"Data centre" does not
mean one design.
Almost every argument about data-centre water is really an argument about cooling architecture. Change the cooling design and the water answer changes completely — while the electricity answer moves in the opposite direction.
A giant AI data centre must need huge cooling towers consuming enormous amounts of local water. Right?
Answer before you scroll. The honest answer is the third one.
Compare the five architectures
- Water
- None None consumed in normal operation — the loop is sealed and refilled rarely.
- Electricity
- Medium Higher than evaporative. Fans replace evaporation, so you pay in electricity.
- Rack density suitability
- High when paired with direct liquid cooling.
- Climate sensitivity
- Strongly favoured by cold climates. Alberta’s winters are a genuine engineering advantage here.
Disclosed in AlbertaMeta states the Sturgeon design uses dry cooling with no operational water use for cooling.
Notice the pattern: the designs that use the least water use the most electricity, and the one that uses the least electricity consumes the most water. There is no free option — only a choice about which resource you would rather spend.
So does a dry-cooled campus have no water footprint?
No.
Meta states the Sturgeon design uses closed-loop liquid cooling with dry heat rejection and no operational water use for cooling. That is a real and meaningful design choice. It is also not the whole water story.
- Inside the fence Domestic and site water Taps, washrooms, kitchens.
- Inside the fence Fire protection Suppression systems charged and tested.
- Inside the fence Construction Concrete, dust suppression, earthworks.
- Outside the fence Electricity generation Thermal plants use water for their own cycles.
- Outside the fence Semiconductor manufacturing Chip fabs are extremely water-intensive.
- Outside the fence Upstream supply chain Steel, copper, concrete, everything else.
Three of six sit outside the site boundary, in other industries and often other countries. That is systems thinking — and it is far more useful than any single water-per-query statistic.