AI has created as much carbon pollution as New York City this year and swallowed as much H20 into water bottles as people consume worldwide, according to new estimates.
This study is based on the relatively limited amount of data currently available to the public, so it probably paints a rather conservative picture of AI's environmental impact. A lack of transparency from technology companies makes it difficult to recognize the potential environmental damage caused by AI becoming part of daily operations, argue study authors who have spent years tracking the power consumption of data centers used for AI and cryptocurrency mining.
“There's no way to put a very precise number on this, but in any case it's going to be very large… In the end, everyone is paying the price for this,” says Alex de Vries Gao, a doctoral candidate at the VU Amsterdam Institute for the Environment, who published a paper today in the journal Environmental Studies. pattern.
“At the end of the day, everyone is paying the price for this.”
To make sense of these numbers, De Vries Gao built on previous research that found that worldwide AI power demand could reach 23GW this year, which could exceed the amount of electricity used to mine Bitcoin in 2024. While many tech companies disclose total numbers for carbon emissions and direct water use in their annual sustainability reports, they typically do not break those numbers down to show how much resources AI consumes. De Vries-Gao found a workaround by using analyst forecasts, company earnings calls, and other public information to measure how much hardware is being produced for AI and how much energy that hardware would use.
Once we know how much power these AI systems are likely to consume, we can now use that to predict the amount of global warming pollution that may be generated. The amount ranged from 32.6 million to 79.7 million tons per year. For comparison, New York City emits about 50 million tons of carbon dioxide per year.
Data centers can also be big consumers of water, which is an issue related to power usage as well. Data center cooling systems use water to prevent servers from overheating. Power plants also require large amounts of water to cool equipment and use steam to turn turbines, making up a large portion of a data center's water usage. The push to build new data centers for generative AI is also driving plans to build more power plants, which will result in increased water use and further increases in greenhouse gas pollution from burning fossil fuels.
De Vries Gao said AI could use between 312.5 billion and 764.6 billion liters of water this year. This is even higher than a previous study conducted in 2023, which estimated that water use could reach 600 billion liters in 2027.
“I think this is the biggest surprise,” said Xiaolei Ren, one of the authors of the 2023 study and an associate professor of electrical and computer engineering at the University of California, Riverside. “[de Vries-Gao’s] “This paper is really timely…especially as opinions on AI and water become increasingly polarized,” Ren added.
Across the United States, which has more such facilities than any other country in the world, local opposition to new data center projects is surging over concerns about water and power usage.
Even with higher water usage projections, Ren said De Vries Gao's analysis is “very conservative” because it only captures the environmental impact of operating AI equipment, and excludes additional impacts that accumulate along the supply chain and at the end of the equipment's life.
The results are quite wide-ranging because companies are not disclosing more accurate data. De Vries-Gao gleaned as much information as possible from sustainability reports, but found that important details were often left out, such as indirect water consumption from electricity demand and how much was specifically used for AI. Emissions and water consumption can vary depending on where a data center is located and how dirty the local power grid is, so being more proactive about where you operate or plan to build new data centers could also shed more light on AI's growing environmental impact.
“We can ask ourselves: Is this what we want? Is this fair?” de Vries-Gao says. “We really need transparency to start a discussion.”
