A new UN report is designed to deliver the most comprehensive view yet of the environmental costs of artificial intelligence – not just its burgeoning electricity use and carbon emissions but also its water and land footprints, its e-waste consequences, as well as the unjust distribution of AI's benefits and burdens worldwide.
According to Environmental Cost of AI’s Energy Use: Carbon, Water and Land Footprints, from the United Nations University Institute for Water, Environment and Health (UNU-INWEH): “One of the most consequential dimensions of AI that remains comparatively under-examined is its environmental footprint and the justice implications that follow."
Its expansion involves “physical infrastructure and supply chains, including data centers, chips, electricity generation, cooling systems, water withdrawals, land occupation, critical minerals, and eventual e-waste.”
This report, “is a step forward in addressing the current gap in AI's environmental governance by assessing its environmental footprints. The investigation goes beyond the carbon-only lens... It examines AI's indirect environmental footprints through energy use, quantifying the carbon, water and land footprints associated with generating the electricity required to operate AI at scale, and highlighting how outcomes vary substantially by location depending on electricity supply mixes.”
“This matters,” the report adds, “because ‘low-carbon’ is not automatically ‘low-water’ or ‘low-land,’ and evaluating sustainability through a single metric can hide trade-offs and shift burdens onto places already facing water stress or land pressure. These asymmetries can reinforce the environmental problems of local communities while strategic advantages of AI flow elsewhere.”
According to the report, expenditures on AI this year are projected to exceed USD 2.5 trillion and the global market is foreseen growing from USD 189 billion in 2023 to nearly USD 5 trillion by 2033, a 25-fold increase in less than a decade.
Reflected in that surge are sobering energy consumption statistics and insights. For example, if data centers, the physical backbone of AI, were a country their estimated 448 terawatt-hours (448 billion kWh) of electricity consumption in 2025 would rank them 11th globally, roughly on par with France.
AI-related workloads accounted for roughly 20% of total data center electricity use in 2025. If that share rises to the expected 40% by 2030, AI-related electricity consumption could reach approximately 374 TWh. On current trajectories that figure could roughly double to 945 TWh by 2030, accounting for almost 3% of projected global electricity use, or enough to supply power to all 1.3 billion people in Sub-Saharan Africa for over 5 years.
Depending on how that electricity is generated, associated emissions could reach 400 million tonnes of CO₂e, comparable to the UK’s emissions from all sectors in 2025.
The associated land footprint of generating that electricity in 2030 would exceed 14,000 km², roughly the area of Northern Ireland.
Meanwhile, the estimated 9.3 trillion liters of water used by data centers, would meet the drinking water needs of Earth’s 8.1 billion people for about 1.6 years.
The report notes that, even when some withdrawn water is returned, “large-scale withdrawals can strain aquifers and river systems, particularly in arid or groundwater-depleted regions.”
Training is only the beginning
Training new AI models requires immense energy. The estimated 100 GWh of electricity required to train Chat GPT-5 roughly equals the annual residential usage of 770,000 people in Sub-Saharan Africa (60% of the region’s population); the associated water footprint is estimated at 1 billion liters and a land footprint of 1.5 km2 of land, or roughly the size of 215 football fields. While these numbers are significant, the UN scientists now warn that the footprint of AI’s daily use is far bigger.
ChatGPT alone is estimated to process around 2.5 billion prompts per day. At a conservative 0.42 Wh per text prompt, that translates into roughly 383 GWh of electricity per year. The related annual water footprint would be equal to the minimum annual domestic water needs of some 500,000 people in Sub-Saharan Africa, and the land footprint exceeds 800 football fields.
“The numbers grow drastically once the AI embedded in mass platforms (such as Google Search) is counted,” the report says. “Crucially, per-use energy varies by orders of magnitude across modalities and output lengths, so product defaults and user choices are footprint determinants.”
It notes that Google processes an estimated 5 trillion searches annually and a conventional search uses about 0.3 Wh. An AI-enhanced generative search uses up to 3 Wh, a 10-fold increase.
As the report explains, “Every kilowatt-hour of electricity used to train or run an AI model carries environmental footprints, including a carbon footprint from the generation mix; a water footprint from electricity production and cooling; and a land footprint from energy infrastructure, reservoirs, and fuel extraction. These three footprints do not always shift in the same direction.”
“For example, switching from coal to bioenergy can, on average, reduce the carbon footprint by 72%, but this comes at the cost of much larger water and land footprints. On average, the water footprint of bioenergy is more than 30 times greater than that of coal and its land footprint is 100 times greater. In different regions and countries, electricity is produced from various sources. The environmental footprint of energy production in a given location depends on the share of each source in its electricity supply portfolio.”
Video generation as an emerging environmental crisis
Meanwhile, a single high-resolution AI video clip can require more than 415 Wh, making it more energy-intensive than the creation of hundreds of AI images. When resolution and frame count are factored in, energy requirements rise quadratically (double the output quadruples the energy used). And as video gets embedded in mainstream platforms, this quickly becomes an infrastructure-scale problem.
The report also underlines the growing problem of AI hardware waste.
“At the end of life, poorly managed e-waste can expose frontline communities to hazardous substances. By 2030, AI infrastructure could generate up to 2.5 million metric tons of e-waste each year, roughly equivalent to discarding 250 Eiffel Towers annually.
The findings show that responsible AI requires full value-chain governance, from mineral sourcing to recycling and safe disposal.
An uneven distribution of benefits and burdens
The minerals powering AI hardware are often extracted in ways that cause concentrated environmental and social harm, particularly in the Global South and in regions with weak regulatory oversight.
The new report underscores a structural inequity at the heart of the AI boom. Frontier AI infrastructure is concentrated in a small number of locations. Countries that lack domestic compute capacity depend on external providers, giving them little control over access, pricing, or data governance. The result is a widening digital divide between nations that build and control AI systems and those that simply consume them while often bearing a disproportionate share of the environmental costs.
(Related: the recent UNU-INWEH report Critical Minerals, Water Insecurity and Injustice).
Further points
Low-carbon is not low-impact
Brazil's hydro grid produces electricity 77% below the global carbon average, but its water and land footprints are nearly triple the global mean.
The UK's grid has a land footprint more than four times the global average. The report directly challenges the assumption that renewable-powered data centers are always green or sustainable, a finding that cuts against a lot of current industry messaging.
The Jevons Paradox trap
The report underlines that efficiency gains alone will not reduce AI's total environmental footprint. Lower costs drive higher volumes of use, potentially erasing all savings. It calls explicitly for resource budgets — caps on tokens, GPU-hours, or kilowatt-hours — not just better hardware.
AI computing is 90% concentrated in two countries
Only 32 nations host AI-specialized cloud infrastructure, and 90% of that capacity is in the US and China. More than 150 countries have no sovereign AI computing at all. The report frames this not just as an economic divide but as an environmental justice issue: excluded countries bear mineral extraction and e-waste burdens while the strategic benefits flow elsewhere.
Ireland as a live cautionary example
Data centers now account for 21% of Ireland's total metered electricity, up from 5% in 2015, exceeding all urban household consumption combined. The national grid operator has paused new approvals around Dublin until 2028. It's a concrete, documented example of what happens when AI infrastructure growth outpaces energy planning — and a preview of what other countries are heading toward.
A roadmap for responsible AI
The report calls for a responsible AI ecosystem built on six principles: transparency; efficiency by design; equity and environmental justice; lifecycle responsibility; global cooperation; and sustainable use. Practical recommendations are directed at each major group of stakeholders:
Governments should integrate AI infrastructure into energy planning, water governance, and land-use permitting, and require standardized environmental footprint reporting.
Industry and AI developers should treat model selection, default outputs, and routing decisions as footprint determinants, and improve efficiency by design.
Users and deploying organizations should adopt fit-for-purpose use — selecting the lightest model and lowest-energy format that meets the task.
Data center operators and utilities should treat siting and energy procurement as environmental footprint decisions, and apply cumulative impact assessment.
Investors should treat electricity, carbon, water and land footprints as material risks in AI infrastructure portfolios.
Communities and civil society should be involved early in data center siting decisions, with enforceable transparency and grievance mechanisms.
International institutions should support harmonized measurement standards, reduce incentives for cross-border burden shifting, and build compute capacity in excluded regions.
"Concise mode"
The report warns that even the language used by AI users can make a huge difference. Simply getting rid of politeness by not saying “please” and “thank you” can reduce the overall footprint significantly by making the prompts more concise. For example, a concise response mode can reduce ChatGPT token output by 30%, saving 87-98 GWh of electricity per year, equivalent to the annual residential electricity of nearly 760,000 people in Sub-Saharan Africa. The report reframes user behavior and product design as environmental governance tools, not just convenience features.
"Technological advancement must remain environmentally manageable," the report states, and that requires measuring, disclosing, and acting on the full footprint, not just the carbon portion.
Less visible public engagement with AI
Netflix, one of the world's largest video streaming services, offers an example of how AI is embedded in daily digital interactions. While users may not associate Netflix with AI directly, the platform uses machine learning models and real-time processing systems for personalized recommendations, content delivery optimization, and dynamic compression to reduce data use.
In the financial sector, generative AI-driven applications automate customer service, but also improve fraud detection and risk assessment. In healthcare, AI is employed in diagnostics, medical imaging, and patient risk prediction—improving speed and precision of care, while reducing treatment costs.
With an estimated 4.5 billion people globally lacking essential healthcare and an expected shortfall of 11 million healthcare workers by 2030, AI has the potential to narrow these critical gaps, particularly in underserved communities where resources are scarce.
Some estimates suggest that partially autonomous vehicles could account for one in 10 new vehicle sales by 2030, as systems become better at interpreting environments and travel routes and customers gain confidence in safety. Robotaxis are already giving 1.3 million rides each month, mostly in the U.S., but also in China, UAE, Singapore, Japan, and other countries, highlighting the potential for deployment worldwide.
An increasingly polarized global workforce
The report warns that, “without deliberate intervention, the global workforce could become increasingly polarized, divided by access to AI technologies and related workforce skills. Those with fewer training opportunities are especially vulnerable to the changes AI is bringing. While job disruption is a visible consequence of AI deployment, the technology's influence extends far beyond the workplace, into realms of warfare, ethics, and even existential risk.”
The report concludes: “AI offers remarkable potential, but fulfilling this promise responsibly requires systemic change. 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. Real progress depends on embedding sustainability at every level, from hardware and model design to deployment, governance, and public use. By committing to transparency, engineering for efficiency, choosing wisely as users and institutions, protecting communities that face disproportionate burdens, and cooperating across borders, society can ensure that progress in intelligence is matched by progress in care. Responsible AI is possible when capability and stewardship grow together within planetary limits.”
