The benefits of AI flow worldwide; the costs concentrate in specific regions.
A new UN University study invites humanity to reckon with a paradox at the heart of its most celebrated technology: the systems we build to expand intelligence are quietly consuming the very resources life depends on. By 2030, AI infrastructure is projected to draw electricity equivalent to nearly three times the combined consumption of Pakistan, Bangladesh, and Nigeria, while its water demands could sustain 1.3 billion people and its land footprint could swallow a region twice the size of greater Jakarta. What makes this moment morally urgent is not merely the scale of consumption, but its geography — the costs fall heaviest on nations that hold the least power to shape the decisions driving them.
- AI's environmental toll is far larger than the carbon debate suggests — water, land, minerals, and e-waste are all being drawn into a system most people interact with daily without seeing its true cost.
- Day-to-day AI usage — billions of prompts processed every hour — drives 80 to 90 percent of total energy demand, dwarfing the one-time cost of training that dominates public concern.
- A rebound effect is quietly undermining efficiency gains: as AI becomes cheaper to run, consumption rises faster than improvements can offset, making technological optimism alone an insufficient response.
- Over 150 nations bear the environmental burden of AI infrastructure — depleted aquifers, toxic e-waste, degraded land — while more than 90 percent of AI computing capacity remains concentrated in the US and China.
- UN University researchers are calling for a 'responsible AI ecosystem' grounded in transparency, lifecycle accountability, and global equity — framing governance choices made now as the decisive factor in whether AI develops within planetary limits.
The machines running artificial intelligence are consuming far more than the world has been told. A new study from UN University maps the full environmental footprint of AI infrastructure, and the picture extends well beyond carbon emissions. By 2030, data centres could draw 945 terawatt-hours of electricity annually — nearly three times the combined usage of Pakistan, Bangladesh, and Nigeria, nations home to over 650 million people.
Electricity is only the beginning. Every kilowatt flowing through a data centre carries hidden costs: water for cooling, land for power generation, minerals extracted to build the hardware. AI-related water consumption could reach the level needed to meet the basic domestic needs of 1.3 billion people each year. The land footprint may exceed 14,500 square kilometres by decade's end. These figures represent real pressure on aquifers in drought-stricken regions and competition for resources in places already stretched thin.
The public debate has focused on the energy cost of training large AI models, but the study reveals that training is almost a rounding error. Daily usage — the billions of prompts processed every single day — accounts for 80 to 90 percent of total energy demand. One widely used AI service alone processes around 2.5 billion prompts daily. The energy cost varies enormously by task: generating a single AI image can demand more than a thousand times the electricity of simple text classification.
The problem is compounded by the rebound effect. As AI becomes more efficient and cheaper to run, people use it more. Efficiency gains are swallowed by rising demand, meaning technological improvement alone may accelerate the crisis rather than resolve it.
The burden is not shared equally. More than 90 percent of the world's AI computing capacity is concentrated in the United States and China, while over 150 nations lack significant domestic infrastructure. Those nations nonetheless bear environmental costs — depleted water supplies, degraded land, and up to 2.5 million tonnes of e-waste annually by 2030, much of it destined for lower-income countries ill-equipped to handle it safely. It is, the researchers suggest, a new form of environmental colonialism.
The UN University team frames their findings not as an argument against AI, but as a call for governance equal to the stakes. They propose a framework built on transparency, efficiency by design, equity, and lifecycle responsibility. Governments must integrate AI infrastructure into energy and land-use planning. Companies must design for minimal resource consumption from the outset. The future of AI, the report argues, will be shaped less by what engineers can build than by the choices made today about who bears the cost — and who reaps the reward.
The machines that power artificial intelligence are thirsty in ways we're only beginning to understand. A new study from UN University has mapped the full environmental footprint of AI infrastructure—and it extends far beyond the carbon emissions that typically dominate the conversation. By 2030, data centres running AI systems could consume 945 terawatt-hours of electricity annually. To put that in perspective: it's nearly three times what Pakistan, Bangladesh, and Nigeria use combined, despite those three nations being home to more than 650 million people.
But electricity is only part of the picture. Every kilowatt that flows through a data centre carries hidden costs—water needed for cooling systems, land required for power generation and supply chains, minerals extracted from the earth to build the hardware itself. According to the UN University research, AI-related water consumption could reach the level needed to meet the basic domestic needs of 1.3 billion people annually by the end of the decade. The land footprint could exceed 14,500 square kilometres, roughly double the size of the Jakarta metropolitan area. These aren't abstract numbers. They represent pressure on aquifers in drought-stricken regions, competition for scarce resources, and environmental degradation in places already stretched thin.
What makes this crisis particularly difficult to see is how we measure environmental harm. The public debate has fixated on greenhouse gas emissions, especially the energy required to train large AI models. But the study reveals that training is almost a rounding error compared to what happens next. Day-to-day usage—the billions of prompts processed every single day—accounts for 80 to 90 percent of total energy demand. One widely used AI service alone processes around 2.5 billion prompts daily, consuming hundreds of gigawatt-hours of electricity each year. The energy cost varies wildly depending on the task: generating a single AI image can demand more than a thousand times the electricity of simple text classification. Video generation pushes the demand even higher.
The problem compounds itself through what economists call the rebound effect. As AI systems become more efficient and cheaper to run, people use them more. Lower costs don't reduce total consumption—they increase it. Efficiency gains get swallowed by rising demand. This means that technological improvements alone won't solve the problem; they may actually accelerate it.
The burden of this infrastructure is not distributed evenly across the globe. While the benefits of AI flow worldwide, the environmental costs concentrate in specific regions. Some countries already see data centres consuming a significant share of their national electricity supply, straining grids that were never designed for this load. In others, expanding facilities are drawing heavily on water supplies during periods of scarcity. The electronic waste problem looms as well: AI infrastructure could generate up to 2.5 million tonnes of e-waste annually by 2030. Much of that waste will end up in lower-income countries with limited capacity to handle it safely. The extraction of critical minerals needed for AI hardware raises additional concerns about environmental damage and social inequity in mining regions.
This uneven distribution reveals a deeper injustice. More than 90 percent of the world's AI-specialised computing capacity is concentrated in just two countries: the United States and China. Over 150 nations lack significant domestic AI infrastructure. This means they bear environmental costs—depleted water supplies, degraded land, toxic waste—without sharing in the economic benefits that AI-driven growth creates. It's a new form of environmental colonialism, where the costs of one nation's technological advancement are exported to another's.
The UN University researchers are careful to frame their findings not as an argument against AI itself, but as a call for urgent action to ensure the technology develops within planetary limits. They propose a framework for a "responsible AI ecosystem" built on transparency, efficiency by design, equity, lifecycle responsibility, global cooperation, and sustainable use. Governments need to integrate AI infrastructure into their energy, water, and land-use planning. Companies need to design systems that minimise resource consumption from the ground up. Users have a role too, choosing lower-impact applications where possible. The future of AI, the report argues, will depend not only on what engineers can build, but on the governance choices made today—choices about who bears the cost and who reaps the reward.
Citações Notáveis
The report is not an argument against AI itself, but a call for urgent action to ensure the technology develops within planetary limits.— UN University researchers
Solutions seen as green in one sense may worsen pressures in others, particularly in regions already facing resource scarcity.— UN University study
A Conversa do Hearth Outra perspectiva sobre a história
When we talk about AI's environmental impact, why has the focus been so narrow—just carbon emissions?
Because carbon is easy to measure and it fits into existing climate frameworks. But AI doesn't just warm the planet; it consumes water and land in ways that hit specific regions hard. A data centre in a drought-prone area doesn't care about global carbon accounting.
You mentioned the rebound effect. Can you explain why efficiency doesn't help here?
When AI gets cheaper and faster, people use it more. You save energy per task, but total usage explodes. It's like making cars more fuel-efficient—people just drive more. The gains disappear.
The study says 80 to 90 percent of energy goes to daily use, not training. That surprised me.
Most people think the big energy cost is training GPT or similar models. But once they're built, millions of people query them constantly. That's where the real consumption lives—in the everyday use that nobody thinks about.
What does it mean that 90 percent of AI computing capacity is in two countries?
It means the United States and China get the economic benefits while countries with water scarcity or fragile grids bear the environmental cost. A data centre in a water-stressed region serves users everywhere, but the local aquifer pays the price.
Is there a way to make this sustainable?
The report doesn't say it's impossible, but it requires real choices: governments planning for this infrastructure, companies designing for efficiency from the start, and honestly confronting who pays and who benefits. Right now, we're just building without asking those questions.
What happens if we don't change course?
By 2030, we're looking at water consumption equivalent to the basic needs of 1.3 billion people, land use the size of a major metropolitan area, and millions of tonnes of toxic waste in countries least equipped to handle it. The technology keeps growing, but the planet doesn't.