A new United Nations report is the most complete accounting yet of what powering AI takes from the planet. The numbers are big — but context is everything. On June 3, 2026, the United Nations University Institute for Water, Environment and Health (UNU-INWEH) released a report titled Environmental Cost of Artificial Intelligence: Carbon, Water and Land Footprints . It's not an activist pamphlet or a corporate talking point. It's a peer-reviewed scientific accounting from a UN research institute, and it deserves to be read straight — not cheered, not dismissed. Here's what it actually says, what holds up, and what the scary headlines leave out. What the report found Electricity. Global data centers powering AI are projected to consume 945 terawatt-hours of electricity per year by 2030 — about 3% of the world's electricity, and nearly triple the combined annual power use of Pakistan, Bangladesh, and Nigeria, countries home to more than 650 million people. AI is the engine: it accounts for roughly 20% of data-center energy today, projected to hit 40% by 2030 . Water. Every unit of that electricity carries a "water footprint" — water used both to cool the servers and to generate the power in the first place. The report projects AI's water footprint at 9.3 trillion liters per year by 2030 , equal to the basic annual domestic water needs of 1.3 billion people. Land. The land footprint from data centers, power generation, and supply chains is projected to exceed 14,500 square kilometers — roughly twice the footprint of metro Jakarta. E-waste. As hardware is replaced, AI infrastructure could generate up to 2.5 million metric tons of electronic waste per year by 2030 — the report likens it to discarding 250 Eiffel Towers annually. Concentration. As of 2025, only 32 countries — about 16% of nations — host specialized AI data centers, and 90% of that capacity sits in just two countries: the United States and China. More than 150 countries have none of their own. One finding cuts against the usual story: inference, not training, drives most of AI's energy use. It's not the one-time cost of building a model — it's the billions of everyday queries, image generations, and video clips. A single high-resolution AI video clip can require more than 415 watt-hours, and energy demand rises quadratically as resolution and frame count climb. What's real — and worth taking seriously This report is not the funded, organized opposition that has shadowed individual data-center projects. It's a credible global source, and three of its points deserve honest engagement rather than spin: E-waste and critical minerals. This is the genuinely underdiscussed problem. The minerals AI hardware depends on are often mined in regions with weak environmental oversight, and much of the discarded hardware ends up in lower-income countries with limited safe-disposal systems. That's a real justice issue, not a manufactured one. Local water stress. The report notes that in places like Uruguay and Mexico, plans to build large data centers during droughts triggered public protest over whether household water would be diverted to industry. In a drought, "a small share of national water" can still mean a real local fight. Context-specific stress is legitimate. Concentration of power and benefit. Ninety percent of AI infrastructure in two countries, while the environmental burdens spread globally, is a real imbalance worth naming. What the headlines leave out Big numbers are designed to land hard. A few pieces of context the viral versions skip: Scale, in perspective. 9.3 trillion liters is about 2.5 trillion gallons — globally, per year, projected for 2030. For comparison, U.S. agricultural irrigation alone uses roughly 43 trillion gallons every year (USGS). So even this alarming global AI projection is a fraction of what a single sector in a single country already consumes. The footprint is real; it is not, by itself, civilization-ending. The footprint isn't fixed — it tracks the energy mix. The report is explicit that water and carbon footprints depend on how the electricity is generated. A data center on a clean grid has a dramatically smaller footprint than one on coal or gas. That means the lever isn't "stop building" — it's "build cleaner power alongside it." "For cooling" is incomplete. Most of that 9.3 trillion liters is tied to generating the electricity , not just to chilling the servers. Closed-loop and air-cooled designs are already shrinking the direct-cooling share — which is exactly where efficiency gains are heading. The honest takeaway The UN didn't say AI is an environmental catastrophe, and it didn't say the worries are overblown. It said something more useful: carbon-only accounting is no longer enough, and AI's water, land, and waste footprints need to be measured, disclosed, and planned for — through what it calls a "responsible AI ecosystem" built on transparency, efficiency, and shared accountability. That's a standard worth holding the industry to. The answer to a 945-terawatt-hour future isn't to freeze the buildout — it's to power it cleanly, cool it smartly, and deal honestly with the waste and the inequities on the back end. The footprint is the price of admission to one of the most consequential technologies in history. The job is to shrink it, not to pretend it isn't there. Sources: United Nations University Institute for Water, Environment and Health (UNU-INWEH), "Environmental Cost of Artificial Intelligence: Carbon, Water and Land Footprints," June 3, 2026; UN News; TIME; PBS NewsHour; Earth.org; EurekAlert!; USGS (agricultural water-use comparison). All figures as reported at time of publication.