A new study estimates that infrastructure for artificial intelligence (AI) will produce carbon emissions globally in 2025 that are roughly equivalent to New York City's emissions for an entire year, while consuming as much water as the world's residents drink from bottled water combined. The study authors pointed out that because the data disclosed by companies is very limited, this assessment is likely to be a relatively "conservative" version, and the actual environmental cost may be higher.
The research was completed by Alex de Vries-Gao, a doctoral student at the Institute of Environmental Studies at VU University in Amsterdam, and published in the academic journal Patterns. He has long tracked the energy consumption of data centers related to AI and cryptocurrency mining. This time, based on previous research, he conducted a comprehensive calculation of AI electricity consumption and the resulting emissions and water use in 2025. He bluntly said that it is almost impossible to come up with extremely precise figures at present, "but no matter what, the scale will be very huge, and in the end it will be everyone who pays the price."
According to earlier research, global AI computing power demand may reach 23 gigawatts in 2025, already exceeding the power consumption of Bitcoin mining in 2024. However, large technology companies usually only disclose overall carbon emissions and direct water consumption in their annual sustainability reports, but rarely break down how many resources the AI business itself consumes. To this end, de Vries-Gao used analyst reports, earnings call minutes and other public information to calculate the production quantity and running power consumption of hardware such as AI chips, and then calculated greenhouse gas emissions and water consumption based on this.
The results show that AI-related systems may emit about 32.6 million to 79.7 million tons of carbon dioxide per year in 2025, with the median value equivalent to New York City's average annual carbon dioxide emissions of about 50 million tons. In terms of water use, AI is expected to consume about 312.5 billion to 764.6 billion liters of water this year, higher than a 2023 study predicting an upper limit of about 600 billion liters in 2027. Ren Shaolei, an associate professor of electrical and computer engineering at the University of California, Riverside, said that the water use estimate in the latest results was "the most surprising" and believed that the analysis was still "quite conservative" in terms of methodology because it only calculated the impact of the equipment during the operation phase and did not include the supply chain and additional environmental costs after the equipment is scrapped.
Data centers are AI’s “big consumers of energy and water.” Servers generate a lot of heat during high-load operation and require a large amount of water to be consumed through cooling systems to prevent overheating. The power plants that power data centers themselves also rely on massive amounts of cooling water. These factors together constitute AI's huge "water footprint." The explosive growth of generative AI has driven the construction of new data centers and the planning of new power plants. If these power plants continue to rely on fossil fuels, they will not only increase water demand, but also further increase greenhouse gas emissions.
In the United States, which leads the world in the number of data centers, many proposed projects there have encountered increasingly strong community opposition. The core focus is on the occupation of power and water resources. Opponents worry that AI data centers will further exacerbate tensions in areas that are already experiencing water shortages or stress on power grids. Researchers point out that even in areas with abundant water resources, the development of concentrated data centers may have long-term impacts on local ecosystems.
Still, the study offers a wide range of forecasts, largely due to a lack of transparency in companies disclosing environmental data. de Vries-Gao found that although many companies publish sustainability reports, they often omit key details, such as the proportion of "indirect water" behind electricity consumption and the specific share of AI business in overall water consumption and emissions. In addition, the power grid structure varies significantly in different regions, and the "cleanliness" of the power source directly affects the emission level corresponding to the same power consumption. Therefore, if the company can more clearly mark the geographical distribution of data centers, it will also help the outside world to more accurately assess the environmental impact of AI expansion.
The study calls for technology companies to be more open and transparent on AI-related carbon emissions and water use data so that the public and policymakers can fully understand the true environmental costs of this technology wave. Ren Shaolei believes that at a time when society's attitude towards AI is becoming increasingly polarized and the debate around water issues is intensifying, this type of work is particularly critical and helps promote fact-based public discussion. de Vries-Gao said that only with more transparent information can society seriously discuss a fundamental question: "Is this the future we want? Is this fair?"