UN warns AI's hidden environmental cost: water, soil, and massive energy consumption

Artificial intelligence has become so woven into daily life that most people never think about what powers it. A quick search, a generated image, a text composed in seconds—these feel weightless, virtual, costless. But behind every interaction sits a vast physical infrastructure consuming enormous quantities of electricity, water, and minerals. A new United Nations report makes clear that this hidden machinery carries a price tag far larger than most realize, and one that falls unevenly across the world.

The report, led by researchers at the United Nations University's Institute for Water, Environment and Health, attempts something ambitious: measuring AI's full environmental footprint beyond carbon emissions alone. The findings are sobering. A single high-resolution video generated by AI can consume more than 415 watt-hours of electricity—roughly equivalent to the energy required to produce hundreds of AI-generated images. An AI-assisted Google search demands nearly ten times the power of a conventional search. ChatGPT alone processes 2.5 billion queries daily, consuming an estimated 383 gigawatt-hours annually. These numbers accumulate quickly into something staggering: if data centers were a country, they would rank as the world's eleventh-largest electricity consumer in 2025, matching France's total consumption at 448 terawatt-hours per year, with AI accounting for roughly one-fifth of that figure.

What makes the UN analysis distinctive is its refusal to treat energy consumption in isolation. The researchers calculated not just carbon emissions but also water depletion and land use tied to electricity generation. Under current growth trajectories, AI's water footprint alone could reach 9.3 trillion liters annually—enough to meet global freshwater needs for 1.6 years. The territorial footprint would exceed 14,000 square kilometers. These are not abstract figures. In Ireland, data centers already consume 21 percent of the country's electricity, more than urban households, forcing the grid operator to halt new connections in Dublin until 2028. If that proportion reaches 40 percent by 2030, consumption could spike to 374 terawatt-hours. The researchers warn that some data centers have been built in water-stressed communities with aging electrical grids and little political leverage to resist.

Yet the environmental cost tells only part of the story. The benefits and burdens of AI are distributed with stark inequality. Ninety percent of the world's specialized cloud infrastructure for AI is concentrated in the United States and China. More than 150 countries lack any sovereign computing capacity. Meanwhile, the minerals essential to building AI hardware—cobalt, lithium, rare earths—are extracted primarily in the Global South, often under conditions that generate severe environmental and social damage. Communities in these regions absorb the pollution from mining operations while gaining almost none of the economic benefits of the technology they help power. The wealthiest nations develop, deploy, and profit from AI; poorer nations supply the raw materials and inherit the environmental wreckage.

The researchers emphasize that renewable energy does not solve this problem. A data center powered by solar panels may produce low carbon emissions, but it still consumes vast quantities of water and land. Shifting infrastructure to regions with renewable energy can simply displace environmental burdens to areas already under stress. Kaveh Madani, who directed the research team, stresses that "AI does not have a single footprint." Its impact depends on the specific task, the size of the model, where computation occurs, and what energy source feeds it. A brief text response carries a lighter load than an image; video is far more intensive. Even user behavior matters: removing polite language like "please" and "thank you" shortens responses and reduces consumption by up to 98 gigawatt-hours annually.

The report identifies a fundamental governance gap. Companies and governments lack standardized, publicly available data on AI's resource consumption. Without measurement, there can be no meaningful regulation. The researchers call for transparent reporting systems that would allow comparison across providers and jurisdictions, enabling users to understand the true cost of their choices. They also propose designing more efficient models tailored to specific tasks, improving data center efficiency, and locating infrastructure in regions where environmental impact is lowest. But transparency alone will not suffice. In regions where water stress or grid capacity is severe, direct limits on AI deployment may become necessary.

The scale of the challenge is expanding rapidly. Global spending on AI could exceed 2.5 trillion dollars this year, with the market projected to grow from 189 billion dollars in 2023 to nearly 5 trillion by 2033—a twenty-five-fold increase in a single decade. Video generation, still in its infancy, promises to become a major driver of consumption. By 2030, AI could generate 2.5 million tons of electronic waste annually. The researchers do not argue for halting AI development. Instead, they insist that the technology must be designed, powered, and governed responsibly, with full awareness of its planetary boundaries and with genuine attention to who benefits and who pays. As one researcher put it: the question is not whether AI is powerful or innovative, but who decides, who profits, and who bears the cost.