New UN University Report: The Environmental Cost of AI's Electricity Use

As AI's enormous energy consumption triggers cascading impacts on land, water, and climate, researchers at the United Nations University are calling for urgent action in a new report.

By 2030, the electricity consumption of data centers running AI worldwide is projected to reach 945 billion kilowatt-hours. This is nearly three times the annual electricity consumption of Pakistan, Bangladesh, and Nigeria combined (with a total population of over 650 million). The associated water consumption is equivalent to the annual domestic water needs of 1.3 billion people in sub-Saharan Africa, and the land use area exceeds 14,500 square kilometers, roughly twice the size of the Jakarta metropolitan area, home to over 32 million people.

These startling findings are detailed in a new report, 'Environmental Cost of AI’s Energy Use: Carbon, Water and Land Footprints,' published by the United Nations University Institute for Water, Environment and Health (UNU-INWEH). While many researchers have previously warned about greenhouse gas emissions from data centers, the UNU scientists now point out that focusing solely on carbon emissions fails to accurately capture the environmental costs of AI and data centers. The report quantifies the carbon, water, and land 'footprints' of AI's electricity use globally and reveals significant variations across 20 major data center hubs worldwide.

'This report is not intended to oppose the technological innovation of AI, which enriches the lives of billions of people around the world,' said Dr. Kaveh Madani, Director of UNU-INWEH, who led the research team. 'We are calling for the responsible use of AI and proactive management of its unintended impacts to make AI sustainable and equitable. AI represents a technological revolution of our time, but we have very little time left to ensure that the infrastructure supporting it develops within the Earth's environmental limits and that the benefits of AI reach people living in regions with mineral reserves needed for AI development, as well as those in areas hosting data centers and electronic waste.'

Mis-measured 'Footprints'

The report also reveals that the environmental costs of AI are generally miscalculated. Many current assessments focus on the carbon emissions associated with training large models. However, AI systems consume more than just electricity. They also generate water footprints from cooling and power generation, and land footprints from energy infrastructure and supply chains. These three footprints are not necessarily proportional. For example, switching from coal to bioenergy can reduce the carbon footprint of electricity by an average of 70%, but it can increase the water footprint by over 30 times and the land footprint by 100 times. The report concludes that being 'low-carbon' does not automatically lead to lower water and land use, warning that evaluating AI sustainability with a single metric risks overlooking hidden trade-offs and imposing additional environmental burdens on regions already facing water and land scarcity.

From an infrastructure perspective, the numbers quickly escalate. In 2025, global data centers consumed an estimated 448 billion kilowatt-hours of electricity. If all the world's data centers were treated as a single country, they would be the 11th largest electricity consumer globally, behind France (10th) and ahead of Saudi Arabia (12th).

'What surprised us most was that options that seemed most environmentally friendly from a decarbonization perspective were often worse for water and land,' said Dr. Miriam Akzel, lead author of the report from UNU-INWEH. 'If we continue to judge AI's sustainability solely by carbon emissions, we will assume that using renewable energy makes AI infrastructure clean. But this means solving one problem while creating another, and the burden often falls on unexpected regions.'

AI's Continuous Use Consumes More Power Than Development

Much of the previous discussion on AI's environmental impact has focused on the energy needed to train large models. Training GPT-3 consumed about 1.3 billion kilowatt-hours of electricity, and GPT-4 is estimated to have consumed 50-70 billion kilowatt-hours. However, the report reveals that this view is already outdated. Once a model is publicly released, 'inference'—the continuous operation of the model to answer everyday user queries—accounts for 80-90% of AI's energy consumption, making it the largest cost.

ChatGPT alone is estimated to process about 2.5 billion prompts per day, consuming approximately 383 billion kilowatt-hours of electricity annually. Offsetting the associated carbon emissions would require growing 2.6 million tree seedlings for 10 years, covering an area roughly the size of Manhattan Island in the United States. The water footprint is equivalent to the minimum annual domestic water needs of about 500,000 people in sub-Saharan Africa, and the land footprint covers over 800 soccer fields.

Video Generation Emerges as a New Environmental Crisis

The energy consumed per AI processing task varies dramatically by task type. A typical chat-based query consumes about 200 times more energy than a basic text classification task. Generating a single image with AI can require about 1,450 times more energy than a classification task. And generating a short video with AI consumes as much electricity as classifying 200,000 spam emails. The model used, prompt length, output format, and resolution all significantly impact the footprint, but many of these settings are determined by default product configurations that users are often unaware of.

Why Efficiency Doesn't Reduce Environmental Impact

The report cites the 'rebound effect' (Jevons paradox), warning that as models become more efficient, costs decrease and usage increases. Without clear limits on token count, resolution, and default output volume, any improvement in per-task efficiency can be easily offset by increased usage.

'Many people think that as AI technology advances and processing becomes more efficient, the environmental footprint of AI will shrink. But that's only one side of the whole problem,' said Professor Madani, co-author of the report and the 2026 Stockholm Water Prize laureate.