Artificial intelligence is consuming vast amounts of energy and water, with new estimates suggesting that in 2025, AI’s carbon emissions could match those of New York City, and its water use could equal the total amount of water contained in every plastic bottle consumed globally in a year. The findings come from a study by Alex de Vries-Gao, a PhD candidate at the VU Amsterdam Institute for Environmental Studies, published in the journal Patterns. The research presents a conservative view of AI’s environmental toll, as it relies on limited public data. Tech companies often do not break down their energy and water use by specific activities like AI, making it difficult to fully assess the impact. De Vries-Gao used a combination of analyst reports, company earnings calls, and public disclosures to estimate the power demand of AI hardware, which he found could reach 23 gigawatts (GW) in 2025—surpassing the electricity used by Bitcoin mining in 2024. Based on this energy use, he calculated that AI’s annual carbon emissions could range from 32.6 to 79.7 million tons. New York City’s annual emissions are about 50 million tons, placing AI’s output in the same ballpark. The water footprint is equally concerning. AI could consume between 312.5 and 764.6 billion liters of water this year, primarily for cooling data center servers and supporting the power plants that supply them. This exceeds earlier projections, including a 2023 study that estimated up to 600 billion liters by 2027. Shaolei Ren, an associate professor at the University of California, Riverside and co-author of the 2023 study, called de Vries-Gao’s work “timely” and noted that the water impact of AI is becoming a growing concern, especially as communities in the U.S. — which hosts the most data centers worldwide — push back against new projects over water and power usage. Ren also emphasized that the new study is still conservative, as it only accounts for the direct operation of AI systems, not the full environmental cost of manufacturing hardware, transporting components, or disposing of devices at the end of their life. The actual impact could be much higher. A major challenge remains the lack of transparency. While companies report total emissions and water use, they rarely specify how much is tied to AI. The environmental cost also varies by location, depending on the cleanliness of the local power grid. More detailed reporting on where data centers are built and how they are powered would help clarify the true impact. De Vries-Gao stresses the need for greater accountability. “We can really ask ourselves, is this how we want it to be? Is this fair?” he says. “We really need transparency so we can start having that discussion.” Without it, the environmental costs of AI’s rapid expansion may continue to grow unseen—while society bears the consequences.