AI & strategy
Sovereign AI: What Quebec Really Controls
· 10 min read

On October 29, 2026, in Montreal, Prime Minister Mark Carney announced an equity investment of up to $1.6 billion in the Canadian company Cohere. His justification fit in one sentence: "Today, true sovereignty depends on our ability to build, adopt, and govern AI on our own terms" (Office of the Prime Minister of Canada, 2026).
Four months earlier, the Government of Quebec had signed two collaboration agreements, first with the same Cohere and then with France's Mistral, also presented in the name of digital sovereignty. These agreements have no financial impact and do not constitute a contractual commitment (Gouvernement du Québec, 2026d, 2026e).
The same word therefore covers very different instruments: an equity stake on one side, exploratory workshops on the other. That gap invites a simple question that public discourse often avoids: when it comes to artificial intelligence, what does Quebec actually control, and what is left to build?
One word, several layers of control
The Government of Quebec defines digital sovereignty as the ability to control and protect its infrastructure, technologies and data against foreign influence, in particular by preventing the laws of other countries from applying to them (Gouvernement du Québec, 2026c). The same page specifies, a telling detail for a political scientist, that digital sovereignty is not national sovereignty (Gouvernement du Québec, 2026c). The state takes care to depoliticize the term and turn it into a management issue.
That definition is worth breaking down. Applied to AI, sovereignty plays out on at least four layers:
- infrastructure: where the servers are, who supplies the compute, where the energy comes from;
- models: who designs them, who updates them, who can withdraw them;
- jurisdiction: which law applies, not to the server, but to the company that operates it;
- competence: can the state assess, audit and, if needed, replace what it buys?
A state can be strong on one layer and dependent on another. That is precisely Quebec's situation.
What Quebec has actually put in place
A policy statement, without a timeline. Approved in January 2026 and presented publicly on February 13, the Policy Statement on Digital Sovereignty and Information Technology Procurement pursues two objectives: increasing Quebec's digital sovereignty and maximizing the economic benefits of the state's IT investments (Ministère de la Cybersécurité et du Numérique, 2026a; Gouvernement du Québec, 2026b, 2026d). Its orientations are relevant: expanding the Quebec Government Cloud, building a sovereign cloud infrastructure that draws on Quebec's electricity and AI expertise, favouring open-source software, and writing data control clauses into contracts (Ministère de la Cybersécurité et du Numérique, 2026a). Yet the document sets no deadline and no quantified target, and it never names the foreign laws with extraterritorial reach against which it seeks to protect data (Ministère de la Cybersécurité et du Numérique, 2026a). The statement in fact describes itself as a strategic foundation meant to anchor a future policy (Ministère de la Cybersécurité et du Numérique, 2026a).
Hosting rules, with an open door. The government's cloud and hosting orientations, updated in June 2026, require highly sensitive data to be hosted on sovereign infrastructure (Ministère de la Cybersécurité et du Numérique, 2026b). The list of such infrastructure is instructive: the Quebec Government Cloud, a data centre owned by the Government of Quebec or of Canada, but also "a public cloud infrastructure designated by the minister" (Ministère de la Cybersécurité et du Numérique, 2026b). An organization unable to apply the orientation may also resort to risk mitigation measures within a documented process (Ministère de la Cybersécurité et du Numérique, 2026b). In other words, the line between sovereign infrastructure and commercial public cloud depends in part on a ministerial designation, and the public page does not specify which infrastructure has been designated.
Exploratory agreements. The memoranda signed with Cohere (announced June 9, 2026) and Mistral (announced June 16, 2026) provide for exchanges, workshops and discussions to better understand the conditions for a supervised deployment of AI in the public administration (Gouvernement du Québec, 2026d, 2026e). Neither specifies where data would be hosted (Gouvernement du Québec, 2026d, 2026e). Exploring before buying is good procurement practice, and it would be unfair to call it a failure. But these agreements should be named for what they are: learning tools, not control levers.
The legal knot the policy statement does not name
The law that weighs most heavily in this debate is American. Enacted in March 2018, the CLOUD Act clarified that providers subject to U.S. jurisdiction can be required to produce data they control, wherever it is stored (U.S. Department of Justice, 2019). The consequence is concrete: a server located in Montreal but operated by a company subject to U.S. law sits physically in Quebec without being governed exclusively by Quebec law.
The clearest testimony came from industry itself. On June 10, 2025, before the French Senate's commission of inquiry on public procurement, the rapporteur asked Microsoft France's director of public and legal affairs whether he could guarantee, under oath, that French citizens' data would never be passed to U.S. authorities without France's consent. His answer: "No, I cannot guarantee that," adding that it had never happened so far (Sénat [France], 2025, our translation).
Canada and the United States have also been negotiating, since March 2022, a bilateral agreement under the CLOUD Act to ease cross-border access to data in criminal investigations (U.S. Department of Justice, 2022). As Barry Appleton points out in a special report published in March 2026, however, the exposure of Canadian data does not depend on such an agreement being signed: it exists as soon as Canadian institutions rely on U.S.-linked cloud services (Appleton, 2026).
This is the heart of the problem: Quebec's policy statement refers to data being inaccessible to foreign third parties (Ministère de la Cybersécurité et du Numérique, 2026a), without naming the legal mechanism that makes them accessible today. A sovereignty policy that does not name the dependency it seeks to reduce struggles to set verifiable criteria.
The Canadian champion has become transatlantic
Cohere's trajectory illustrates a reality that sovereignty rhetoric often hides. In December 2024, Ottawa committed up to $240 million to a Cohere data centre project, the first investment under the Canadian Sovereign AI Compute Strategy; the facility was to be built with the American firm CoreWeave (Verdict, 2024). In April 2026, Cohere announced its combination with Germany's Aleph Alpha; the definitive agreement, signed on September 16, provides for dual headquarters in Berlin and Toronto (Unite.AI, 2026). Germany's Schwarz Group, owner of Lidl, committed to lead the Series E funding round (Unite.AI, 2026). That is the round the federal government joined on October 29, in exchange for specific commitments: at least 1,000 technical jobs in Canada and at least $1.7 billion in computing infrastructure in the country by 2030, with the investment paid out in stages (Office of the Prime Minister of Canada, 2026).
Mistral, Quebec's other partner, follows a similar logic. Its €1.7 billion Series C, announced in September 2025, was led by the Dutch company ASML, which became its main shareholder, with participation from existing investors including NVIDIA and the U.S. fund Andreessen Horowitz (Mistral AI, 2025).
The lesson matters. For a middle power, AI sovereignty does not take the form of autarky. It takes the form of chosen interdependence among allies. The federal AI for All strategy, published on June 4, 2026, embraces this explicitly with its "build, partner, buy" approach: build in Canada where possible, partner with trusted allies otherwise, buy on the market when appropriate (Innovation, Science and Economic Development Canada, 2026; Blake, Cassels & Graydon, 2026). For Quebec, the right question is therefore not "Quebec supplier or foreign supplier," but rather: which dependency, on whom, and reversible under what conditions?
The political price: when Washington calls sovereignty a barrier
Digital sovereignty now carries a diplomatic cost. On March 31, 2026, the Office of the United States Trade Representative published its annual report on foreign trade barriers (Office of the United States Trade Representative, 2026). For the first time, the term "sovereign cloud" appears in it, and Canada is named, alongside countries with far more restrictive policies (Geist, 2026). The report targets the federal sovereign cloud initiative, which would require data to be processed, transmitted and stored exclusively in Canada, by providers not subject to foreign laws allowing access without Canada's prior consent (Geist, 2026).
Ottawa did not back down: Shared Services Canada invoked the National Security Exception, which exempts every phase of its sovereign cloud procurements from applicable trade agreements (Shared Services Canada, 2026). This standoff is part of the CUSMA joint review that Washington is conducting in 2026 (Appleton, 2026).
For Quebec, the irony is real: the most concrete limit on its digital sovereignty may not be technical, but commercial. A provincial policy now sits inside a continental negotiation that the province does not control.
The serious objection: control rather than ownership
Any honest analysis must confront the best counterargument. In a Macdonald-Laurier Institute paper republished by ITIF in September 2026, Lawrence Zhang argues that Canadian ownership is a poor proxy for security (Zhang, 2026). He notes that actors linked to China compromised at least twenty Government of Canada networks over four years, and that a Canadian-owned data centre would not have prevented those intrusions (Zhang, 2026). His recommendation: achieve control through architecture and contracts (data kept in Canada, administration restricted to cleared personnel, customer-held encryption keys) rather than requiring Canadian ownership (Zhang, 2026).
The argument is strong, especially on cost: building a complete national infrastructure is expensive, and AI's productivity gains lie mostly downstream, in its uses (Zhang, 2026). Customer-held encryption keys also sharply reduce legal exposure, since the provider then stores data it cannot read (Zhang, 2026).
It does, however, have a limit. A contract does not neutralize the law the provider is subject to, and encryption protects confidentiality without guaranteeing continuity of service. The most reasonable synthesis distinguishes two perimeters: for the vast majority of data, contractual and technical control is enough; for a narrow core of critical data and systems, the operator's jurisdiction becomes decisive.
Five questions before signing
Rather than a label, AI sovereignty can be translated into verifiable questions, useful to any public decision-maker at the moment of purchase:
- Which law applies to the operator? Not to the server: to the company that runs it, and to its parent company.
- Who holds the encryption keys? If it is the provider, confidentiality rests on the provider's policy, not the state's.
- Can we leave? Open formats, portability of fine-tuned models and data, exit costs written into the contract.
- Who supplies compute and energy, and at what price? The policy statement counts on Quebec's electricity (Ministère de la Cybersécurité et du Numérique, 2026a), even as Hydro-Québec has asked the Régie de l'énergie for a dedicated rate for data centres above 5 MW, averaging about 13 cents per kWh, which would double what they currently pay; reviewed in public hearings in fall 2026, the application indicates that their peak demand, about 200 MW, could reach 1,000 MW by 2035 (Le Soleil, 2026a, 2026b). The energy advantage is real, but it is now being negotiated.
- Can the state assess what it buys? The Artificial Intelligence Expertise Centre of Quebec's Ministry of Cybersecurity and Digital Affairs exists for that purpose (Gouvernement du Québec, 2026a); its capacity to audit models supplied by third parties is a sovereignty lever in its own right.
Conclusion: from vocabulary to criteria
In 2026, Quebec acquired a vocabulary of digital sovereignty and its first hosting rules. That is real progress. But its AI instruments remain exploratory for now, while Ottawa has moved on to equity investment and the National Security Exception, and Washington treats sovereignty as a trade irritant.
The next step would be to turn intent into criteria: attach deadlines and indicators to the policy statement, publish the list of infrastructure designated as sovereign, and build questions like the five above into AI calls for tenders. Sovereignty is not a supplier's nationality. It is a set of controls, and controls can be verified.
(The author)

Strategy, Digital and AI Consultant
Trained in political science (from an honours bachelor's to a master's at Université de Montréal), I'm now completing a graduate diploma (DESS) in private- and public-sector management at HEC Montréal.
Keep reading
The AI ROI Paradox: Why Adoption Isn't Showing Up in the Bottom Line Yet
Nearly nine in ten organisations use AI, yet its effect on financial results is flat. What the data say, and the decisions that set apart the few organisations that truly benefit.
November 13, 2026 · 7 min
AI Act Delayed: What Europe's Retreat Teaches Canada
The European Union has pushed the AI Act's high-risk rules back to 2027 and 2028. An analysis of a calculated retreat, and what it means for Canada and Quebec, which still have no AI law.
November 12, 2026 · 8 min
Can Canada Really Break Free from Washington?
From the "rupture" declared at Davos to the Alliance for the Future signed with the European Union, Canada's middle-power doctrine tested against international relations theory and the reality of asymmetric interdependence.
November 14, 2026 · 10 min