Canada’s rush to host hyperscale AI data centres has exposed a stubborn blind spot: the sector’s thirst for water is large, unevenly reported, and increasingly political — and municipalities that welcomed jobs and investment now face complex technical and governance choices to protect local drinking supplies.
Modern data centres are no longer just quiet server rooms; they are industrial campuses built to host thousands of high‑performance chips that run AI training and inference around the clock. Those chips generate enormous heat, and removing that heat reliably at scale is often most efficiently done with systems that use water — evaporative cooling towers, chilled‑water loops, or heat‑rejection circuits tied into local utilities. The result: cooling demands that can reach millions or even hundreds of millions of litres of water per year for a single large campus under certain designs and operating regimes.
That reality, and the industry’s mixed record on disclosure, has put local governments and residents on a collision course with cloud providers. In Canada, the federal government has actively promoted on‑shore AI infrastructure (including a $2 billion sovereign compute strategy that contains a $700 million AI Compute Challenge to catalyze Canadian data‑centre capacity), which has accelerated land buys and campus proposals. Yet water regulation and metering remain primarily municipal responsibilities — and many municipalities lack the technical capacity or legal levers to demand the exacting, auditable reporting that communities now want.
Municipal actions (recommended, in order):
Canada’s federal programs and the appetite of hyperscalers make the country a likely growth market for AI‑ready campuses. That opportunity can be preserved only by pairing attraction policies with binding, transparent water governance: dedicated metering, enforceable operational limits, audited WUE reporting, and financial instruments that ensure communities are not left footing the bill for scarce water. Municipalities that insist on those safeguards will protect their water, preserve public trust, and keep open the economic benefits without paying their communities’ water bills as the price of progress.
Source: bahiaverdade.com.br AI-related data centres use vast amounts of water. But gauging how much is a murky business - Bahia Verdade
Modern data centres are no longer just quiet server rooms; they are industrial campuses built to host thousands of high‑performance chips that run AI training and inference around the clock. Those chips generate enormous heat, and removing that heat reliably at scale is often most efficiently done with systems that use water — evaporative cooling towers, chilled‑water loops, or heat‑rejection circuits tied into local utilities. The result: cooling demands that can reach millions or even hundreds of millions of litres of water per year for a single large campus under certain designs and operating regimes.That reality, and the industry’s mixed record on disclosure, has put local governments and residents on a collision course with cloud providers. In Canada, the federal government has actively promoted on‑shore AI infrastructure (including a $2 billion sovereign compute strategy that contains a $700 million AI Compute Challenge to catalyze Canadian data‑centre capacity), which has accelerated land buys and campus proposals. Yet water regulation and metering remain primarily municipal responsibilities — and many municipalities lack the technical capacity or legal levers to demand the exacting, auditable reporting that communities now want.
How data‑centre cooling works — and why water matters
The basic physics (quick primer)
Servers and accelerators turn electrical energy into computations and, inevitably, heat. To keep processors within safe operating temperature windows, that heat must be moved off chip into a heat sink and then ultimately rejected to the environment.- Air economization (or free‑air cooling) exchanges outside air with conditioned indoor air when the climate allows; it is water‑free but viable only where ambient temperature and humidity permit.
- Closed‑loop liquid cooling (direct‑to‑chip or cold‑plate systems) circulates dielectric fluids or water in sealed circuits; it reduces energy and can minimize water use but still needs a heat sink system that may use water to reject heat.
- Evaporative cooling / cooling towers use water evaporation to shed heat to the atmosphere; they are inexpensive and effective in hot climates but are the most water‑consumptive option.
Metrics to understand
- PUE (Power Usage Effectiveness) measures energy efficiency, not water use.
- WUE (Water Usage Effectiveness) — liters of water consumed per kWh — is the emerging metric to benchmark water intensity. Voluntary industry pacts are converging on WUE targets, but reporting is patchy and inconsistent.
What the numbers say (and why headlines can mislead)
A handful of recent academic and industry studies have been crucial in translating thermodynamics into everyday headlines — but important caveats apply.- A 2023 technical analysis that gained wide coverage estimated that generating between 10 and 50 medium‑length ChatGPT responses corresponds to roughly 500 millilitres of fresh water when you account for both the water used to produce the electricity and the water used on‑site for cooling; the paper used GPT‑3/GPT‑class examples and provided detailed methodology. That 500 ml figure captured public attention because it is easy to analogize (one small bottle of water), but it is highly sensitive to where the compute runs, the local electricity mix, and the cooling architecture chosen.
- Broader sector tallies published in media accounts and analyst reports have placed annual global on‑site cooling water usage in a range that runs into the low hundreds of billions of litres — numbers that aggregate very different geographies, accounting methods, and what’s included (direct evaporative loss versus indirect water embedded in electricity generation). Some reports attributed a figure of roughly 140 billion litres in 2023 to an International Energy Agency (IEA) compilation; however, the IEA’s public reports emphasize energy trends and electricity‑related water intensity rather than a single unambiguous global water‑for‑cooling number, and a straightforward primary IEA document explicitly showing the 140 billion‑litre line was not readily locatable in public IEA archives during verification — treat that specific headline as reported rather than as a directly traceable IEA table. That uncertainty is important: the range and methodology matter more than a single sound bite.
- Industry averages are poor substitutes for facility plans. A single 100‑megawatt water‑cooled facility can, under continuous water‑intensive operation, consume on the order of a million to two million litres per day; by contrast, a facility that uses free‑air economization most of the year will have drastically lower potable water draws. The variance is enormous and context‑dependent.
Case studies: where the opacity and the politics collide
Nanaimo, B.C.: a local fight with national implications
In Nanaimo, a rezoned wooded lot proposed for a roughly 200,000‑square‑foot data centre has morphed into a flashpoint. Residents, led by local figures such as Kathryn Barnwell, have challenged municipal leaders to weigh the tradeoffs between promised economic gain and potential strain on a drought‑prone municipal supply. Opponents point to precedent abroad where early usage estimates proved optimistic once facilities ramped up, and they cite permit allocations expressed as high volumetric flows that, when annualized, convert to millions or even billions of litres per year. The Nanaimo dispute captures the governance gap: energy interconnection and taxation incentives are debated at provincial and federal levels, while granular water risk is usually decided at the municipal planning table — often with little technical capacity to interrogate complex cooling forecasts.Etobicoke / Toronto: large permitted allocations on planning documents
Planning documents for a Microsoft facility (often discussed under the moniker YTO 40 in local reporting) disclosed an allowed potable draw figure in the planning filings that was reported publicly as 39.75 litres per second — a flow rate that, if sustained continuously, converts to roughly 1.2 billion litres per year. Microsoft counters that its Canadian builds are engineered to rely primarily on free‑air cooling and rainwater capture, and that municipal potable water would be used only under defined extreme conditions. Even if the company’s design intentions are genuine, the permit allocation and the lack of continuous public meter data in some locations have fueled local concern.Varennes, Quebec: metering gaps
An Amazon Web Services campus near Montreal (Varennes) has been cited in reporting as lacking a dedicated potable‑water meter for part of its operational history, which left the municipality with only a flat commercial payment and limited visibility into hourly or seasonal peaks. That absence of metering is emblematic of a governance shortfall: when large industrial water draws are not metered, neither municipalities nor residents can independently verify whether corporate commitments to low potable use are being honored. Municipal officials later indicated a meter would be installed.Noordoostpolder / Netherlands: the cautionary international precedent
Dutch municipal records and investigative reporting revealed that a large campus in North Holland — operated by Microsoft alongside other hyperscalers in the same zone — used tens of millions of litres of potable water in a single year, well above earlier public estimates and municipal briefings. Local media and industry coverage documented figures on the order of 75–84 million litres for a single site during a hot year; Microsoft replied that some of the water figures included construction‑phase consumption and that operational estimates reported earlier referred to conditions at a specific point in time. The episode underlines two recurring features: (1) early, optimistic public figures can omit construction and ramp‑up phases; and (2) real weather variability (heatwaves) can stress designs that depend on air economization or conditional potable draws.Industry responses and engineering options
Operators and vendors are responding with a palette of technical and contractual approaches designed to reduce potable draws and to address community concerns:- Air‑first design: prioritize free‑air economization and only allow water for heat rejection above defined temperature/humidity thresholds.
- Closed‑loop and immersion cooling: seal liquid circuits and minimize evaporative makeup water; these architectures can dramatically cut make‑up water but need reliable non‑potable heat sinks or cooling ponds for full heat rejection.
- Reclaimed and non‑potable water: use treated wastewater, stormwater capture, or industrial water that does not intrude on municipal potable supplies — but this requires local approvals and treatment infrastructure.
- On‑site reuse: capture condensate, reuse process water, and deploy heat‑recovery systems to lower thermal loads.
Governance gaps and a practical action checklist
Local fights show common failures: optimistic early estimates, absent metering, flat‑rate billing that hides volumetric use, and permits that lack enforceable operational modes. Policymakers at municipal, provincial and federal levels can close these gaps with concrete, implementable requirements.Municipal actions (recommended, in order):
- Require a dedicated volumetric potable‑water meter and public reporting (hourly or daily) as a baseline condition of permit approval.
- Make approvals conditional and auditable: air‑first default, potable draw only above verifiable thresholds, and third‑party audits of reported monthly volumes for the first two years of operation.
- Separate construction‑phase water use from steady‑state operational use in permit documents and require reporting of both.
- Negotiate community benefit agreements that internalize the cost of any necessary upgrades to local water infrastructure.
- Tie public incentives and grant money to mandatory, audited WUE reporting and to commitments on non‑potable water sourcing where feasible. Canada’s federal AI sovereign compute strategy already includes grant programs targeted at building on‑shore compute; conditioning those funds on transparency and water‑safety requirements would align public policy with stewardship.
- Provide funding and technical assistance to small municipalities so they can evaluate complex water and cooling models during project reviews.
- Publish facility‑level water metrics (withdrawal, consumptive loss, source type) and make them auditable.
- Disclose contingency plans for heatwaves and droughts, including fallback cooling modes, non‑potable sourcing, and thermal‑load shedding strategies.
Risks and trade‑offs — what to watch for
- Timing and peaks matter: even modest annual volumes can threaten supply if withdrawals coincide with drought peaks or seasonal lows. Municipal engineers care less about annual totals and more about peak day and hourly demand profiles.
- Embedded water via electricity: a non‑water‑intensive on‑site design might still have a large indirect water footprint if the local electricity is generated with water‑intensive thermal plants; comprehensive assessments must include indirect effects.
- Reputational and legal costs: operators that understate or obscure water use risk project cancellations, litigation, and community backlash that can be far costlier than the compliance measures that would have prevented the problem. Examples in Europe and the U.S. show expedited community action can stop or delay projects.
When the numbers don’t add up: verifiable claims and warning signs
Several commonly quoted figures require careful vetting:- The 500‑ml per 10–50 ChatGPT responses figure comes from a peer‑review track preprint and related analyses that explicitly document their methods; it is a useful illustrative estimate but not a universal constant. Use it as a conversation starter, not as a permit condition without location‑specific recalculation.
- The widely reported 140 billion litres global cooling figure has appeared in multiple media summaries attributed to energy‑sector compilations. However, direct primary documentation from the IEA that spells out that exact number in a single public table was not found during verification; municipalities and advocates should therefore avoid treating that headline figure as a precise baseline and should instead rely on facility‑level audits and conservative engineering scenarios. Flag this number as reported but not independently traceable in public IEA tables at the time of verification.
- Company promises that “we’ll only draw municipal water above X°C” are meaningful only if they are codified in a permit, metrically verifiable, and paired with third‑party audits. PR statements without meter‑backed reporting are insufficient. The Netherlands case shows how conditional claims can be complicated by construction phases and unanticipated heatwaves.
A realistic roadmap for towns weighing an AI campus
- Conduct a full‑capacity, worst‑case water‑demand model that separates construction, ramp‑up, and steady‑state operations.
- Demand dedicated metering and integrate meter data into public dashboards.
- Require binding cooling‑mode clauses in zoning/permits (air‑first, potable only above clear thresholds).
- Insist on third‑party annual audits during the first five years of operation.
- Coordinate regionally: neighbours must share data and watershed planning to prevent cumulative impacts.
Conclusion
The AI infrastructure boom is a real economic story: massive capital, construction work, and long‑term compute capacity that governments legitimately want to host. But that boom cannot be decoupled from basic resource stewardship. Water is a public good, and in an era of hotter summers and more frequent droughts, permitting decisions that concede potable supplies to conditional corporate promises without enforceable metering and audits invite conflict and potential harm.Canada’s federal programs and the appetite of hyperscalers make the country a likely growth market for AI‑ready campuses. That opportunity can be preserved only by pairing attraction policies with binding, transparent water governance: dedicated metering, enforceable operational limits, audited WUE reporting, and financial instruments that ensure communities are not left footing the bill for scarce water. Municipalities that insist on those safeguards will protect their water, preserve public trust, and keep open the economic benefits without paying their communities’ water bills as the price of progress.
Source: bahiaverdade.com.br AI-related data centres use vast amounts of water. But gauging how much is a murky business - Bahia Verdade