UN Warns AI's Water Footprint Will Surge 129% by 2030

While the world is fascinated by what Artificial Intelligence (AI) can do, this boom has a thirst problem, and the latest research suggests it is only set to grow. The latest findings from the United Nations University (UNU) warns that the global data centre infrastructure underpinning today's AI systems could consume 945 terawatt-hours of electricity annually by 2030. Put that number in context, and it's worrisome. This means it's nearly triple the combined annual electricity consumption of Pakistan, Bangladesh, and Nigeria three nations that are home to more than 650 million people between them.

The worrying part is that this happens before you even get to the water. Every kilowatt-hour that flows into a server rack carries a hidden cost. Cooling systems and power plants need water. The mining and manufacturing that build AI hardware in the first place chew through land and raw materials. In other words, electricity is just the headline figure, but the water and land footprints tell a much bigger and thirstier story.

Rethinking How Sustainability Is Measured

According to the UNU study, AI's water consumption could climb high enough by the end of the decade to match the basic domestic water needs of roughly 1.3 billion people. Its land footprint, meanwhile, could stretch past 14,500 square kilometres, an area about twice the size of greater Jakarta.

Most of the public conversation around AI's environmental toll has fixated on the energy-hungry process of training large models. But the researchers found that this isn't actually where most of the damage is done. Day-to-day usage, like every single prompt, each image generated, every single chatbot query, makes up somewhere between 80 and 90 per cent of total energy demand.

The scale of that usage is hard to wrap your head around. One popular AI service alone is thought to field around 2.5 billion prompts a day, burning through hundreds of gigawatt-hours of electricity annually just to keep up. Not all AI tasks are created equal, either. Asking a model to classify a snippet of text is relatively cheap. Asking it to generate a single image can cost over a thousand times more energy. Video generation pushes the bill higher still.

There's a temptation to assume better chips and smarter algorithms will solve the problem on their own. The report pours cold water on that idea. It points to a well-documented phenomenon known as the rebound effect: as AI gets cheaper and faster to run, people simply use more of it, and total resource consumption climbs rather than falls.

The pain isn't shared equally, either. AI's benefits are global, but its environmental costs tend to land hardest on specific regions. In some countries, data centres already gobble up a significant slice of the national power grid. In others, new facilities are draining local water supplies, sometimes in places already gripped by drought.

Then there's the waste. The report projects that AI infrastructure could generate up to 2.5 million tonnes of electronic waste every year by 2030, much of which is likely to end up in lower-income countries that lack the infrastructure to dispose of it safely. Add to that the environmental and social fallout from mining the critical minerals needed to build AI chips in the first place, and the picture gets murkier still.

A Widening Digital and Environmental Divide

The most uncomfortable finding in the report is just how lopsided AI's infrastructure buildout has become. More than 90 per cent of the world's specialised AI computing capacity is concentrated in just two countries: the United States and China. Meanwhile, over 150 nations have little to no domestic AI infrastructure of their own. This imbalance goes beyond simple economics. It raises hard questions about environmental justice, namely, whether it's fair for some countries to shoulder the environmental burden of AI's growth while others reap most of the rewards.

Building a Responsible AI Ecosystem

Despite the grim statistics, UNU researchers are careful to frame the report as a wake-up call rather than an indictment of AI itself. Their argument isn't that AI should be abandoned; it's that the technology needs to grow within the planet's actual limits, rather than despite them.

To that end, the study lays out a blueprint for what it calls a "responsible AI ecosystem," built around six core principles: transparency, efficiency by design, equity, lifecycle responsibility, global cooperation, and sustainable use.

The recommendations that follow are aimed at everyone with a stake in AI's future. Governments are being urged to fold AI infrastructure planning into broader energy, water, and land-use strategy, rather than treating data centres as a separate issue. Companies are being pushed to design systems that sip resources instead of guzzling them. While, ordinary users, the report suggests, have a role too, choosing lower-impact tools and applications where they can. According to UNU, engineering improvements alone will not determine AI's sustainability. The choices policymakers, industry leaders and regulators make now will play an equally significant role.