Data Centres and Water: The Disclosure Black Hole

The input nobody measures

Power at least gets counted. Grid operators publish queues, developers file megawatt figures, regulators rule on gas plants in the open. Water is different. Across our dataset of Canadian AI data centres, not one publishes a measured, audited water-intensity figure, no litres per megawatt-hour, no annual withdrawal, no consumption number you could check against a meter. What exists instead is a scatter of licence ceilings, planning estimates, and relative marketing claims, each of which needs its own caveat.

The pattern repeats project to project. QScale’s Q01 campus in Quebec markets direct liquid cooling and up to roughly 80 percent free-air cooling, but publishes no litres per day and no water-use efficiency figure. TELUS describes its B.C. sovereign AI cluster as sealed closed-loop with “up to 90 percent” water reduction, a relative claim with no absolute litres and no stated baseline. The Beacon and VoltaGrid proponents in Saint John told the community they simply could not state their closed loop’s water volume, and deferred the figure to a future environmental impact assessment. Disclosure this thin is not an accident of a young industry. It is the norm.

Two numbers, and what they really are

Only two Canadian projects give you anything project-specific to hold, and both come with heavy asterisks about what kind of number they are.

• Wonder Valley (O’Leary Ventures, near Grande Prairie) holds a water licence for up to 6 million cubic metres a year from the Smoky River, plus a preliminary certificate for up to 24 million cubic metres a year. Converted to a daily rate, the 6 million cubic metre licence works out to about 16.4 million litres a day, and the 24 million cubic metre certificate to about 65.8 million litres a day. Those are annual allocation ceilings divided by 365, not metered consumption. The company frames the withdrawal as roughly 0.0556 percent of the Smoky River’s annual flow and under 1 percent of daily flow, restrictable in low-flow periods. Sturgeon Lake Cree Nation appealed the licence; the appeal was dismissed on April 17, 2026.
• Beacon AI Indus (near Calgary) describes “closed-loop glycol cooling with minimal process-water use” in its filings. Yet those same regulatory filings, surfaced by Canada’s National Observer, put the campus draw at up to about 1.5 million litres a day, plus roughly 750,000 litres for steam generators and fire safety. That is a filing estimate, not a meter reading, and it undercuts the “minimal water” framing on the project’s own paperwork.

Notice how much work the units are doing. A licence cap in cubic metres per year, a filing draw in litres per day, a benchmark in litres per megawatt-hour, and a U.S. comparator in millions of gallons per day are four different kinds of number. Reporting that flattens them into one headline figure is the fastest way to be wrong about this file.

Cooling method drives the order of magnitude

The reason water use is so hard to pin down is that it is set almost entirely by a design choice that projects rarely detail: how the heat gets rejected. The spread between methods is not a rounding difference. It is roughly an order of magnitude.

• Evaporative or open-loop cooling, the traditional workhorse, runs around 1,800 to 2,900 litres per megawatt-hour on benchmark figures, and up to roughly 4,000 in hot, dry climates. It loses water to the atmosphere by design.
• Closed-loop and direct-to-chip systems recirculate a fixed charge of coolant. Hyperscale water-use efficiency across climate zones spans roughly 90 to 1,800 litres per megawatt-hour, and the newest builds push toward near-zero ongoing draw, fill plus makeup only.
• Air-cooled and dry systems use almost no direct water, but shift the cost to electricity through a higher power-usage effectiveness.

For scale, vendor fleet averages sit at the low end: Microsoft reports a fleet water-use efficiency near 0.30 litres per kilowatt-hour, about 300 litres per megawatt-hour, down from 0.49 in 2021, with new builds designed for zero water. A secondary estimate puts AWS near 0.19 litres per kilowatt-hour, about 190 litres per megawatt-hour. These are vendor and benchmark numbers, self-reported and not Canada-specific, and they are useful only as reference points. None of them is a measured value from a Canadian AI project, because no such published value exists.

“Closed-loop” is not the same as zero-water

The most common reassurance from proponents is that their cooling is closed-loop, as if that settled the question. At small scale it nearly does. At hyperscale it does not, and the evidence for that is worth stating plainly.

A closed-loop Meta campus in Indiana is estimated, in an academic preprint, to still require about 30.3 million litres a day (8 million gallons) of makeup water for peak heat rejection at full buildout. It is a U.S. comparator from a lower-credibility source, so treat the exact figure as indicative. But the direction is the point: a system marketed as closed-loop still consumes tens of millions of litres a day when it is large enough. Alberta’s own Beacon AI Indus filings make the same point domestically, with a closed-loop design that still draws around 1.5 million litres a day. And in New Brunswick, Maine legislators reviewing a related VoltaGrid project flagged that even “closed-loop” systems can produce unaddressed wastewater discharge. Closed-loop reduces water use. It does not zero it out, and it does not remove the discharge question.

Three-quarters of Alberta’s sites are in stressed basins

Disclosure would matter less if the projects sat where water is plentiful. Many do not. Canada’s National Observer analyzed 38 proposed Alberta data-centre campuses against the WRI Aqueduct 4.0 stress index and AESO connection data, and the result is the sharpest single finding in this chapter.

• Roughly three-quarters of the 38 campuses sit in basins rated high stress (at least 40 percent of available supply already used) or extremely high stress (at least 80 percent). Only 3 of the 38 sit in low-stress areas.
• The Bow and Oldman basins have been closed to new surface-water licences since 2006 because they are over-allocated. Yet 13 data centres totalling 5,744 megawatts are proposed within them, and 4 already have grid approval.
• Under Alberta’s “first in time, first in right” system, developers can buy or inherit senior water rights to get around a closed basin, which turns water allocation into a market rather than a public planning decision.

The province has also widened its own room to maneuver. Alberta’s 2025 Water Amendment Act lets the government approve “low-risk” inter-basin water transfers by ministerial order, moving water between watersheds administratively rather than through the older, slower licensing path. Combined with over-allocated basins and near-total non-disclosure, that is why water, not power, is often the flashpoint at the local level, from Wonder Valley to the Saint John opposition to residents in Olds who openly questioned Synapse’s closed-loop cooling claims.

The counter-view, for balance

It would be one-sided to leave it there. Not everyone thinks data-centre water use is a crisis in the Canadian context, and the strongest counter-argument deserves a fair hearing.

The Canadian Centre for Policy Alternatives has argued, through analyst Hadrian Mertins-Kirkwood, that broad data-centre water concern is overblown for a water-rich country like Canada, while explicitly flagging water-stressed basins as the exception that does warrant scrutiny. That is a defensible position, and the numbers partly support it. Nationally, Canada is not Arizona. For a sense of the U.S. baseline, the Lawrence Berkeley National Laboratory estimated American data centres used about 17 billion gallons (roughly 64 billion litres) of direct cooling water in 2023, projected to double by 2028, and Ceres found 32 percent of U.S. data centres sit in high or extremely high stress areas. Canada’s overall endowment is far more comfortable than that.

But national abundance is not the metric that decides whether a specific campus stresses a specific river. The honest synthesis is a split one. In aggregate, Canada has water to spare. In the exact basins where Alberta developers most want to build, it does not, and the absence of any measured, published water figure means the public is being asked to trust the reassurances without the meter. Both things are true at once, and the CCPA framing captures the first without dissolving the second.

What we can and cannot say

Strip away the marketing and the fair conclusion is narrow but firm.

1. No Canadian AI data centre publishes a measured water-intensity figure. Every project-level number in circulation is a licence cap, a filing estimate, or a relative claim, and should be labelled as such.
2. Cooling method sets water use to within an order of magnitude, and “closed-loop” at hyperscale is not zero-water. The reassuring word is doing more work than the physics allows.
3. Roughly three-quarters of planned Alberta sites sit in high or extreme water-stress basins, several of which have been closed to new surface-water licences for nearly two decades.
4. Nationally the water endowment is genuinely large, which is why the concern is best framed as basin-specific, not country-wide.

The fix is not exotic. It is the meter reading itself: mandatory, audited, per-facility disclosure of withdrawal and consumption, reported in consistent units and separated from licence ceilings. Until that exists, water will stay what it is today, the least-disclosed input in the buildout and, not coincidentally, the one that keeps ending up in front of a tribunal.

Frequently asked questions

Sources

Primary and reputable secondary sources: Canada’s National Observer (the 38-campus Alberta water-stress investigation, using WRI Aqueduct 4.0 and AESO data, and the Wonder Valley licence reporting); CBC News and The Narwhal (Wonder Valley, Saint John, and Olds opposition); Data Center Knowledge (Alberta water challenges); the WRI Aqueduct water-risk index; Microsoft and reporting on AWS (vendor fleet water-use efficiency benchmarks); an arXiv academic preprint (the Meta Indiana closed-loop makeup-water estimate); the Lawrence Berkeley National Laboratory and Ceres, via The Conversation (U.S. aggregate water use and stress exposure); and the Canadian Centre for Policy Alternatives (the counter-view on Canadian water risk). Licence caps, filing draws, and benchmark or vendor averages are labelled distinctly throughout; none represents a measured value from a Canadian AI data centre.