Chapter 6: Infrastructure Reality

When Prime Minister Carney's AI for All strategy committed to "Powering AI adoption" as one of its six pillars, the language was abstract: compute…

When Prime Minister Carney's AI for All strategy committed to "Powering AI adoption" as one of its six pillars, the language was abstract: compute, capacity, sovereign foundations, Canadian leadership. This chapter goes inside the abstraction. A data centre is a physical building, sited in a specific place, drawing real electricity from the grid, using water (or air) to cool silicon, owned by a corporation that operates under the law of a particular jurisdiction. Until you've worked through those specifics, the policy framing tells you almost nothing about what's actually being built in Canada or by whom. Canada has, as of late 2025, over 300 existing data centres with another 60-plus in active development. Roughly half of all current capacity is concentrated in three city-regions, whose data-centre electrical demand Natural Resources Canada's Best Practice Guide for Canadian Data Centres estimated at about 320 megawatts for Toronto, 520 megawatts for Montréal (already larger than Toronto), and 46 megawatts for Vancouver, a much smaller cluster. Calgary is the fastest-growing emerging hub, driven mostly by Amazon Web Services' second Canadian region and by Alberta's combination of cheap land, cheap electricity, and provincial policy favourable to large-scale data-centre development. One caveat worth holding, because it doubles as a lesson in reading numbers. Ask a commercial data-centre analyst which Canadian city is biggest and you'll often hear Toronto, not Montréal — firms like CBRE rank Toronto's colocation supply (the third-party space tenants rent rather than build for themselves) well ahead. NRCan's estimate runs the other way because it counts total electrical demand, including the hyperscale and cryptocurrency-mining loads that parked themselves on cheap Hydro-Québec power and never showed up in anyone's colocation tally. Both numbers are real; they are counting different things. When a source tells you which city "leads," the first question is which load it decided to count. Federal data-centre electricity consumption grew from 1.4% of total Canadian electricity in 2021 to 2.1% in 2023, a 50% increase in two years, before the most recent wave of generative-AI demand had fully landed. The International Energy Agency projects that global data-centre electricity will roughly double from about 415 terawatt-hours in 2024 to more than 940 TWh by 2030, surpassing the entire national power use of Canada by the late 2020s. The Canadian-specific trajectory inside that global projection is what this chapter is about. We’ll look inside a data centre, then put the Canadian facilities we can identify on the map. Wonder Valley brings the questions about power, water, work and the duty to consult to the same patch of ground. ---

What's actually in there

Lire cette section en français

View receipt #39

A data centre is a building optimized for one purpose: to keep large quantities of computing hardware powered, cooled, and connected to the internet, twenty-four hours a day, three hundred and sixty-five days a year, with the highest possible reliability.

Walk inside one and you'll find rows of metal racks, each rack holding stacks of servers — flat horizontal computers, looking nothing like a desktop or laptop, designed to maximize density. Each server has CPUs (central processing units, the general-purpose computing chips) and, increasingly in AI-focused facilities, GPUs (graphics processing units, originally designed for video games, now the central compute substrate for almost all serious machine learning work). NVIDIA's H100 and H200 GPUs are the chips that most large language models, including those running at Cohere, OpenAI, and Anthropic, are trained and operated on. NVIDIA shipped an estimated 3.76 million data-centre GPUs in 2023, roughly a million more than the year before, per analyst firm TechInsights — despite simultaneous efficiency improvements, and shipments have kept climbing since (the same firm pegs 2024 at almost four million). The chips are physical objects, mostly made in Taiwan (TSMC), with their critical materials (rare earths, gallium, germanium, cobalt for batteries, copper, tungsten) sourced globally. A majority of the extraction projects for such minerals are located on or near Indigenous and peasant lands, a finding from Owen et al. (2023) that we introduced in Chapter 1 and will return to in Chapter 9.

The servers run continuously. Continuous operation means continuous electrical demand. A modern hyperscale AI data centre — the kind being built by Cohere, by Microsoft for its Azure platform, by Amazon for AWS — requires 100 megawatts or more of electrical capacity. That is ten to twenty times the demand of a traditional enterprise data centre at 5–10 MW. The Canadian Union of Public Employees, in its March 2026 Senate submission, used a concrete frame: the annual electricity demand of one major AI data centre is roughly equivalent to the electrical demand of 350,000 electric vehicles.

Almost all the electricity used by the servers ends up as heat. A hall running 100 MW of IT equipment needs to remove roughly that much heat from the equipment while it runs. Cooling, power conversion and other building systems add to the electricity drawn at the site boundary.

There are three broad approaches to that removal, and which one a data centre uses determines most of its environmental footprint. Free-air cooling uses outside air, blown through the building, to carry heat away. Canada's cold climate is a structural advantage here. Natural Resources Canada's own technical guide notes that "in cooler and dryer climates such as in many parts of Canada," free-air cooling can be used for substantial portions of the year, sharply reducing energy overhead. Evaporative cooling uses water, which absorbs heat efficiently but consumes significant volumes: for AI-scale facilities, often hundreds of thousands of litres per day per facility, sometimes millions during summer peaks. Mechanical refrigeration (chillers, similar in principle to large air conditioners) uses no water but consumes substantial additional electricity, sometimes increasing total energy use by 30–50% over the computing load itself.

Most large Canadian and US facilities use some blend of all three depending on weather, with evaporative cooling dominant during warm months and free-air during cold months. The choice between water and electricity is not abstract. A facility built in a drought-prone zone (as more and more US and Mexican facilities are) faces real water-supply pressure, while one in a cold Canadian climate trades that water consumption for the simpler logistics of moving outside air through the building. Either way the environmental cost is real; it just lands differently, and on different communities.

The headline metric the industry uses to describe data-centre efficiency is Power Usage Effectiveness (PUE), the ratio of total facility energy use to the energy used by the computing equipment itself. A perfect PUE of 1.0 would mean every watt entering the building powers a chip. Real-world facilities range from about 1.05 (Google's industry-leading global fleet average is 1.09; Microsoft Azure and Meta are in similar territory) to 2.5 or worse for older facilities. The global average is around 1.55–1.58. Lower is better; the difference between 1.1 and 1.5 represents enormous quantities of electricity going to cooling, lighting, and overhead rather than to computing.

Start with who was counted. Natural Resources Canada’s 2024 guide cites a Canadian market assessment that put PUE reporting at about 22%. The guide does not explain who received those reports or show how to reproduce the count. In Uptime Institute’s 2023 global survey, 71% of 716 respondents to a multiple-choice question said they compile and report PUE for corporate sustainability purposes. That is a share of survey respondents. The two figures cannot give us a ratio of public disclosure.

🧌 GOBLIN CHECK — Before you compare two percentages, ask who was counted and who can see the numbers. A facility can measure its efficiency, send it to head office and never publish it. If the survey doesn’t tell you which happened, the goblin leaves that box blank.

A second metric: Water Usage Effectiveness (WUE), litres of water consumed per kilowatt-hour of IT energy delivered. The US Department of Energy's average baseline is about 1.8 litres per kWh. Specific Canadian disclosures are even sparser than PUE disclosures. Google has reported a per-prompt figure of 0.26 millilitres of water (about five drops) for the median Gemini text query, a number from Google's August 2025 inference paper, with the caveats we worked through in Chapter 1: corporate self-disclosure, unverified by independent third party, scope limited to median text prompts (excluding image and video generation, where consumption is far higher).

Here's a working example. Start with 100 MW of IT load running continuously. With a power usage effectiveness (PUE) of 1.15–1.25, the whole site draws 115–125 MW and uses about 1.01–1.10 TWh per year. If 100 MW is already the site's total draw, continuous operation uses 0.876 TWh per year. A connection rating still needs a utilization assumption. For the same continuous IT load, our illustrative on-site water scenarios use near zero to 0.18 million m³ per year for predominantly dry heat rejection, or 1.6 to 2.5 million m³ per year for evaporative cooling. These are scenario inputs. Weather, heat rejection design, operating hours and the water source determine how a particular site compares.

GOBLIN FACTS — megawatts are policy. A 100 MW campus is not just a building. It is a large industrial load that competes for grid capacity, water planning, and local land-use decisions.

Before comparing a proposed campus with a city's existing demand, ask whether both numbers measure the same thing. IT load, whole-site draw and grid capacity reserved for a project can all appear in megawatts. A grid planner still needs to know which projects will run, where they connect and when their loads overlap.

That's the technical layer. Now to who owns what.


The hyperscaler footprint in Canada, mapped

Lire cette section en français

View receipt #116

Bell’s initial 300 MW build at Sherwood, near Regina, shows what construction looks like on the ground. Saskatchewan reported on October 1, 2026 that all four buildings were taking shape. BCE reported about $475 million in capital expenditures and associated cash outflows as of September 30 and expects the first data hall in the first half of 2027. A separate, non-binding agreement proposes up to 900 MW more.

When AI for All and Minister Solomon talk about "sovereign AI infrastructure," the implicit map most listeners have is something like: Canadian companies operating Canadian data centres on Canadian land, governed by Canadian law. The actual map is more complicated.

Amazon Web Services entered Canada in 2016 with its first cloud region, AWS Canada (Central), in Montréal. AWS's own self-published economic-impact study (November 2021) claims that direct investment of $1.4 billion over the first five years supported approximately $1.1 billion in gross value added and 687 full-time equivalent jobs in Canada. The second Canadian region, AWS Canada West, in Calgary, was announced in 2021 and expected to launch in late 2023 or early 2024. Future investment, per the same AWS study: $21 billion (CA$21B) by 2037, projected to support 5,195 full-time jobs and contribute $39 billion to GDP. (A separate AWS-commissioned report by the research firm Public First frames the macro picture: $8.5 billion in 2021 economic value, $40B-plus GDP potential by 2030; Chapter 7 returns to that firm.) Edge locations operate in Toronto, Montréal, and Vancouver. Total Amazon employment in Canada is around 39,500 people, including approximately 2,800 AWS employees.

Microsoft Azure opened two Canadian regions in 2016, in Toronto and Quebec City, both contemporaneous with AWS's Montréal launch. Microsoft does not publish equivalent commissioned-research investment-impact reports for its Canadian footprint, but the regions are comparable to AWS's in scale and customer base. Microsoft has committed roughly $19 billion to Canadian cloud and AI infrastructure over 2023–2027, and in April 2026 Ontario welcomed an Azure Canada Central expansion adding data centres in Markham, Vaughan, and Toronto, billed at about 1,250 jobs.

Google Cloud operates two Canadian cloud regions — Montréal (older) and Toronto (newer). Like Microsoft, Google has not published a comparable commissioned-research impact report on its Canadian footprint.

Oracle, IBM, and Tencent all operate or are actively expanding Canadian cloud regions, with smaller footprints than the big three but real presence. IBM emphasizes Toronto and Montréal multizone region capability specifically positioned for AI and data-sovereignty workloads.

The colocation operator layer, distinct from the hyperscaler cloud regions, is where much of the actual physical capacity sits. Three operators dominate Canadian colocation: eStruxture positions itself as the largest Canadian colocation provider, with facilities in Montréal, Toronto, Calgary, and Vancouver. Cologix operates 22 Canadian sites totaling 94 MW and over 1 million square feet, with 12 facilities in Montréal alone. Vantage Data Centers operates campuses including 89 MW combined at Montréal III and 86 MW at Québec City. QScale's Q01 campus in Lévis is marketed at 142 MW with a 60 MW expansion phase that began construction in June 2026, making it one of the largest AI-oriented data centre projects in Canada. These are the physical layer that the hyperscaler regions and Canadian-controlled "sovereign cloud" offerings actually run on top of.

Put together: every major foreign hyperscaler operates Canadian regions, with Montréal (because of its hydropower-dominated grid and cold climate) and Toronto (because of its financial-services proximity) as the primary nodes. The geographical concentration in Quebec and Ontario is not accidental: those are the provinces with the lowest electricity costs and the most favourable grid characteristics for high-load data-centre operation. Alberta is the rising third node, primarily because of Calgary's combination of cheap electricity (until recently dominantly natural gas, increasingly with renewable additions) and provincial policy incentives.

Now to the Canadian-controlled side. The thread to hold through the names and dollar figures below is one question: who owns the building, who owns the chips, and who owns the model. Cohere, founded in Toronto in 2019 and headquartered there, is the most-named Canadian AI company in AI for All, the strategy's principal "national champion" example. In December 2024, the federal government committed $240 million to support Cohere's Canadian compute build-out, the first investment under the $2-billion Canadian Sovereign AI Compute Strategy. That money backs a Cambridge, Ontario data centre built and operated by the US cloud provider CoreWeave, with Cohere as anchor tenant, a sovereignty caveat worth naming. Separately, Bell announced its AI Fabric project in May 2025, a national network starting with a plan for upwards of 500 MW across six B.C. facilities, and in July 2025 Bell and Cohere partnered: Bell as Cohere's preferred Canadian infrastructure provider, Cohere as AI Fabric's preferred model provider. Cohere itself develops foundation models (its Command series competes with OpenAI's GPT, Anthropic's Claude, and Google's Gemini, primarily on enterprise rather than consumer markets). The Bell AI Fabric represents the most concrete operational expression of the "sovereign AI infrastructure" pillar: a Canadian telecommunications incumbent building infrastructure intended to host Canadian AI workloads in Canadian jurisdiction, with Canada's flagship model developer signed on as partner. That partnership moved from announcement to operation in June 2026, when Bell, Cohere, Hypertec, and BUZZ HPC unveiled a deal to run Cohere's models on Bell's new Merritt, British Columbia data centre (roughly 6.5 MW, slated to roughly double by early 2027), on NVIDIA DGX servers assembled in Canada by Hypertec and financed through a reported US$220-million GPU contract. It is the most concrete sovereign-AI buildout yet, and it also marks exactly where the sovereignty stops: the Canadian data centre and the Canadian-assembled racks still run on NVIDIA chips designed in the United States and fabricated in Taiwan. Canadian land, Canadian building, Canadian assembly, foreign silicon. The deal is the companies' own announcement, so read the "sovereign" framing as their claim rather than an audit.

Telus and Bell are the two Canadian telecommunications incumbents most active in the "sovereign cloud" framing, both marketing data-centre and cloud services to enterprise customers (especially federal and provincial governments) on the basis that Canadian-controlled infrastructure is safer for sensitive data than US-controlled hyperscaler infrastructure. Both companies' environmental and sovereignty claims are corporate self-disclosure and should be read accordingly. The guide flagged this in its bias methodology, and we'll return to it in Chapter 7 when we look at who actually holds AI power in Canada.

The asymmetry the chapter has to be honest about. Foreign hyperscalers already operate Canadian regions, and Canadian-controlled operators have facilities and projects at different stages. The inventory here gives examples. To put a national share on either group, we'd need operating megawatts measured on the same basis, with announcements and builds counted separately and with an explicit definition of control. For a buyer, the useful next question is who actually operates the capacity on offer today.

The political case for all this got a vivid illustration in June 2026. After a US export-control directive forced Anthropic to suspend its newest models (Fable 5 and Mythos 5) for foreign nationals, and the company disabled them for everyone, Prime Minister Carney pointed to the episode as proof of the risk of leaning too heavily on American AI providers. It was the rare moment of a sitting prime minister making the guide's own argument out loud: that "sovereign AI" is not branding but a hedge against exactly this kind of dependence, where a foreign government's decision can reach through a foreign vendor and switch off tools Canadians rely on. The catch is the one this section already named. The compute Canadians actually use remains overwhelmingly foreign, so the dependence Carney warned about is, for now, the operating condition rather than a risk on the horizon.


The Wonder Valley case study

Lire cette section en français

In December 2024, O'Leary Ventures (yes, that Kevin O'Leary) announced, in cooperation with the Municipal District of Greenview, plans for Wonder Valley: billed as the world's largest AI data-centre industrial park, a proposed multi-gigawatt campus in the Greenview Industrial Gateway near Grande Prairie, in Treaty 8 territory in northwestern Alberta, drawing electricity primarily from dedicated natural-gas generation rather than from the provincial grid.

What happened next is the actual case study. Sturgeon Lake Cree Nation, whose Treaty 8 territory the project sits within, learned about Wonder Valley the way you probably did: from the news. On January 13, 2025, Chief Sheldon Sunshine issued an open letter to Premier Danielle Smith and O'Leary demanding the project halt until consultation occurred. The non-consultation was reported through January 2025; The Narwhal connected the project to the broader AI-infrastructure buildout that fall. In February 2026, O'Leary Digital announced a joint venture to advance a 7.5 GW "WonderValley" campus in Utah while Alberta permitting continued. And in June 2026, the same week AI for All launched, Sturgeon Lake Cree Nation was in court, arguing that the Crown's duty to consult had not been met. A national AI strategy launched its "Sovereign Foundations" pillar in the same news cycle as a First Nation litigating whether the most basic constitutional consultation obligation had been honoured on the country's highest-profile AI infrastructure project.

The structure of the project is worth understanding, because it puts almost every conversation in this guide on top of one site.

On environment, Wonder Valley represents what's sometimes called a "behind-the-meter" or "bring-your-own-grid" architecture: dedicated natural-gas generation, sized to the data centre's load, not connected to the broader Alberta grid in the way a traditional data centre would be. The advantage from a permitting and reliability perspective is real: the facility doesn't have to wait for grid expansion or compete with other electrical users. The carbon implication is also real: natural-gas combustion produces roughly 400–500 grams of CO₂-equivalent per kilowatt-hour, compared to roughly 30 grams for Quebec or Manitoba hydropower and roughly 800–1000 grams for coal. A gigawatt-scale data centre on dedicated natural gas is, by carbon arithmetic, on the order of millions of tonnes of CO₂ annually. Independent assessment of Wonder Valley's specific carbon footprint has been contested between project proponents (using lower-end estimates and emphasizing offset commitments) and environmental advocates (using higher-end estimates and questioning the offset accounting).

EXAMPLE — the generator in the backyard. "Behind-the-meter" power sounds like a technicality. It just means a data centre builds its own power plant on site instead of plugging into the public grid — the industrial version of running a generator out back so you never have to ask the utility. Faster to build, and much harder for anyone outside to see.

On consultation and Indigenous rights, the layer an earlier draft of this guide got wrong, and the correction is worth making in public. Early framings of Wonder Valley (including this guide's own working draft) described Sturgeon Lake Cree Nation as an equity participant with governance roles. The documented record says otherwise: the Nation's open letter states it was not consulted before the announcement; subsequent reporting confirmed the Nation learned of the project through media; and as of June 2026 the Nation is in court over the duty to consult. That duty is constitutional, grounded in section 35 and developed through the Haida Nation line of cases, and it is triggered by Crown conduct that may adversely affect treaty rights. Whether Alberta met it here is for the court to decide. That the project was announced without prior consultation is, on the public record, not seriously contested.

🧌 GOBLIN CHECK — a correction, filed in public. An earlier draft of this chapter called Sturgeon Lake Cree Nation a Wonder Valley equity partner. Then the receipts came in: an open letter demanding a halt, news reports of zero consultation, a courtroom. When the narrative and the documents disagree, the narrative loses — that's the whole rule of this book, and it applies hardest to the book itself. The corrected record stays. So does this note, as the worked example.

On jobs, Wonder Valley's stated employment projections are in the range of hundreds of operational positions plus several thousand construction positions over the build period. Independent labour-economic assessment of those numbers (how many are local, how many are skilled trades from outside the region, what wages, what duration) has been less prominent in coverage than the headline figures. The Canadian Union of Public Employees' general critique of AI infrastructure investments (that public funding shouldn't be allowed to support facilities that cut jobs elsewhere) applies obliquely; Wonder Valley is largely private capital with federal alignment rather than direct public funding.

On Indigenous data sovereignty, the sequence matters. OCAP, NISR, and CARE are frameworks for governing data, but data governance presupposes a relationship, and the documented record here shows the relationship was skipped at layer zero: the land itself. A project that did not consult before claiming territory for the world's largest AI campus was never going to engage Indigenous data governance over what the facility processes. Chapter 9 returns to Wonder Valley at exactly this layer.

The documented read of Wonder Valley is that it is the clearest Canadian case study of AI infrastructure outrunning the constitutional consultation framework: real industrial ambition, real carbon implications, and a First Nation in court asking whether the country's AI buildout answers to Treaty 8 or simply happens on top of it. It remains the most important Canadian AI infrastructure story to watch through the rest of the decade, now as a test of whether AI for All's sovereignty framing can survive contact with section 35, rather than as the Indigenous-participation model earlier coverage (and an earlier draft of this guide) wanted it to be.

The siting backlash spreads

Lire cette section en français

Wonder Valley is the most documented Canadian siting fight, but by mid-2026 it had stopped being the only one. Community opposition has become a real variable in where these buildings go. In Armour Township, Ontario, a developer withdrew a proposed 2 MW AI data centre in June 2026 after a petition passed a thousand signatures, with water use, power draw, and noise the recurring objections; hundreds turned out against a proposed Hamilton campus, and protesters marched against new data centres in Vancouver. The common complaint is not Luddism. It is that the public was handed a fait accompli without the numbers (how much water, how much power, whose grid, whose bill) and started asking for them. Hamilton's council went a step further, directing staff to draft a moratorium bylaw pending a local framework. The developers are not silent in this: the head of one Hamilton proposal argued that building cleaner data centres is better than building none, and that swearing off the sector cedes both the infrastructure and the chance to do it responsibly. It is a fair point, and it is also the argument every applicant has an interest in making; the live question is whether "cleaner" is a binding commitment or a brochure.

Meanwhile the provincial scramble pulls in the opposite direction from the federal clean-power framing. Alberta is openly pitching cheap natural gas to win the buildout: of roughly 100 hyperscale data centres in the works nationally, about 90 percent are slated for Alberta, whose grid runs several times the national emissions intensity, and the province's own technology minister has called gas "really the only option for the next three to five years" — though the grid mechanics, AESO's connection cap, and Alberta's formal AI Data Centres Strategy are the next section's subject. Saskatchewan is courting the same rush: Bell is building what it bills as Canada's largest AI data centre near Regina, a roughly $1.7-billion project approved amid protest, and Saskatoon's chamber of commerce wants the next one. It is the siting paradox in real time, the compute racing toward the cheapest and dirtiest power while the cleanest provinces ration access.


The grid question

Lire cette section en français

Canada's electricity-supply story is a structural advantage that the AI infrastructure conversation often takes for granted, and which deserves clearer naming.

Roughly 82% of Canadian electricity generation in 2023 came from non-emitting sources — hydropower (about 60% of total generation), nuclear (about 15%), wind and solar (combined about 7%). The remaining 18% is fossil-fuel-based, concentrated heavily in Alberta (natural gas) and Saskatchewan (a mix of gas and remaining coal). Quebec, Manitoba, British Columbia, Newfoundland and Labrador, and Yukon are predominantly hydropower — meaning a data centre sited in any of these provinces operates on a grid with very low carbon intensity per kilowatt-hour.

This is the structural reason Quebec hosts Canada's largest data-centre cluster by demand (Montréal at 520 MW) and is the preferred Canadian destination for hyperscalers that have made public renewable-energy commitments. A data centre in Montréal running on the Hydro-Québec grid has roughly one-fifteenth the carbon intensity of a comparable data centre in Alberta running on the gas-dominated provincial grid.

This is also the structural reason the AI for All sovereignty framing intersects awkwardly with Canadian energy politics. Alberta's data-centre attractiveness (cheap electricity, favourable policy) is partly a function of its natural-gas-dominant grid. Increasing Canadian AI compute in Alberta — which is happening, including with AWS Canada West and Wonder Valley — increases the carbon intensity of Canadian AI workloads compared to siting the same load in Quebec or British Columbia. The strategy doesn't commit to siting requirements that would direct AI compute toward lower-carbon provinces. Whether it should is a contested policy question (the BC + AI critique surfaced something like this; Alberta's provincial government would push back hard against any such requirement on federal-jurisdiction grounds).

ALIGNMENT — follow the megawatts. When a data centre lands in your region, the press release talks jobs and innovation. The question that tells you more: where does the power come from, and who pays for the grid it leans on? Compute always has to plug into something. Find the something.

A finding worth foregrounding: provincial grid operators are no longer treating data-centre demand as routine industrial load. Across 2024-2026, four major provinces moved toward explicit allocation regimes for AI and data-centre electricity demand, in ways that materially constrain where new capacity can land:

  • Ontario. The Independent Electricity System Operator (IESO) projects that data centres will represent 13% of new electricity demand by 2035 and 4% of total anticipated Ontario demand. Ontario reached 145.6 TWh of demand in 2025. The province is formalizing new data-centre connection requirements that go beyond standard industrial-load processes.
  • Alberta. The Alberta Electric System Operator (AESO) received 29 proposed data-centre projects totaling more than 16 GW of requested load in 2025 — a scale that exceeds the province's entire current peak demand. AESO imposed an interim 1,200 MW large-load connection cap while developing a longer-term integration framework. Alberta also announced an explicit AI data-centres strategy.
  • British Columbia. BC Hydro has moved to competitive allocation of electricity, setting aside 300 MW for AI projects and 100 MW for data centres in the first two-year call. Cryptocurrency-mining loads have been permanently excluded from the allocation framework — a notable policy choice the guide will return to.
  • Quebec. Hydro-Québec proposed a dedicated large-data-centre tariff in 2026 designed to reflect the value of renewable electricity rather than treating data centres as standard industrial customers. The utility already maintains a large-power tariff regime for 5 MW+ customers.

This is a substantive change in the Canadian electricity-policy landscape. Grid access for AI infrastructure is becoming a policy allocation problem, not a real-estate problem. Where data centres go in Canada will be increasingly determined by provincial regulators making explicit choices about which loads get to connect, in what queue order, at what tariff, with what flexibility commitments — not by where the cheapest land happens to be available.

There is a critical nuance here. Quebec and BC are routinely framed as Canada's lowest-carbon AI siting destinations because of hydropower dominance. The Canada Energy Regulator documents that this advantage has structural limits: Quebec's electricity exports fell from 25.9 TWh in 2019 to 13.3 TWh in 2023 amid lower precipitation and changing domestic demand patterns; British Columbia has increased imports from the US Pacific Northwest for similar reasons. Hydropower output varies year-to-year with precipitation in ways that natural-gas generation does not. The framing "clean power" and the framing "surplus power" are not equivalent. Quebec's hydro grid is low-carbon; its capacity to absorb new large loads is not unlimited, and is subject to climate-driven variability the AI infrastructure conversation has largely not engaged. Site C in BC strengthens long-term supply; it does not eliminate the variability.

On water, ask for the facility’s cooling design, annual water intake and water source. Cooler weather can help a facility use outside air for cooling, but the design and local operating conditions decide what it saves. PUE measures energy. It does not tell you how much water the facility takes from the local system.

The international comparison context: the global average for data-centre electricity is around 2% of total demand (IEA, 2024). Canada's 2.1% (NRCan, 2023) is roughly at the global average. The US is closer to 4–4.5%: Lawrence Berkeley National Laboratory put US data centres at 4.4% of US electricity in 2023, with projections of 6.7–12% by 2028. Canada is currently smaller as a share of demand than the US but growing fast on a domestic-share basis (the 50% jump from 1.4% to 2.1% in two years).


The transparency gap, in operational terms

Lire cette section en français

Source: Natural Resources Canada’s data centre guide

If you remember one thing from this chapter for the rest of the guide, make it this: most Canadian data centres do not publicly report their basic operational metrics.

The comparison worth making is between records. Can we see comparable, measured results for individual Canadian sites, with a reporting period and a definition of what was measured? The NRCan assessment and the Uptime survey tell us about reported practices. Neither supplies that facility-by-facility record.

The EU requires facility-level reporting. Under its Energy Efficiency Directive and Delegated Regulation (EU) 2024/1364, covered centres with at least 500 kW of installed IT power demand submit annual data, including energy and water inputs. The public database shows aggregates for Member States and the Union. Canada’s September 2026 development principles call for transparency with communities. The principles themselves set no comparable annual reporting schedule or public database.

This matters because the rest of the guide's environmental, sovereignty, and accountability conversations depend on data the operators are not currently required to share. When Telus or Bell claims that their Canadian-sovereign-cloud offerings are environmentally favourable, that claim is currently unverifiable, not because the claim is necessarily false, but because the data that would let an independent observer verify it is not in the public record. When a Toronto data centre says it uses free-air cooling for 80% of the year, you have the operator's word and no third-party measurement.

This is one of the specific silences the guide's bias-mapping methodology asks readers to notice. The strategy says it values sovereign infrastructure. The operational metrics that would let citizens verify whether the sovereign infrastructure is delivering on its environmental and reliability claims are mostly not published. The gap between the framing and the disclosure is the gap to track.


The infrastructure, totalled

Lire cette section en français

CHAPTER RECAP — you now have: - A working physical model of what's actually in a data centre — racks, servers, GPUs, electrical demand, cooling regimes, the PUE and WUE metrics that describe efficiency. - Canada's actual current footprint: 300+ existing facilities, 60+ in pipeline, the city-scale numbers (Toronto ~320 MW, Montréal ~520 MW, Vancouver ~46 MW, Calgary rising), and the 50% two-year growth in data-centre electricity share (1.4% to 2.1% of total Canadian demand). - Foreign hyperscalers operate Canadian regions, and Canadian-controlled operators have facilities and projects at different stages. Compare like-for-like operating capacity before assigning either group a share of the Canadian market. - The federal 500 MW compute access objective, Bell’s initial 300 MW Sherwood build and the earlier $240 million Cohere commitment for Cambridge. The Wonder Valley consultation dispute is a separate case. - The grid story: 82% Canadian electricity from non-emitting sources, Quebec and Manitoba as the lowest-carbon hosts, Alberta's data-centre growth carrying a carbon intensity 10–15× higher per kWh, and the absence of federal siting requirements in AI for All. - How to distinguish a market assessment, a survey on corporate reporting and a public database of regulatory aggregates.

The map tells us where the facilities are. In Chapter 7, follow the contracts, the money and the research that turns a company’s footprint into an economic-impact claim.

You can now read claims about Canadian AI infrastructure in a way that distinguishes operational reality from strategic framing — which is most of what reading Canadian AI policy requires.


Bias label for this chapter: technical-operational reportage with Canadian-policy framing. Author lean: skeptical of corporate self-disclosure unverified by independent third party; sympathetic to mandatory-reporting regimes (EU model) as a baseline for accountability; cautious about treating individual Canadian Indigenous-equity projects as generalizable models without evidence. Government technical sources (NRCan) treated as authoritative within the scope they cover and as leaning toward international-best-practice framing. Hyperscaler economic-impact data labelled as commissioned-research (Public First) and read accordingly. Indigenous-equity case studies (Wonder Valley) reported with both proponent and critic framings represented.

Primary sources cited or relied on in this chapter: Natural Resources Canada, Best Practice Guide for Canadian Data Centres (November 2024); IEA, World Energy Outlook 2024 data-centre projections; Public First, AWS Canada Impact Report; Google Cloud Blog, "Measuring the environmental impact of AI inference" (Vahdat & Dean, August 21, 2025); Luccioni, Strubell & Crawford (FAccT 2025); CUPE Senate Brief (March 2026); BC + AI "AI for All Has To Mean All of Us" (June 5, 2026); Owen et al. (2023); Canada Energy Regulator provincial profiles and 2023 hydro variability documentation; Independent Electricity System Operator (IESO) Ontario data-centre demand projections; Alberta Electric System Operator (AESO) 2025 large-load queue documentation and 1,200 MW interim cap; BC Hydro competitive allocation framework documentation including crypto exclusion; Hydro-Québec large-data-centre tariff proposal (2026); Uptime Institute Global Data Center Survey 2024; eStruxture, Cologix, Vantage, and QScale operator disclosures including QScale Q01 Lévis campus announcement (June 2026). Detailed citations in the Sources appendix.


🧌 GOBLIN CHECK — Before you compare two percentages, ask who was counted and who can see the numbers. A facility can measure its efficiency, send it to head office and never publish it. If the survey doesn’t tell you which happened, the goblin leaves that box blank.

🧌 GOBLIN CHECK — a correction, filed in public. An earlier draft of this chapter called Sturgeon Lake Cree Nation a Wonder Valley equity partner. Then the receipts came in: an open letter demanding a halt, news reports of zero consultation, a courtroom. When the narrative and the documents disagree, the narrative loses — that's the whole rule of this book, and it applies hardest to the book itself. The corrected record stays. So does this note, as the worked example.

Recap

  • A working physical model of what's actually in a data centre — racks, servers, GPUs, electrical demand, cooling regimes, the PUE and WUE metrics that describe efficiency.
  • Canada's actual current footprint: 300+ existing facilities, 60+ in pipeline, the city-scale numbers (Toronto ~320 MW, Montréal ~520 MW, Vancouver ~46 MW, Calgary rising), and the 50% two-year growth in data-centre electricity share (1.4% to 2.1% of total Canadian demand).
  • Foreign hyperscalers operate Canadian regions, and Canadian-controlled operators have facilities and projects at different stages. Compare like-for-like operating capacity before assigning either group a share of the Canadian market.
  • The federal 500 MW compute access objective, Bell’s initial 300 MW Sherwood build and the earlier $240 million Cohere commitment for Cambridge. The Wonder Valley consultation dispute is a separate case.
  • The grid story: 82% Canadian electricity from non-emitting sources, Quebec and Manitoba as the lowest-carbon hosts, Alberta's data-centre growth carrying a carbon intensity 10–15× higher per kWh, and the absence of federal siting requirements in AI for All.
  • How to distinguish a market assessment, a survey on corporate reporting and a public database of regulatory aggregates.

Sources

  • Natural Resources Canada, Best Practice Guide for Canadian Data Centres (November 2024)
  • IEA, World Energy Outlook 2024 data-centre projections
  • Public First, AWS Canada Impact Report
  • Google Cloud Blog, "Measuring the environmental impact of AI inference" (Vahdat & Dean, August 21, 2025)
  • Luccioni, Strubell & Crawford (FAccT 2025)
  • CUPE Senate Brief (March 2026)
  • BC + AI "AI for All Has To Mean All of Us" (June 5, 2026)
  • Owen et al. (2023)
  • Canada Energy Regulator provincial profiles and 2023 hydro variability documentation
  • Independent Electricity System Operator (IESO) Ontario data-centre demand projections
  • Alberta Electric System Operator (AESO) 2025 large-load queue documentation and 1,200 MW interim cap
  • BC Hydro competitive allocation framework documentation including crypto exclusion
  • Hydro-Québec large-data-centre tariff proposal (2026)
  • Uptime Institute Global Data Center Survey 2024
  • eStruxture, Cologix, Vantage, and QScale operator disclosures including QScale Q01 Lévis campus announcement (June 2026).