The artificial intelligence (AI) boom is often discussed in terms of innovation, productivity and economic growth. A new United Nations report suggests it should also be discussed in terms of water.
According to the report, produced by the United Nations University Institute for Water, Environment and Health (UNU-INWEH), the water footprint associated with the world’s growing network of AI-driven data centres could meet the annual domestic needs of all 1.3 billion people living in sub-Saharan Africa.
The report finds that every prompt, image and video is backed by vast networks of data centres, cooling systems, electricity grids, water withdrawals, land use and mineral extraction. This material footprint is growing at extraordinary speed.
“One of the most consequential dimensions of AI that remains comparatively under-examined is its environmental footprint and the justice implications that follow,” the report said, arguing that AI is not merely software, but a physical system embedded in energy, water and land use.
That footprint is expanding alongside one of the fastest technological adoptions in history.
Since ChatGPT’s launch in 2022, generative AI has moved from novelty to mainstream infrastructure, with hundreds of millions of users now relying on AI tools for everything from search and writing to coding and customer service.
The report is described by its authors as a step toward closing a major gap in AI governance by moving beyond a carbon-only lens. It emphasises that “low-carbon is not automatically low-water or low-land,” warning that single-metric assessments can obscure trade-offs and shift environmental burdens onto already water-stressed or land-stressed regions.
It points to Brazil as an illustration of the trade-offs involved. While hydropower dominates the country’s grid, with a carbon footprint 77% below the global average, its water and land footprints are nearly triple the global mean.
Those trade-offs are already significant. In 2025, data centres, the physical backbone of AI, consumed an estimated 448 terawatt-hours (TWh) of electricity, ranking them 11th globally if treated as a country.
On current trends, that demand could rise to 945 TWh by 2030, nearly triple the combined electricity use of Pakistan, Bangladesh and Nigeria, which together are home to more than 650 million people.
The associated impacts are equally striking: up to 399 million tonnes of CO₂e (carbon dioxide equivalent), a water footprint of 9.3 trillion litres and land use exceeding 14 000 km².
Last year, AI accounted for roughly 20% of data centre electricity use, a share projected to double to 40% by 2030. At that level, AI alone could consume about 378 TWh annually, enough to power the residential needs of sub-Saharan Africa for more than two years.
Factory initiative to expand Europe’s AI capacity, with next-generation GPUs, expanded storage, and energy-efficient cooling
to support large-scale AI training and advanced machine learning for startups, researchers, and public institutions across
Europe. Photo by Steve Jurvetson (Flickr).
The report warns that even where water is eventually returned to the environment, large withdrawals by data centres can place additional pressure on aquifers and river systems, particularly in already water-stressed regions.
The footprint is not just about infrastructure build-out; it is also about how AI is used. Once deployed, systems shift from energy-intensive training to continuous inference at massive scale. In many models, inference now accounts for 80% to 90% of total energy consumption.
Daily usage patterns matter. ChatGPT alone processes around 2.5 billion prompts per day. At an estimated 0.42 Wh per text prompt, that translates into roughly 383 GWh of electricity annually, alongside water and land footprints equivalent to the domestic needs of hundreds of thousands of people and hundreds of football fields of land use.
As the report said, “every prompt, image, video, and query accumulates when multiplied by billions of users.”
Even small design choices become system-wide environmental decisions. A typical ChatGPT-style query can use far more energy than basic text classification, while long responses or image generation can multiply consumption by hundreds or thousands of times.
Video, the report said, is emerging as the most resource-intensive frontier, with a single AI-generated clip consuming energy on a scale comparable to hundreds of images or hundreds of thousands of simple computations.
Efficiency gains alone are not enough. “If lower per-use impacts drive higher volumes of use, total impacts may still rise,” it notes, describing a rebound effect consistent with the Jevons paradox.
The environmental consequences extend beyond energy use. AI hardware depends on critical minerals whose extraction often concentrates environmental and social harm in the Global South.
At the end of life, e-waste is expected to reach 2.5 million metric tons annually by 2030 — roughly equivalent to discarding 250 Eiffel Towers each year — raising concerns about toxic exposure and weak recycling systems in frontline communities.
The report also highlights a stark geopolitical imbalance. Only 32 countries host AI-specialised cloud infrastructure, with 90% concentrated in the United States and China, while more than 150 countries lack sovereign compute capacity.
This, it argues, reinforces a widening digital and environmental divide in which some states capture the benefits of AI while others absorb many of its material costs.
These concerns are echoed in parallel research from the University of Cape Town (UCT). Grant Oosterwyk, a senior lecturer and PhD candidate in Information Systems, warns that the rapid expansion of AI data centres is generating “significant environmental, social and political tensions” that are often obscured by narratives of innovation.
His research identifies five interlinked pressures: an energy paradox, rising water strain, hyperscaler dominance, sovereignty erosion and urban displacement. The “energy paradox” describes how AI simultaneously improves efficiency while dramatically increasing electricity demand.
Estimated footprints per standard-resolution AI image: 2.9 Wh electricity, 1.22 g CO₂e, 28.6 mL of water, and 0.45 cm² of
land, based on literature benchmarks and global average electricity footprint factors. Photo UNU-INWEH
Water strain arises from the heavy use of cooling systems in already stressed regions. Meanwhile, the dominance of large technology firms, or hyperscalers, raises concerns about market concentration and inequality.
Oosterwyk also highlights “sovereignty erosion,” where reliance on foreign-owned infrastructure weakens national control over data governance and digital policy.
“Our findings suggest that AI infrastructure is increasingly becoming a geopolitical issue,” he said, noting that questions of ownership and control are becoming as important as technological capability itself.
Urban displacement adds another layer, as large-scale data centres reshape land use and place pressure on surrounding communities, often with limited local benefit.
“The expansion of AI infrastructure is not occurring in isolation,” Oosterwyk said. “It intersects directly with broader concerns around energy access, climate resilience and environmental sustainability.”
Both UNU-INWEH and UCT researchers converge on a central conclusion: AI is not an abstract digital force, but a materially intensive system embedded in electricity grids, water systems, land use and global supply chains.
The UNU-INWEH report calls for a shift toward lifecycle governance and six operational principles: transparency, efficiency by design, equity and environmental justice, life-cycle responsibility, global cooperation and sustainable use.
It urges governments to integrate AI infrastructure into energy, water and land-use planning, and to require standardised environmental footprint reporting.
It also calls for “fit-for-purpose” AI use — selecting the lightest model and lowest-energy format capable of completing a task — and for treating default settings, output length and routing decisions as environmental variables, not neutral design choices.
UCT’s Oosterwyk similarly argues for integrative policy frameworks that balance innovation with sustainability. “Balancing AI’s transformative potential with environmental sustainability and social responsibility will be one of the defining policy challenges of the coming years,” he said.
The UNU-INWEH report concludes that AI’s promise can only be realised sustainably if its environmental costs are made visible and managed.
“Every interaction draws on finite resources and the total environmental footprint depends on how AI systems are designed, how often they are used and what tasks they perform,” it states.
“Responsible AI is possible when capability and stewardship grow together within planetary limits.”
AI’s rapid growth is driving major water, energy, land and e-waste impacts, raising environmental justice concerns worldwide