There is a version of the AI story the industry tells — democratizing knowledge, curing cancer, accelerating science. Then there is the version the UN University Institute for Water, Environment and Health just put in writing, and it does not have a feel-good pitch deck. According to the institute's formal assessment, AI data centers could consume as much water annually as 1.3 billion people by 2030. That is not a metaphor. That is a direct volumetric comparison between the cooling demands of server infrastructure and the survival needs of human beings on the same planet.

The scale of the resource claim is worth sitting with. A single query to a large-scale generative AI model can require roughly half a liter of water to keep the servers that process it from overheating. Multiply that by billions of daily interactions globally, add the exponential growth trajectories that every major AI company is currently chasing, and the compound figure becomes genuinely difficult to minimize. The UN report does not minimize it. It describes a resource trajectory that, if unaddressed, puts AI's expansion in direct competition with agriculture, municipal water systems, and ecological reserves — particularly in regions already under water stress.

Energy is the other half of the equation, and it is no less stark. Data centers running the current generation of AI workloads are among the most electricity-intensive facilities ever built. The International Energy Agency has separately documented that global data center electricity consumption is on track to more than double by the end of the decade. AI inference and training workloads are the primary driver of that acceleration. Several of the largest technology companies have already acknowledged, in their own regulatory filings and sustainability reports, that their energy consumption is rising faster than their renewable procurement — meaning the gap between their green commitments and their actual grid draw is widening, not closing.

What makes the UN assessment particularly pointed is what it says about geography. The energy and water costs of AI are not distributed evenly across the world. Data center expansion is increasingly targeting regions with cheaper land, looser regulatory environments, or more favorable tax structures — which frequently correlates with regions where the local grid is more carbon-intensive and where water scarcity is already a documented stress. Sub-Saharan Africa and parts of South and Southeast Asia are flagged in the broader discussion of AI infrastructure buildout, raising the uncomfortable question of who bears the environmental externalities of a technology whose primary economic benefits accrue to wealthy markets.

Land is the third dimension of the footprint, and the least-discussed. The physical footprint of hyperscale data centers — the cooling infrastructure, power substations, access roads, and exclusion zones they require — consumes significant tracts of land that are permanently converted from other uses. Unlike a factory, a data center does not produce a physical good that can be shipped and used. It produces computation, the benefits of which are intangible and unevenly distributed, while the land conversion is permanent and local.

The technology industry's standard response to environmental critiques is a combination of renewable energy commitments, efficiency improvement pledges, and the argument that AI will ultimately solve climate problems faster than it creates them. That last claim is speculative by definition. The renewable energy commitments, meanwhile, are complicated by the scale and timing mismatch: building new renewable capacity takes years, AI infrastructure is being deployed in months, and in the interim the grid power being consumed is whatever the regional mix happens to be — which in many jurisdictions is still predominantly fossil fuel. Pledging to be net-zero by 2030 while signing power purchase agreements that stress coal-heavy grids today is a form of temporal accounting that benefits the company's marketing while the atmosphere experiences the emissions in real time.

One detail buried in the broader discussion of AI energy consumption deserves more daylight: the marginal cost of politeness. Researchers examining prompt structure found that AI models consume meaningfully more energy when users include social courtesies — please, thank you, could you — because the additional tokens extend the computational sequence. This is a minor data point in isolation, but it is illustrative of something larger: the resource cost of AI is embedded in behaviors so normalized that users never think of them as resource consumption at all. Every interaction is an infrastructure event. The cumulative effect of billions of infrastructure events daily is what the UN just described.

What the report asks of governments is straightforward in principle and politically difficult in practice: mandate environmental disclosure from AI companies, require data centers to account for water consumption in water-stressed regions, and integrate AI infrastructure into national energy and climate planning rather than treating it as a tech-sector matter outside the scope of environmental regulation. Canada, where the UN institute is based, is specifically noted as a country whose AI strategy contains no binding new protections for water or climate tied to AI infrastructure expansion — a gap that is not unique to Canada but is pointed given the host nation context. The warning is formal, the evidence is documented, and the window for getting ahead of this is narrowing at roughly the same speed as the AI investment cycle is accelerating.