A framework for Canadian AI infrastructure — where we are today, and where we should aim.
Sovereignty has become the organizing word of Canada's AI moment. It runs through the federal compute strategy, the multi-billion-dollar push into public supercomputing, and the first wave of large-scale data-centre announcements. The ambition is right, and the building is worth applauding: Canada needs this infrastructure, and every facility that breaks ground here means construction and operating jobs, capital investment, demand for Canadian power, and another piece of a domestic digital backbone the country will depend on for decades. What is missing is a definition. Sovereignty is invoked nearly everywhere and pinned down almost nowhere — and for AI infrastructure in particular, it is treated as a property of geography when it is really a property of control.
A data centre can be entirely Canadian in physical form — Canadian site, Canadian power, Canadian crews — while the control over what runs inside it, and the ownership of what it produces, sit elsewhere. That is not a failure and it is not a trick. It is what happens when "sovereign" is left undefined: the easiest thing to deliver, location, quietly stands in for everything harder. The purpose of this framework is to separate the layers, so the country can see clearly which ones it already holds, which ones it should hold, and which ones it can sensibly share with partners.
Contents
I. What sovereignty actually means
II. The stack: seven layers from soil to sovereignty
III. The opportunity for Canada
IV. Conclusion
I. What sovereignty actually means
Start with location, because it is the layer most easily dismissed and it deserves the opposite. A data centre on Canadian soil is jobs in construction and operations, capital put to work, new demand for Canadian power, and an anchor for a domestic digital economy. It establishes Canadian jurisdiction over the physical plant. None of that requires a single further layer to be worth having; locating this infrastructure in Canada is a real and sufficient good on its own terms.
But location is not the same as control, and sovereignty is a question of control. In the AI context it is the degree to which Canadian hands — governments, institutions, and companies — hold the decisions that matter over a system: who it serves, who can interrupt or switch it off, whose laws can reach inside it, and who owns what it produces and the value that flows from it. It is a property of decision rights, not coordinates. The more of those rights sit in Canada, the more sovereign the infrastructure, whatever flag flies over the building. Canada's own data-sovereignty analysis already concedes the narrow version of this — data stored in Canada can still be reached by foreign law if the provider is subject to that law — but the principle is broader than data.
And it runs through a stack. Control is not one thing, held or lost in a single place; it is distributed across the physical plant, the operating software, the corporate structure, and the law. Which is why the only useful question to ask of any "sovereign" claim is: sovereign in which layers?
II. The stack: seven layers from soil to sovereignty
The layers are not equal — the upper ones govern the lower ones. A project can be complete at the bottom and still cede control at the top. Climbing:
Layer 1 — Location. The site is in Canada. As above, this is the foundation and valuable in its own right; it is also the beginning of the question rather than the end of it.
Layer 2 — Construction. Canadian builders, trades, and engineering, and vetted equipment. The consideration here is supply-chain trust, applied with precision: core networking, firmware, remote-management, and monitoring systems warrant scrutiny that commodity steel and copper do not. Canada drew this line already in telecommunications. The question is trusted components in the systems that could be turned against a facility — not country-of-origin for everything inside it.
Layer 3 — Power. Who generates the electricity, on whose grid, and under whose law. This is the layer where Canada has the most to gain, and the one the "energy superpower" ambition rightly points at — but it is more complicated than a national slogan allows, because the Canadian grid is really a set of regional grids, several of them wired north–south into US markets as much as to the rest of the country. In Quebec, British Columbia, Manitoba, and Labrador, power is abundant, overwhelmingly hydro, and under provincial Crown control: genuinely sovereign, and a rare comparative advantage. In much of the Maritimes and the Prairies it is fossil-heavy, thinly interconnected, or reliant on cross-border flows — closer to continental power than sovereign power. Even Canada's large exporters sit inside an integrated North American system. Power sovereignty, in other words, is a siting and contracting decision, not a fact about the map; the opportunity is to point the genuinely sovereign, clean power Canada does have at this infrastructure.
Layer 4 — Operations. Who runs the facility and maintains it, and — distinct from physical operations — who holds the control plane: root on the scheduler, identity and key management, firmware authority, the ability to patch the systems that orchestrate the compute. Physical operations are readily Canadian — but the people who keep the cooling and the electrical plant running are not the people who control the cluster. The control plane is the layer most often retained abroad, and the one that determines who can actually direct or halt the system. It is the distinction the word "sovereign" most often blurs.
Layer 5 — Ownership. Ownership runs from the land and the building, to the compute inside it — the accelerators, the capital that depreciates — to the connectivity that carries the data out. Foreign ownership of the compute is normal and, for much of the market, entirely appropriate; the point is to know where it sits, not to resent it. But ownership is also where Canada holds an advantage few countries can match: some of the largest and most sophisticated infrastructure investors in the world are Canadian pension funds, and they have spent two decades deploying that capital abroad for want of comparable assets at home. AI infrastructure — long-lived, capital-hungry, backed by contracted cash flows — is precisely the asset class they want. The capital to own these facilities domestically already exists; it has been waiting for the opportunities to arrive. And importantly, in Canada, ownership increasingly includes Indigenous equity participation alongside that capital, which makes the ownership layer both a sovereignty question and a vehicle for shared economic benefit.
Layer 6 — Jurisdiction. The legal perimeter — the layer you cannot photograph. It is set the day the operating company is incorporated and the day the contract is signed, and it determines whose courts, whose subpoenas, and whose emergency powers can reach inside the facility, regardless of where the facility physically sits. A US-incorporated operator carries US legal reach onto Canadian soil; the legal nationality of the operator travels with it. This is a property to be chosen, not a grievance to be aired: for commercial inference it rarely matters, and for government, health, and defence data it is close to the whole question. Because it follows from corporate structure rather than geography, it is also the cheapest layer to get right — which is most of the opportunity, below.
Layer 7 — Data and IP. This is the layer most often misframed. The usual worry is that a foreign tenant will train on Canadian power and carry the model — the appreciating asset — home. That gets the problem backwards. There is little hyperscale frontier-training demand in Canada to capture in the first place; the labs training the largest models are elsewhere and will remain so. The real questions at this layer are who is buying the compute, where the data and the resulting weights are stored, and whether Canadian law makes it attractive to keep them here. That last point does the most work and gets the least attention: a clear, favourable copyright and text-and-data-mining regime is what makes Canada a place where models can be trained and where the data and weights can legally and safely reside — and an uncertain one is what sends that activity offshore. The value Canada can actually capture here is threefold: a legal regime that anchors training data and model weights onshore; Canadian champions — Cohere foremost — that own their models and IP outright; and Canadian-controlled compute serving sovereign and strategic workloads. If Canadian inference is flying a submarine, or running inside a hospital or a department of government, the value has been captured, whoever built the chips.
III. The opportunity for Canada
One layer is not on the table, and pretending otherwise wastes effort: the chips. The leading accelerators are American-designed, fabricated in Taiwan or Korea, dependent on a single Dutch source of lithography, and fed by a handful of Asian memory suppliers. A leading-edge fab is a multi-decade, tens-of-billions undertaking that no near-term Canadian program will replicate, and federal policy sensibly accepts foreign hardware while asking for Canadian integration around it. The right standard for the hardware layer is not "made in Canada" but allied-trusted: designed and built in trusted jurisdictions, free of adversary content — the same conclusion the United Kingdom, Japan, Australia, and Germany have reached. Concede it cleanly; it frees capital and attention for the layers Canada can actually hold.
And those layers are a real opportunity, not a consolation prize. Two of them are decisive. The first is power: where it is clean, abundant, and under provincial control, it is a genuine and scarce advantage, and the ambition to be an energy superpower is the right instinct — provided the infrastructure is sited where the power is actually sovereign rather than merely Canadian-flagged. The second is almost free and routinely overlooked: the legal structure of the operating entity and the tenant. An operating company organised as Canadian-controlled, holding its administrative keys beyond the reach of any foreign court, with Canadian legal process written into its data terms, converts a physically Canadian facility into a jurisdictionally Canadian one without an additional yard of concrete. The reason it is rarely done is that it is rarely asked for — exactly what a national framework could change. Around those two sit the assets that make the opportunity Canada's to take: deep domestic pools of pension capital that have lacked a reason to invest at home; champions like Cohere that keep model ownership and IP onshore; an IP and copyright regime that, written well, would make Canada a place to train and store rather than a place training merely passes through; and the sovereign demand — government, health, defence — that turns Canadian-controlled compute into Canadian strategic capability. None of this requires owning the whole stack. It requires owning the parts that decide who is served, who is in control, and where the value settles.
IV. Conclusion
Much of the capacity now being proposed and built in Canada will be Canadian in its site, its power, and its construction, and foreign in its compute ownership, its tenant control, and its IP. That is a strong place to start, not a shortfall to apologise for — and for a great deal of commercial compute it is exactly the right answer. The ambition is to climb higher where the stakes warrant it: to hold the control plane, the legal perimeter, and the model itself for the regulated, public, and strategic workloads where control genuinely matters, and to leave the rest to partners. Sovereignty is not a bar every facility must clear; it is a height the country gains where the climb is worth it.
Soil is not sovereignty. It is the foundation sovereignty is built on — and Canada has both the foundation and the things that matter most above it: clean power where it counts, the capital to own what gets built, and the law and the champions to keep the value at home. The work ahead is not to domesticate every layer. It is to be clear-eyed about which ones decide who is served, who is in control, and where the value settles — and to build, deliberately, from the soil upward.