A new study warns that the rapid expansion of artificial intelligence could carry a significant and largely hidden environmental cost. By 2025, AI systems running in data centres could generate between 32.6 million and 79.7 million tonnes of carbon dioxide annually, placing the AI carbon footprint on par with the total emissions of major cities or even entire countries. At the upper end of the estimate, AI-related emissions could approach 80 million tonnes of CO₂, comparable to New York City’s reported 52.2 million tonnes in 2023. Even the lower bound aligns closely with Norway’s national emissions, estimated at around 31.5 million tonnes, highlighting the scale of AI’s growing climate impact.
The study focuses on data centres, the backbone of AI, cloud computing, and digital services such as video streaming. These facilities house vast numbers of servers that consume large amounts of electricity and generate intense heat. To function safely, many rely on water-based cooling systems, linking AI growth not only to energy demand but also to water use. As AI adoption accelerates across sectors, the demand for new and expanded data centres has surged, amplifying both electricity consumption and water requirements. This dual pressure makes the AI carbon footprint a complex challenge that extends beyond emissions alone.
In addition to carbon emissions, the research estimates that AI-related water consumption in 2025 could range from 312.5 to 764.6 billion litres—an amount comparable to the entire global bottled water industry’s annual use. This figure includes direct water use for cooling data centres and indirect water consumption associated with electricity generation. Crucially, the study finds that indirect water use can be up to four times higher than direct use, yet this is rarely disclosed by technology companies. As a result, the true scale of AI’s environmental demands remains poorly understood, reinforcing concerns about the underestimated AI carbon footprint and its associated resource impacts.
Geography plays an important role in shaping AI’s environmental intensity. Europe hosts around 15 percent of the world’s data centres, second only to the United States at roughly 45 percent. The report highlights Europe’s relative advantage due to cleaner electricity grids. European power systems average about 174 grams of CO₂ per kilowatt-hour, less than half the global average and significantly lower than the United States. This cleaner energy mix means that data centres in Europe generally have a smaller carbon footprint per unit of electricity consumed, partially mitigating the overall AI carbon footprint compared with regions more reliant on fossil fuels.
A central finding of the research is the lack of transparency among major technology companies. An examination of environmental reports from nine firms—including Amazon, Apple, Google, Meta, Microsoft, and others—found that none report AI-specific environmental metrics. While several companies acknowledge that AI is driving sharp increases in electricity use, they do not distinguish between AI and non-AI computing activities. This forces researchers to rely on top-down estimates that combine public sustainability reports with assumptions about AI energy demand and grid carbon intensity, creating significant uncertainty.
The study concludes by calling for stronger policies requiring detailed disclosure of AI-related environmental data. Recommended measures include reporting the locations of AI systems, the scale of operations at each site, and water usage effectiveness values for individual facilities. Without such transparency, accurately measuring and managing the environmental impact of AI—and responsibly addressing its rapidly growing carbon and water demands—will remain a major challenge.