Artificial intelligence is rapidly transforming economies, industries and everyday life, but a new report from the United Nations University Institute for Water, Environment and Health (UNU-INWEH) warns that the technology's environmental footprint is expanding alongside its adoption.
The report, Environmental Cost of AI's Energy Use: Carbon, Water and Land Footprints, examines the environmental impacts associated with the electricity needed to train and operate artificial intelligence systems. While discussions around AI often focus on carbon emissions, the study highlights that water use and land occupation are also important dimensions of the technology's growth.
The authors note that water demands are linked not only to electricity generation but also to cooling systems used in data centers
According to the report, the global network of data centers that supports AI and other digital services consumed an estimated 448 terawatt hours of electricity in 2025. The associated water footprint reached approximately 4.5 trillion liters, an amount the authors say could meet the annual basic domestic water needs of more than 600 million people in Sub Saharan Africa.
Looking ahead, the report projects that global data center electricity consumption could exceed 945 terawatt hours by 2030. Under that scenario, the associated water footprint would rise to 9.3 trillion liters, enough to meet the minimum annual domestic water needs of the entire population of Sub Saharan Africa for one year.
The authors note that water demands are linked not only to electricity generation but also to cooling systems used in data centers. As AI deployment expands, these requirements could place additional pressure on water resources, particularly in regions already facing water stress.
Training advanced models requires substantial resources
The study emphasizes that environmental impacts vary significantly depending on where the electricity is generated
The report estimates that training frontier AI models can require very large amounts of electricity and water.
GPT 4, for example, is estimated to have consumed between 50 and 70 gigawatt hours of electricity during training. The associated water footprint was approximately 600 million liters, enough to meet the minimum annual domestic water needs of about 81,000 people in Sub Saharan Africa.
Future generations of large AI models may require even more resources. The report estimates that a GPT 5 scale model could require around 100 gigawatt hours of electricity and about 1 billion liters of water during training.
The study emphasizes that environmental impacts vary significantly depending on where the electricity is generated. Different energy sources carry different carbon, water and land footprints, meaning that the location of AI infrastructure can strongly influence its overall environmental performance.
Daily AI use now drives most energy demand
While training large models attracts attention, the report suggests that the operational phase, known as inference, accounts for most AI related energy consumption.
Inference refers to the process of generating responses after a model has been deployed. According to the report, this stage represents roughly 80% to 90% of total AI energy use because billions of interactions occur every day.
The report calls for greater transparency and standardized reporting of carbon, water and land footprints
ChatGPT alone is estimated to process around 2.5 billion prompts daily. The study also finds that energy demands vary considerably across different AI tasks. Image generation requires far more energy than text based tasks, while video generation represents one of the most resource intensive applications currently available.
The authors argue that choices made by developers and users, including model selection, output formats and the frequency of use, can influence overall environmental impacts.
Call for broader environmental accounting
UNU-INWEH argues that AI should not be viewed solely through a carbon emissions lens. The report calls for greater transparency and standardized reporting of carbon, water and land footprints associated with AI infrastructure and operations.
The authors recommend that governments integrate AI infrastructure planning into energy, water and land management strategies. They also encourage industry to adopt efficiency by design approaches and consider environmental impacts throughout the lifecycle of AI systems.
As investment in artificial intelligence continues to accelerate, the report concludes that understanding and measuring its environmental footprint will be essential to ensuring that technological progress remains compatible with sustainability objectives and responsible resource management.
