As demand for generative artificial intelligence continues to surge, the physical infrastructure that powers it is placing increasingly heavy demands on natural resources. A recent analysis by Canadian Geographic indicates that a typical 100-megawatt data centre — the kind used for large-scale AI workloads — uses around two million litres of water per day to cool its servers.

Water enters the picture primarily because AI processing generates substantial amounts of heat. Servers convert electrical energy into heat as they perform computations; to maintain operational temperatures, cooling systems are required. Many use open-loop cooling towers that rely on constant evaporation of water. Some of the water draw is indirect — for instance, via electricity generation, which itself may require water.

In addition, the researchers behind the coverage calculated that for a moderate number of AI model responses — between 10 and 50 medium-length replies generated by GPT-style models — the water consumption could equate to roughly 500 millilitres of water. While that per-interaction figure may seem small in isolation, scaling up to millions of queries magnifies the impact across many data centres.

In Canada there are currently 239 operational data centres. With expanding investment and implementation in AI infrastructure, that number is projected to grow. For example, a telecommunications provider is planning multiple new AI-oriented centres in British Columbia, and some regions are positioning themselves as preferred hosts for AI-enabled investment.

A critical caveat is that 80 to 90 per cent of the water used in many data-centre cooling systems is classified as potable water. Yet many of these facilities are located in regions already experiencing water stress. For instance, an industry investigation in 2025 found that most data-centre sites are in areas of water scarcity.

Among strategies for mitigation, shifting to recycled or non-potable water sources, designing closed-loop or air-cooled systems and placing data-centre facilities in cooler climates with lower evaporation rates are under discussion. Meanwhile, the environmental footprint of AI becomes not just a matter of electricity consumption and carbon emissions, but also one of major water budgets.

In sum, while a simple query to an AI-model may seem trivial, it sits atop a complex network of infrastructure that consumes large volumes of energy and water. Recognizing the hidden resource cost behind digital convenience underscores the need for more sustainable planning of the AI ecosystem.