The GPU's environmental cost goes far beyond AI data centers

The Verge article, written by Justine Calma, starts from an uncomfortable question: how do you decide which uses of a GPU's energy are worthwhile and which are not? The visual trigger is Microsoft's Fairwater data center in Mount Pleasant, Wisconsin, built on the site of a failed Foxconn LCD display project and described as the most powerful AI data center in the world. There, two-story racks are stacked with 72 Nvidia B200 Blackwell GPUs each, chips that in their maximum configuration can consume 1,200W. The text compares that figure with other everyday GPUs: the Nvidia GeForce RTX 5090 for gaming (575W maximum TDP, also based on the Blackwell architecture), the AMD GPU in the PlayStation 5 (about 220W, with nearly 92 million consoles sold), and Apple's A19 chip, with an integrated GPU and Neural Engine, which in an iPhone may consume just a few watts while browsing or 5W in demanding games, but which, multiplied by the roughly 1.5 billion iPhones active worldwide, also represents significant energy use from battery charging.

The report reviews the technical origin of the GPU: it was born as a companion to the CPU to offload graphics rendering work, and thanks to its parallel processing capacity, Nvidia began about 20 years ago to enable general-purpose computing (GPGPU), which accelerated the training of neural networks and, ultimately, modern AI. Nvidia claims to have launched "the world's first GPU" in 1999, although some trace the origins of graphics hardware back to arcade machines of the 1970s. Catherine Flick, a professor of ethics and games technology at Staffordshire University, stresses that many hardware developments with significant ethical implications—from virtual reality to AI itself—began as innovations tied to video games.

A chart of Nvidia's quarterly revenue by market (fiscal years 2018-2026) shows the shift in the business: data center revenue went from $606 million in 2018 to $62.314 billion in 2026, far surpassing the sum of the other segments (gaming, professional visualization, automotive and robotics, original equipment manufacturing and intellectual property), which in 2026 barely reached $5.813 billion.

The article addresses the debate over why AI attracts so much environmental scrutiny when other industries, such as fast fashion, also pollute. Some venture capital investors have gone so far as to blame bad press about AI's environmental impact for consumers' slow adoption of the product. Flick responds with irony ("isn't it nice to have the environment as a scapegoat?") and Ashley Striblet, who works in product strategy and consumer AI, notes in a TikTok video that people keep buying cheap clothing from brands like Shein despite knowing it is not sustainable, because they perceive real value in return; by contrast, many consumers still do not see in generative AI the benefits promised by executives and investors.

Unlike textile or semiconductor manufacturing, historically concentrated in Asia, the AI data center boom is happening inside the United States, driving up local electricity bills and generating pollution near where the very consumers the tech companies are trying to attract live. The text mentions that, in some cases, these centers are installed in low-income neighborhoods and communities of color that already historically face disproportionate environmental burdens: the NAACP has sued xAI (which operates as SpaceXAI) over air pollution generated by gas turbines installed next to its data centers, and the organization has warned other communities across the country to stay alert to similar projects.

A central part of the article draws on the research of Shaolei Ren, an associate professor of electrical and computer engineering at the University of California Riverside, who grew up in a coal-mining region of northern China during the 1980s, where his family kept the windows closed because of the soot and rationed water given the poor water infrastructure. Ren now researches the impact of data centers on the air quality and water resources of nearby communities, and proposes what he calls a "community-integrated data center approach" to prevent these facilities from harming residents, although he acknowledges that this is not the norm today.

Among the figures the report cites: a 2024 preprint study by Ren together with researchers from UCR and Caltech estimated that the energy needed to train a model the size of Meta's Llama 3.1 can generate as much air pollution as 10,000 round trips by car between Los Angeles and New York, the equivalent of about 93 years of driving. The same study calculated that the public health costs associated with the growing adoption of AI could exceed $20 billion in 2028 and cause 1,300 premature deaths per year from air pollution in 2030 (Meta declined to comment and referred to its sustainability report).

For context, the article recalls that video games consumed about 34 terawatt-hours (TWh) a year in the United States, with emissions equivalent to about 5 million cars on the road (24 million tons of CO2), according to a 2019 study, although the most recent consoles are more energy-intensive. By contrast, the consumption of AI servers with GPUs in the U.S. went from 2 TWh in 2017 to more than 40 TWh in 2023, according to a 2024 study by the Lawrence Berkeley National Laboratory, which projects consumption of between 165 and 326 TWh annually in 2028; the lower figure would be equivalent to the annual consumption of more than 8.7 million American households, or—according to the EPA's equivalency calculator—to burning 72.1 billion pounds of coal or charging 5.25 trillion smartphones.

In 2025, AI probably surpassed the electricity consumption of Bitcoin mining, coming to represent nearly half of all electricity used by data centers worldwide, according to a study by Alex de Vries-Gao, a doctoral candidate at the Institute for Environmental Studies at the Vrije Universiteit in Amsterdam, who estimates emissions of between 32.6 and 79.7 million tons of CO2 annually—for comparison, the climate pollution of New York City is around 50 million tons of CO2 a year.

Water consumption is another central axis: de Vries-Gao calculates that AI may have used between 312.5 and 764.6 billion liters of water in 2025, a figure on the order of what humanity consumes in bottled water each year, exceeding even Ren's 2023 projection, which estimated up to 600 billion liters by 2027. Ren warns that annual totals conceal consumption peaks: while a household may use between 1.5 and 2.5 times more water at moments of maximum demand, a data center can use between 6 and 10 times more, and some large-scale projects up to 30 times more, precisely when temperatures—and therefore drought—are most severe. According to a preprint co-authored by Ren, U.S. data centers could need up to 1.451 billion gallons per day of new peak water capacity through 2030, at a cost of up to $10 billion, a burden difficult to bear for small community water systems with limited funding. Ren also questions whether corporate promises to recycle water or replenish nearby sources are sufficient, and calls for more transparency about peak-hour water use so that communities can plan alongside tech companies without being the ones stuck with the bill.

The report closes (incompletely in the available material) by addressing the "cradle" environmental impact, that is, from the very manufacturing of the GPUs. Sophia Falk, a doctoral candidate at the University of Bonn, and David Ekchajzer, a doctoral candidate at the Université Paris-Saclay, have ground up tens of thousands of dollars' worth of donated chips in industrial blenders to analyze their material composition. Their study of the Nvidia A100 found that 90% of its materials are heavy metals—mainly copper, iron, tin and nickel—in addition to silicon; the text was beginning to detail the role of copper, also widely used in electrical transmission lines, when the downloaded content cuts off. Falk insists that AI's impact goes beyond carbon and that much of this material cost remains invisible to those who simply type a query into a chatbot.

Overall, the article does not offer a single moral conclusion on whether the energy expenditure of GPUs (in AI, video games or smartphones) is justified or not, but rather documents, with figures and testimony from researchers, the real size of that cost—in air, water, public health and materials—and calls for transparency mechanisms and community planning before the expansion of data centers continues to accelerate.