AI's Growing Thirst Could Outdrink Humanity's Supply

By 2030, servers could be gulping water at a rate that outpaces what every person on Earth needs to drink in a year. That image is not science fiction. It comes from a sober United Nations estimate that ties AI's future to energy grids, water supplies and a hard lesson from economic history.

Call it the Jevons problem: when a technology becomes cheaper and more efficient, humans tend to use more of it, not less. The 19th-century economist who named the effect noticed that better steam engines made coal cheaper to use — and consumption rose. The UN now warns the same dynamic could play out with AI. Efficiency gains in models and data centers may simply unlock new uses, larger-scale deployments and explosive growth in demand.

Numbers help the point land. The report estimates AI could consume up to 3% of the world’s electricity by 2030 — roughly double today's footprint. In 2025 data centers already used as much electricity as Saudi Arabia, itself the world's 11th-largest power consumer. If AI's power draw doubles, the carbon burden would be enormous: offsetting it for ten years would require planting about 6.7 billion trees.

Water is the other hidden cost. Servers run hot. Cooling them requires vast quantities of water, and the UN projects that by 2030 AI-related infrastructure might demand more water for cooling than the global population uses annually for drinking. The report translates these needs into stark figures: data centers could require roughly 9.3 trillion liters of water and land roughly ten times the area of Mexico City. Those are not abstract numbers; they are supplies and places where people live and work.

There is a geopolitical angle, too. Only 32 countries host AI-specific cloud infrastructure today, and nearly 90% of that capacity sits in two nations: the United States and China. The result is a growing digital divide. Countries that build and control large-scale AI systems reap economic benefits while others shoulder environmental costs — from mineral extraction to e-waste — without sharing in the gains.

So what would responsible AI look like? The UN argues for a lifecycle view: governance that follows a model from the mine to the landfill. That means environmental disclosures at both the model and task level, planning AI demand into national climate strategies, and designing systems with efficiency and justice in mind. It's not just about making models faster. It's about designing them to use less, tracking their true footprint and ensuring that the benefits are widely shared.

Policy experiments are already underway. Aotearoa New Zealand has rolled out a national AI strategy and a public service AI framework. Australia’s national AI plan highlights use cases in government: the National Film and Sound Archive’s Bowerbird, a machine-learning engine for mass transcription, and a proof-of-concept tool at the Department of Veterans Affairs that tests whether AI can speed claim processing. Both countries favour a light-touch, principles-based approach — flexible, innovation-friendly, and deliberately hands-off.

That flexibility has a cost. Neither jurisdiction currently mandates environmental disclosures or keeps a national tally of AI energy use or emissions. In other words, the frameworks encourage AI deployment without forcing an accounting for environmental trade-offs. Efficiency improvements alone may not be enough when the Jevons effect and fast adoption push demand higher.

Efficiency without limits can be a mirage: lower per-unit costs often fuel higher total consumption.

Real change will require two simultaneous moves. First, bake environmental transparency into development cycles so every new model ships with a clear footprint estimate — energy, water, minerals and end-of-life plans. Second, fold projected AI demand into broader energy and climate policy so that grids and water systems plan for growth rather than react to crisis.

The stakes are simple. The natural environment underpins economies, cultures and wellbeing. Treating it as an afterthought in the rush to deploy smarter systems risks handing the upstream burdens — mining, water stress, pollution — to communities already most vulnerable.

We can keep inventing astonishing capabilities. We can still reap the social benefits of automation and discovery. But if history teaches us anything, it is that human behaviour responds to incentives. Without careful governance, the incentives around AI will keep turning efficiency into appetite. Planning and transparency are the levers we still have; it’s time to use them.