The government of Canada has set a national baseline for how artificial intelligence data centres may be built, and twenty-three of the largest AI and cloud companies have signed it. The framework is voluntary, carries no penalties, and rests on an uncomfortable admission: the country that signed the pledge still rents most of the compute its courts, hospitals, banks and ministries depend on. The sovereignty question is no longer whether capacity gets built, but who is allowed to switch it off.
Featured image: a data centre corridor used to illustrate Canada’s responsible data centre framework. Credit: BetaKit, 3 September 2026.
What the original report says
On 3 September 2026, the Government of Canada published Canada’s Responsible Data Centre Development Principles, announced in Markham, Ontario, by the Honourable Evan Solomon, Minister of Artificial Intelligence and Digital Innovation, alongside Tim Tierney, President of the Federation of Canadian Municipalities. The document sets five expectations: data centres must create lasting local benefits; must not shift electricity costs onto Canadians; must minimise water use and environmental impacts; must be transparent about local impacts; and must bring strategic value to Canada, including legal protection, cybersecurity and reliable Canadian access for sensitive workloads. Twenty-three companies signed on, among them Amazon Web Services, Google, Microsoft, OpenAI, Anthropic, Meta, Cohere, Bell, TELUS and OVHcloud. The principles are voluntary, contain no compliance mechanism or penalties, and state that Canada currently relies significantly on infrastructure and services based outside the country for the compute, cloud and data storage that power AI.
THE SCENARIO: A hospital network signs a five-year compute agreement with a hyperscaler running in a neighbouring province. The regional data residency clause is satisfied. Eighteen months later, the provider’s home government issues an export directive restricting an advanced model tier to named jurisdictions. The contract is valid, the servers are nearby, and the clinical triage tool stops working on a Tuesday morning. The board discovers it has strong data governance and no execution governance at all.
The Question
How does a nation build AI infrastructure fast enough to compete while keeping the legal and operational ability to keep its critical systems running when a foreign government decides otherwise?
That is the question Canada’s Responsible Data Centre Development Principles raise without answering. The framework is a genuine improvement on silence. It names five conditions that host communities can hold operators to, and it commits signatories to disclose power use, water use and local impacts in independently verifiable terms. The fifth principle goes further than most national AI strategies by stating that facilities supporting sensitive data, public services and critical infrastructure must meet appropriate requirements for Canadian legal protection, cybersecurity, continuity of service and reliable Canadian access.
Those last three words carry the whole problem. Reliable Canadian access is not a feature a company can promise. It is a capability a state either possesses or does not. When compute, model weights and runtime privileges sit under another jurisdiction’s legal authority, continuity is borrowed rather than owned. The federal government says as much in its own document, noting that Canada currently relies significantly on infrastructure and services based outside the country for the compute, cloud and data storage that power AI.
The question is not whether voluntary principles are better than nothing. They are. The question is whether a pledge can substitute for capability. That matters because the gap between a signed commitment and an enforceable right of access is precisely where sovereignty is lost, and it is usually lost quietly, inside a procurement decision that looked entirely reasonable on the day it was approved.
What Happened and Why It Matters
The framework was launched on 3 September 2026 in Markham, Ontario, by Minister Evan Solomon, joined by Tim Tierney of the Federation of Canadian Municipalities. Reporting by BetaKit records that twenty-three signatories agreed to the principles and that the framework, in Solomon’s own framing, asks rather than compels. There are no binding definitions and no penalties for non-compliance. Enforcement is left to provinces and municipalities through their own approvals, with the federal contribution being data rather than authority.
The Logic reported from the announcement that the accord emerged from months of consultation with municipalities, provincial and territorial governments, Indigenous groups and industry, and that signatories include American cloud providers Amazon, Google and Microsoft, US AI developers Anthropic, Meta and OpenAI, and Canadian firms including Cohere, Bell, TELUS, QScale, eStruxture and Hypertec. The federal government’s own text on the Innovation, Science and Economic Development site sets out the five numbered expectations in full, including the requirement that proponents pay the costs of connecting and serving their projects and that projects not compromise grid reliability.
The water and electricity provisions are more specific than the sovereignty ones. Principle two states that data centre growth must not shift project-driven electricity system costs onto households and existing businesses, and that proponents must cover new generation, transmission, substations and grid upgrades directly attributable to their development. Principle three prefers closed-loop or high-efficiency water technologies and waste heat recovery. Principle four requires clear, project-appropriate and independently verifiable information on expected local impacts including power and water use, infrastructure requirements, sound and emissions.
Principle five is where the sovereignty interest sits, and it is deliberately less defined. It asks that investments create clear and lasting benefits for Canada through investment, tax revenue, Canadian supply-chain participation, research partnerships, workforce development and access to compute for Canadian businesses and institutions. For facilities supporting sensitive data and critical infrastructure, it says those must meet appropriate requirements for Canadian legal protection, cybersecurity, continuity of service and reliable Canadian access. The word appropriate is doing substantial work in that sentence.
That matters because the surrounding policy environment is tightening, not loosening. Texas governor Greg Abbott issued a directive in September 2026 barring the state environmental agency from issuing new permits for data centre projects until an audit of grid interconnection requests is complete, effectively freezing permitting in the fastest-growing data centre market in the United States. California governor Gavin Newsom signed seven bills in the same week requiring operators to pay for grid upgrades, shoulder a greater share of wildfire mitigation and liability costs, disclose power and water demand, and incorporate on-site clean energy. Maryland governor Wes Moore signed an executive order creating a single review standard for large data centres and calling on lawmakers to repeal a 2020 tax exemption passed under his predecessor.
Utility Dive reported on 23 September 2026 that more than 170 billion dollars in AI data centre capacity has been blocked, withdrawn or stalled by community opposition since January 2024, and that Texas has paused new interconnections pending an audit of a 474-gigawatt queue, more than five times the state’s all-time peak demand. The direction of travel across North America is a tighter regulatory perimeter around compute. Canada’s principles sit inside that trend rather than ahead of it.
Meanwhile the dependency the framework acknowledges is concentrated. Research cited in coverage of the announcement notes that close to ninety per cent of global compute, AI investment and leading foundation models sit in the United States and China. A voluntary national framework cannot shift that distribution. It can only set the terms on which capacity lands. For leaders, the practical issue is that the principles govern the building of infrastructure while leaving the control of it largely where it already was.
The Sovereignty Risk
The principle’s own text describes the risk precisely: facilities supporting sensitive data, public services and critical infrastructure must meet appropriate requirements for Canadian legal protection, cybersecurity, continuity of service and reliable Canadian access. Each of those is a legal and operational property, not a social one. None can be secured by a signature.
Canadian legal protection means the data is subject to Canadian law in practice, not merely stored on Canadian soil. A facility owned by a foreign parent can hold Canadian data in Ontario while remaining subject to a foreign legal orders regime that reaches the parent. Data residency and legal jurisdiction are different properties, and only one of them appears in a standard procurement checklist.
Continuity of service is the harder property. Continuity means the state can keep critical systems running through a commercial outage, a corporate restructuring, a sanctions event or a foreign policy reversal. The documented pattern is that interruption arrives with almost no notice. In early 2026 the US administration required Anthropic to suspend access to its most advanced models, giving the company roughly ninety minutes to comply, an event that Bloomberg and the Financial Times reported as triggering an urgent UK government review into the economic and security risks of losing access to frontier AI.
Reliable Canadian access is the point where principle five runs out of instruments. A voluntary accord can request disclosure. It cannot compel a foreign provider to keep serving a Canadian hospital if that provider’s home regulator decides otherwise. What it can do is make the exposure visible early enough for institutions to design around it, and that is a genuine, if limited, contribution.
The accountability gap is structural. The entities that control the intelligence, the information and the machines sit in three different layers, and the state holds leverage over none of them directly. Model sovereignty asks who controls the intelligence. Data sovereignty asks who controls the information. Compute sovereignty asks whose machines it runs on. Canada’s principles operate almost entirely at the third layer and gesture at the first two.
There is also a capability risk that rarely appears in the framework documents. If critical systems run on foreign platforms under foreign terms, domestic institutions lose the muscle memory of governing them. Procurement teams negotiate without understanding inference economics. Risk committees cannot assess a model they cannot inspect. Regulators write guidance for systems they cannot test. Over time the state retains formal authority and loses practical competence, which is the quietest form of sovereignty loss on record.
Concentration compounds all of it. When nearly ninety per cent of compute and frontier model capacity sits in two jurisdictions, diversification options are thin, and every alternative carries a price premium or a capability penalty. That is not a reason to abandon the framework. It is a reason to treat the framework as the beginning of an industrial capability programme rather than the end of a policy debate.
The TEE Method Response
The TEE Method reads any infrastructure question through three connected lenses: Talent, Enterprise and Ecosystem. Applied to Canada’s principles, each lens reveals a different gap between what was signed and what was secured.
Talent is the capacity to design, procure, operate and audit sovereign infrastructure. Principle one commits signatories to jobs, skills training and apprenticeships, which addresses construction and operations labour. It does not address the scarcer capability: people who can write a sovereign compute clause into a contract, evaluate a model’s jurisdictional exposure, and run an exit test on a live dependency. A country can build data centres at scale and still lack the thirty or forty specialists each major institution needs to govern them. Canada’s research strength in AI is well documented; converting it into institutional governance talent is a separate programme with a separate budget line.
Enterprise is the discipline of treating AI systems as governed assets with named owners, documented dependencies and tested failure paths. The principles ask operators to disclose power, water and local impacts, all of which are estate-level facts. Enterprise sovereignty asks a firmer question: which specific workloads would stop if a named provider withdrew access next quarter, and what is the documented fallback. Most institutions cannot answer that today, because the answer sits across four teams that do not share a model inventory.
Ecosystem is the surrounding structure of suppliers, standards, regulators, financiers and community institutions. Canada’s framework is an ecosystem instrument: it convenes twenty-three firms, gives municipalities verifiable data, and creates a common baseline for negotiation. The missing ecosystem layer is demand aggregation. Twenty-three signatories competing individually for grid connections and provincial support will not produce the leverage that a coordinated public buyer could. Where states have secured better terms, it has usually been by acting as a single buyer for public sector compute, not by asking nicely.
Read together, the lenses show the framework’s real function. It is a Talent and Ecosystem instrument wearing a compute policy hat. It improves the quality of information available to host communities and it signals a national standard. It does not yet build Enterprise capability inside institutions that will be exposed to interruption, and that is where the sovereignty work remains undone.
Sovereignty Test Matrix
The following matrix scores the Canadian framework across five sovereignty domains, each rated on a one-to-five scale where one indicates no sovereign capability and five indicates full sovereign control. The test is not whether the framework is good policy. The test is where it leaves the state’s actual leverage.
| Domain | Score | Weak signal | Stronger signal |
|---|---|---|---|
| Political | 3 of 5 | Voluntary accord with no penalties; enforcement deferred to provinces and municipalities | Statutory obligation with named regulator, published compliance findings and remedy power |
| Economic | 2 of 5 | Cost-shifting rules are clear on electricity, silent on compute demand concentration | Public anchor demand, coordinated procurement, domestic supplier share targets with reporting |
| Cultural | 3 of 5 | Local benefit commitments and Indigenous engagement named but self-reported | Host communities hold veto-adjacent rights and independently verified benefit agreements |
| Intellectual | 2 of 5 | No requirement for Canadian model evaluation, training data provenance or inspection rights | National evaluation capability with the legal right to test models used in public services |
| Technological | 3 of 5 | Compute capacity is being built on foreign-owned platforms under foreign legal orders regimes | Critical workloads run on infrastructure the state can keep running under any external decision |
The pattern is consistent. Canada scores highest where the framework imposes costs on operators and lowest where it would need to assert authority over foreign providers. That is the signature of a consumer-side framework in a producer-side market. It protects ratepayers and host communities effectively, and it leaves the legal protection, continuity and reliable access commitments as aspirations rather than entitlements.
For leaders, the matrix is useful as a self-assessment rather than a scorecard for Ottawa. An institution that scores two or below on the intellectual and technological rows is carrying dependency it has not yet quantified. The threshold that matters is concrete: any workload that supports health, safety, payment, identity or statutory reporting, and that cannot be run elsewhere within thirty days, is a sovereignty exposure regardless of where its data is stored.
What Leaders Should Do This Week
The framework gives institutions a legitimate reference point. These seven actions convert it into operating decisions that can be assigned and closed.
First, build a thirty-day dependency register. Name every production AI system, the provider behind it, the jurisdiction whose legal orders reach that provider, and whether the system could be replaced or rerun elsewhere within thirty days. The register should fit on two pages and be owned by a single named executive. Most organisations discover during this exercise that at least one critical workflow depends on a model tier they are not contractually guaranteed.
Second, check the contract for the three properties that matter. Locate the data residency clause, then read it against the parent company’s jurisdiction rather than the server’s location. Look for a continuity-of-service commitment with a defined notice period and a remedy. Where the standard terms are silent, request a side letter. A provider that will not put continuity in writing has told the buyer something important.
Third, apply the thirty-day rule as a decision threshold. Any critical workload that cannot be run elsewhere within thirty days moves to a hardened tier: an open-weight model hosted on infrastructure the institution controls, a second provider in a different jurisdiction, or a documented manual fallback that has been rehearsed rather than assumed. The threshold is deliberately blunt because a more nuanced one will not survive a steering committee.
Fourth, test the fallback once, in production conditions. Run the alternate path during a low-traffic window with the primary disabled and record what breaks. An untested fallback is a plan, not a capability. Institutions that have done this in the public sector generally find the gap is not compute but data pipelines and evaluation fixtures.
Fifth, put a sovereignty clause into the next procurement, not the current renewal. The next award is the cheapest moment to change terms. Specify jurisdiction of legal orders, continuity commitments, exit assistance, model-version change notice, and access to evaluation evidence for any model touching public services or regulated data. Pair this with a standing requirement that any new AI system be entered in the dependency register before go-live.
Sixth, fund the governance talent explicitly. Identify who owns model evaluation, who owns jurisdictional risk, and who owns exit planning, and give each a budget line. Where the institution cannot hire, it should contract a named specialist and rotate staff through the work. The capability gap is smaller than it appears, but only if it is funded rather than assumed.
Seventh, publish the disclosure the framework asks for even where there is no obligation to. Independent, verifiable reporting on power, water and dependency is now a credibility asset with regulators, boards and host communities. Institutions that disclose voluntarily before a mandate arrives set the terms of the conversation rather than inheriting them.
The Question Revisited
How does a nation build AI infrastructure fast enough to compete while keeping the legal and operational ability to keep its critical systems running when a foreign government decides otherwise?
The answer is that it builds capability at the same time as it builds capacity. Canada’s Responsible Data Centre Development Principles are a real contribution: they give host communities a common baseline, they force disclosure that did not previously exist, and they name continuity and reliable Canadian access as national requirements for the first time in a federal document. That last point matters, because a requirement that has been named can be legislated. What the framework does not do is create the authority to enforce any of it against providers headquartered elsewhere.
The TEE Method resolves the gap by treating sovereignty as something institutions practise daily rather than something governments declare. Talent supplies the people who can write and audit the clauses. Enterprise supplies the inventory, the ownership and the tested fallbacks. Ecosystem supplies the demand aggregation and standards that make domestic alternatives viable at commercial scale. A nation whose major institutions have all three can sign a voluntary framework and remain safe, because the signature is not the source of its protection. A nation without them can sign the same document and remain exposed, because a pledge is not a switch.
The sovereignty test is not whether the right principles were adopted. It is whether any institution could name, this week, the one workload it would lose first and what it has already built to keep running without it.
[tt_consultation_booking]
Editorial note: This is a news-led Sovereignty Briefing analysis produced with Hermes-assisted research and source checks.
Sources
This article draws on the TEE Method™ framework from SOVEREIGN: Who Owns the Future? For the complete framework and advisory programmes, see tonishatagoe.com.
Sign in to save