UN Demands AI Transparency on Hidden Environmental Costs as Data Centers Race Toward 945 TWh

AI companies have spent three years publishing carbon commitments. The United Nations has now published the numbers that make those commitments look, at best, incomplete — and at worst, a distraction from the bigger problem.

A report from the United Nations University Institute for Water, Environment and Health, published June 3 and receiving widespread global attention this week, finds that the environmental cost of artificial intelligence is being “systematically mismeasured” because virtually every existing assessment focuses on carbon emissions from training large models while ignoring the water AI consumes and the land it occupies. The numbers, once all three footprints are counted together, are substantially larger than the carbon-only framing suggests — and the report’s central finding is that switching to renewable energy, the industry’s preferred response to carbon criticism, can cut one footprint while multiplying the other two.

The report calls on governments to require standardized environmental disclosure from AI providers across all three dimensions simultaneously, in ways that allow regulators, investors, and communities to make genuinely comparable assessments. Currently, none of the major AI companies are required to disclose their water or land footprints to the same degree they report carbon emissions, and most choose not to.

What Happened

The 56-page UNU-INWEH report, led by director Kaveh Madani and researcher Miriam Aczel, quantifies the carbon, water, and land footprints of AI’s electricity use across the globe, analyzing data from the 20 largest data center hubs and covering both training and inference workloads. In 2025, global data centers consumed 448 TWh of electricity, making them the world’s 11th largest electricity consumer if treated as a country, with AI accounting for roughly 20 percent of that demand. By 2030, that AI share is projected to reach 40 percent, and total data center consumption is expected to hit 945 TWh — enough energy to rank fifth globally among countries.

On the water and land dimensions, AP’s reporting via PBS noted the report finds that producing that electricity already consumed 4.5 trillion litres of water in 2025; by 2030 that water footprint is projected to equal the basic annual domestic water needs of all 1.3 billion people in Sub-Saharan Africa. The land footprint — from energy infrastructure and supply chains — could exceed 14,500 square kilometres, roughly twice the Jakarta metropolitan area. E-waste adds a separate burden: AI infrastructure is projected to generate up to 2.5 million tonnes of electronic waste annually by 2030.

The Mechanism: Why Low-Carbon Is Not the Same as Low-Impact

The report’s most counterintuitive finding directly challenges the AI industry’s preferred sustainability narrative. Switching a data center from coal-fired to bioenergy power can reduce its carbon footprint by approximately 70 percent. But the same switch increases the water footprint of that electricity more than thirty-fold, and increases the land footprint by a hundred-fold, because growing bioenergy crops requires both substantial land and irrigation. “What surprised us most,” lead author Aczel told reporters, “is how often the choices that look greenest from a carbon perspective end up worse for water or for land.”

The report also argues that improving AI’s energy efficiency per task does not reduce total resource consumption — it increases it, because cheaper and faster AI drives more usage. A 30 percent reduction in words per prompt can reduce energy consumption by 25 percent for that query, but the same efficiency improvement that reduces cost per query also accelerates adoption, resulting in more total queries and a larger aggregate footprint. The report also emphasizes that inference — running AI in response to user queries — now accounts for 80 to 90 percent of total AI energy demand. GPT-3’s training run consumed about 1.3 billion watt-hours; ChatGPT’s daily inference operation is estimated to use 383 gigawatt-hours.

The Backstory

The report is the first United Nations-level analysis to quantify all three environmental footprints of AI’s energy use together. The transparency gap is stark: Google’s data centers consumed 27 billion litres of potable water in 2024, a 28 percent year-on-year increase — a figure disclosed only because Google voluntarily includes it in its environmental reports. No binding disclosure standard currently requires AI companies to publish water or land footprints alongside carbon, which means the data available to regulators, investors, and communities is structurally incomplete. A March 2026 Gallup poll found 71 percent of Americans oppose AI data centers being sited in their area, with opposition centered specifically on water, energy, and quality-of-life concerns.

The scale of data center investment feeding this demand has no precedent. As Nvidia’s sweeping infrastructure deals across South Korea and Asia show, the five largest hyperscalers committed an estimated $660–690 billion in capital expenditure in 2026 alone, approximately 75 percent tied directly to AI infrastructure — the largest single-cycle infrastructure investment in private-sector history. That capital commitment is what underpins the trajectory the UN is now quantifying.

Reactions

Madani framed the report’s purpose as translating physical reality into the terms policymakers and investors actually act on: “AI is not just a virtual thing. We’re talking about something that has physics, something that has real impacts. There is infrastructure there. There is energy that is being used,” he said. “A lot of hardware is behind all these operations that to us seem very, very clean because we don’t see smoke out of our devices.” He also offered a surprisingly practical piece of advice: cutting word use in requests to AI systems by 30 percent can reduce the energy used by that AI interaction by 25 percent — saving roughly the same amount of electricity as what about 700,000 people in Africa use in a year. “If you’re too polite, then that extra ‘please’ you put there can make a huge difference,” Madani said. “You’ve got to be very precise and be short.”

Industry responses were measured. The Data Center Coalition said the industry “takes its environmental impact seriously” and remains committed to working with policymakers. The National Artificial Intelligence Association argued the energy return on investment of AI development is “transformative for our world and therefore more than worth it.” Australian Catholic University associate professor Walayat Hussain offered the more precise pushback: “AI is adding new demand and in some cases accelerating it sharply, but it is not scientifically accurate to place the whole burden of the digital infrastructure problem on AI alone.” Jean Su, director of the Energy Justice Program at the Center for Biological Diversity, described the report as significant because it is the first UN, or even global, report “that shines a light on the environmental harms of AI.”

The Dispute: Disclosure Standards and Who Pays the Environmental Bill

The report’s recommendation for standardized three-dimensional environmental disclosure faces a practical obstacle: there is no agreed international methodology for measuring water and land footprints from electricity use across jurisdictions. The carbon accounting frameworks that underpin emissions reporting took decades to develop and still suffer from well-documented inconsistencies; applying the same rigor to water and land footprints in a way that is genuinely comparable across the US, China, Ireland, and Uruguay simultaneously is a substantially harder governance problem than the report’s recommendations imply.

There is also a distributional justice dimension. As of 2025, only 16 percent of countries host AI-specialized data centers, and 90 percent of that capacity is concentrated in just the US and China, meaning the environmental burden is being concentrated in specific geographies while the economic benefits of AI are distributed globally. In Ireland, data centers accounted for 21 percent of total metered electricity in 2023, a figure that has since grown further even as the country’s grid operator paused new approvals around Dublin until 2028. That dynamic is directly relevant to Ireland’s new AI Office, which begins enforcement work this summer — one of the few regulatory bodies in the world enforcing AI rules precisely in the country carrying some of the heaviest per-capita infrastructure burden the UN report quantifies.

What Happens Next

The report is addressed to five distinct audiences: governments, companies, users, data center operators, and investors — each with separate, concrete recommendations. For governments, the ask is to integrate AI infrastructure into energy, water, and land-use permitting alongside carbon, and to mandate standardized three-footprint disclosure. Whether any of these recommendations translate into binding regulation in the near term will depend on which governments treat the report as a mandate for action rather than a reference document. Meanwhile, as the broader question of whether enterprise AI investment is generating measurable returns remains open, the UN’s findings add a new dimension: the resource cost of the experimentation itself is no longer invisible.

Why It Matters

The UNU-INWEH report’s core contribution is the argument that evaluating AI sustainability through any single metric, whether carbon, water, or land, systematically hides trade-offs and shifts environmental burdens onto the communities least able to push back. For the AI industry, that argument has a direct commercial implication: the transparency and disclosure standards the report calls for would, if implemented, make the full resource cost of running large AI models visible to the investors, regulators, and communities currently making decisions about where data centers get built. The companies best positioned for that future are the ones that start treating footprint as a product design constraint now — not waiting for a regulator to mandate what the Microsoft Maia 200 and similar cloud-native AI chip investments suggest is already becoming economically rational anyway: build the infrastructure to use less energy per useful computation.

Sources

United Nations University Institute for Water, Environment and Health (UNU-INWEH); Associated Press / PBS NewsHour; Time Magazine; ScienceAlert; The Guardian; earth.org.