How Much Water Does AI Use? Consumption Now Exceeds World’s Bottled Water, Suggests New Study

Artificial intelligence (AI) has a drinking problem, but it is not what you think. A 2025 study has claimed that the water consumption by AI systems via data centres might have exceeded the total global consumption of bottled water. Its carbon footprint is also said to be equivalent to that of New York City this year. These are bold claims, and despite the lack of pinpoint accuracy due to disclosures by companies, if this is true, it highlights the massive environmental impact driven by the global AI demand.

AI Systems Could Have Critical Environmental Impact, Says Study

The peer-reviewed study, titled “The carbon and water footprints of data centres and what this could mean for artificial intelligence,” was published recently. Led by Dutch academician Alex de Vries-Gao, the research aimed to find the environmental impact of AI systems, primarily the data centres running AI workloads. While the paper acknowledges that finding accurate numbers could be difficult since companies do not distinguish between AI and non-AI workloads in environmental reports, there are ways to estimate the global power demand of AI.

The researcher resorted to generalised emission and water consumption data and the environmental reports of data centres released by companies such as Google, Meta, Amazon, and others to get an estimate of the power generation that is being used for AI workloads.

Based on different models, the researchers found that the carbon footprint of AI systems alone could be between 32.6 and 79.7 million tons of CO2 emissions in 2025, which is equivalent to New York City in the same year. This is significant since the city is among the top when it comes to carbon footprint. In 2018, the World Economic Forum found the US city to have the third-highest emissions, after Seoul and Guangzhou.

Similarly, the water footprint of AI workloads powered by data centres is said to be between 312.5 and 764.6 billion litres, a number exceeding the total bottled water consumed on Earth in a year. This means that AI is not just an energy issue, but also a water-security problem.

As per the study, the major driver of environmental impact is not the training of AI models but running inference. Inference, or the compute required to find responses to users' prompts, is causing a higher amount of emissions and requires a high volume of water consumption. This seems rational given millions of daily queries, image and video generations, and assistants running around the clock.

Surprisingly, the rising environmental impact also highlights that despite efforts being taken to make data centres more energy-efficient, the rise in AI demand far surpasses the measures and reduces the overall effectiveness. Put simply, better tech is leading to more usage, not lower impact.

The study makes two critical points. First, due to the vast requirement of energy-hungry infrastructure, AI should no longer be treated as software, and it should warrant a similar level of regulatory environmental oversight as industries such as telecom, aviation, and heavy industries.

Second, the author suggests that the lack of disclosures and transparency of AI workloads in environmental reports is leading to obscurity when it comes to measurement and conservation efforts. The study highlights that AI companies should start revealing these numbers so that policymakers and researchers are not reacting based on outdated data.