goes beyond carbon [emissions], and it’s kind of out of sight, out of mind.” The boom in new data centers reminds me of some of the same challenges these warehouses brought. National environmental groups rally around the climate impact of fast fashion and online shopping, while residents mount campaigns over local air pollution. Retirees suddenly see their quiet neighborhoods transformed by blocks of noisy industrial complexes. The pollution and noise complaints just come from the whirring of generators and cooling systems at data centers rather than truck traffic surrounding warehouses. But data centers raise an added concern, especially in dry regions out west.  Powering AI is an incredibly thirsty business, requiring water for electricity generation and cooling systems at data centers. De Vries-Gao estimates that AI could have used between 312.5 billion and 764.6 billion liters of water in 2025, in the range of how much people consume globally in water bottles each year. It could be an even bigger jump in water demand than Ren predicted in a 2023 study, when he and his colleagues estimated that water use could climb to as high as 600 billion liters in 2027. As eye-popping as those numbers can be, looking at an annual total can still be an incomplete picture, Ren asserts. That’s because the amount of water a facility uses is “spikey.” They tend to use the most water for cooling when temperatures rise — a demand spike that can suddenly put a ton of stress on a local water district at a time when the community might also be more prone to drought and water scarcity.  During water demand peaks, a home might use 1.5 to 2.5 times more water than it typically would. A data center, on the other hand, might use 6 to 10 times as much — with some massive new data center projects potentially needing 30 times more water, Ren says. If these trends continue, US data centers could require up to 1,451 million gallons per day of new peak water capacity through 2030, according to a recent preprint paper Ren coauthored. Meeting that demand would cost as much as $10 billion. Those are difficult costs to swallow for small community water systems that are often underfunded and may struggle to upgrade their infrastructure without additional support. Even tech companies’ pledges to recycle water at data centers or replenish water sources nearby miss the mark, Ren argues. The public needs more transparency about water usage during peak demand so they can plan ahead. Communities should be able to work with tech companies to build up the infrastructure needed to increase capacity, he says, making sure residents aren’t left with the tab.  From the cradle In order to better understand the environmental impact of GPUs, Sophia Falk, a PhD candidate at Bonn University, David Ekchajzer, a PhD student at the Université Paris-Saclay, and their colleagues have thrown tens of thousands of dollars’ worth of the chips into industrial blenders. They’re still hoping for more donated graphics cards to study what they’re made of and the footprint they have on the planet. While data centers have been in the spotlight lately, they’re far from the only way GPUs can take an environmental toll. There are the metals and chemicals used to manufacture them, and pollution across the entire supply chain.  “When we use AI we tend to forget that this has an impact on the real world around us,” Falk says. “This goes beyond carbon [emissions], and it’s kind of out of sight, out of mind.”  Falk has studied the material footprint of the Nvidia A100, 90 percent of which she and her coauthors found to be composed of heavy metals — primarily copper, iron, tin, and nickel — as well as silicon. Copper, a material that conducts electricity well and is widely used in power lines and electronics, makes up the biggest chunk. All the hype around generative AI and the wave of new data centers being built is exacerbating a global shortfall in copper supply (also fueled by the electrification of homes, buildings, and transportation).  Demand for the material could grow by 24 percent over the next decade, and new mines would have to open twice as fast as they did a decade ago, Wood Mackenzie estimates. One of the biggest risks with large-scale metal mines, including copper, is acid drainage that can contaminate nearby water sources. Sulfides exposed during the mining process react with water and air to form sulfuric acid. While a single A100 GPU might require about 1.4 kilograms of copper, the impacts multiply when used to train or run a generative AI model. Training GPT-4, for example, could have required between 1,174 an