Sovereign AI  ›  Canada

Canada's Sovereign AI Test

Turning research, power and compute into choice.

Canada has the researchers, electricity, companies and public institutions to shape an AI economy. The harder task is to connect them into systems Canadians can govern, use and replace when circumstances change.

🇨🇦 CanadaProgramme funding: Rung 1Strategic-autonomy builderManaged interdependence, applied

The question is choice, not ownership

Canada's AI debate is no longer only about whether the country can attract another data centre or train another model. It is about whether Canadian institutions and businesses can choose how important systems are built, where they run, which laws and people govern them, and what happens when a supplier, policy or network changes. A Canadian organization can use foreign chips and still retain meaningful control over a workload; a facility inside Canada can remain subject to decisions made elsewhere. What matters is the full chain: legal authority, operations, data access, software, network paths, trained staff, continuity and a credible route to another provider.

This analysis applies the managed-interdependence argument of the companion essay The Sovereign AI Paradox. Canada has globally recognized AI research centres, a growing group of technology firms, a relatively clean electricity system, public research infrastructure and access to a large North American market — but it has historically had limited public compute, relies on overseas suppliers for advanced accelerators and much cloud software, and has struggled to turn government technology spending into fast, repeatable delivery. The test is whether policy turns strong inputs into services that are useful, trusted and replaceable.

Canada's strengths begin above the chip

Canada's strongest position is a combination of research, institutions, applied industries and people. Toronto, Montréal, Edmonton and Waterloo have built deep research and startup communities; Canadian companies work in enterprise software, health, financial services, logistics, natural resources, public administration and language technology — real problems on which to develop and test AI.

Electricity

Natural Resources Canada reports 82.5% of Canadian generation comes from non-emitting sources, led by hydro. That creates lower-carbon computing opportunities in some provinces — but a national average cannot answer a community's questions about a specific site's grid connection, transmission, water or consent.

Photonics & packaging

The Canadian Photonics Fabrication Centre supports compound-semiconductor and photonics R&D. At IBM's Bromont site, Ottawa announced up to C$210M toward a C$662M project with C2MI to expand semiconductor packaging — specialized steps, not substitutes for leading-edge fabrication; judge by installed equipment and demonstrated throughput.

Critical minerals

Canada's Critical Minerals Strategy identifies lithium, graphite, nickel and cobalt. Reserves alone do not guarantee leverage — refining, processing, contracts and environmental standards matter too. Minerals plus clean electricity are negotiating assets, not completed capability.

The policy goal

Develop strengths in layers where Canada can plausibly compete: model development, photonics, packaging, safety evaluation, secure systems, specialized data services and industry applications — rather than promising a domestic replacement for every global supplier.

A national strategy built around choice

Launched in June 2026, AI for All presents Canada's direction through six pillars: protecting Canadians, building skills, expanding adoption and shared prosperity, developing sovereign foundations, scaling Canadian companies and deepening trusted partnerships. Its “build, partner, buy” logic is sensible: build where there is durable advantage or unacceptable risk in losing control; partner where scale is out of reach alone; buy where global suppliers fit best — while preserving alternatives and governance rights.

Each verb carries operational content. Build means reliable access, trained operators, power and network connections, refreshes and a viable operating model — more than constructing a facility. Partner means clear terms for access, data, continuity and exit. Buy means procurement that records dependencies, evaluation and the cost of leaving. The strategy also links sovereignty to adoption: infrastructure has little value if Canadian organizations cannot use it — Ottawa's first AI Missions Program directs funding toward health-care projects, and success will be measured by evaluated services and repeat customers, not announcements.

Three compute programs, three different jobs

Budget 2024 set aside C$2 billion over five years for the Canadian Sovereign AI Compute Strategy. It is a portfolio, not one national machine:

COMMERCIAL

AI Compute Challenge

Up to C$700M toward Canadian AI data-centre projects, including up to C$240M to Cohere's C$725M expansion — anchor tenant at a Cambridge, Ontario facility built and operated by U.S.-based CoreWeave (program, project summary).

ACCESS

AI Compute Access Fund

Helps Canadian SMEs pay for compute; in May 2026 the government said it had approved 44 projects worth C$66M. Allows Canadian and non-Canadian clouds — but a grant for access does not transfer operational control to the renter (program).

PUBLIC RESEARCH

Sovereign Compute Infrastructure Program

~C$890M over seven fiscal years from 2026–27; applications closed 1 June 2026. Defines sovereignty through Canadian location, governance, data residency and decision-making authority. Next milestones: signed awards, commissioned systems, actual access (program).

The Cohere–CoreWeave project illustrates managed interdependence: a Canadian model developer, a facility in Canada, an American operator. The concrete questions: Who administers the system? Who can access data in an emergency? Who controls updates? Can a Canadian customer move its workload if the relationship ends? Public answers would come from operating evidence — data-processing terms, subcontractor lists, admin-access diagrams, security assessments, audit rights, tested exit plans — not from the announcement.

Baselines, honestly stated: using November 2023 TOP500 listings, The Dais counted ~41 petaflops of Canadian compute (0.7% of the listed global total); an ISED evaluation found the Digital Research Alliance allocated 20% of requested GPU core-years in 2022–23. Those describe access at that time, not 2026 supply. Commercial plans extend this portfolio: Ottawa welcomed Bell's proposed Saskatchewan expansion (up to 900 MW within a 1.2 GW hub) and is advancing work with TELUS — prospective projects, tracked separately from facilities already operating.

Training is not the same job as inference

Frontier pre-training requires dense accelerator clusters, coordinated software, high-speed networking and capital few countries can sustain alone. Canada can pool pre-training with allies and negotiate access to resulting models — it does not need to train every global-scale model at home. Inference — the repeated, production use of trained models — is where sovereignty stakes are direct: where operational data goes, who can change a deployed model, and whether a service keeps running during an outage or export restriction.

The operational layer matters too: moving model weights means little if you remain tied to a provider's serving APIs, vector database, identity tools or MLOps platform. Open components like Kubernetes, vLLM and MLflow improve portability; Cohere's Command A+ under Apache 2.0 can run on open inference frameworks (NVIDIA H100 or Blackwell minimum configurations). But open weights are a portability feature, not a security approval — buyers still check provenance, licensing, vulnerabilities and test against their actual threats. Portability should be demonstrated with a real workload: data export, weights, prompts, identity and secrets, logs, vector indexes, dependencies, performance under load and recovery time.

Cloud sovereignty has legal and operational layers

Under 18 U.S.C. § 2713 (CLOUD Act), a covered U.S. provider must disclose records within its possession, custody or control pursuant to lawful process, regardless of storage location; a contract cannot cancel that obligation. FISA Section 702 is a separate authority — Congress let it lapse on 12 June 2026, with transition rules preserving existing orders. Encryption and customer-held keys reduce operational exposure but create no categorical legal exemption; it is useful to distinguish a domestic juridical arrangement from contractual sovereignty, and match the required form of control to workload sensitivity — the Munk School's Sovereign by Design risk tiers.

Domestic capacity exists

Shared Services Canada's public-cloud framework lists eight providers — seven U.S.-based, with ThinkOn the Canadian provider certified for federal Protected B workloads. TELUS opened its Sovereign AI Factory in Rimouski in September 2025 (NVIDIA and HPE equipment; supplier-chain dependencies remain). QScale's Q01 campus in Lévis reports 142 MW of renewable power. In October 2025 ThinkOn, Hypertec, Aptum and eStruxture announced a Sovereign Cloud Consortium.

Verification & networks

The federal cloud security profile accepts SOC 2, CSA STAR and ISO/IEC 27001 evidence; NIST's AI RMF can structure AI risk controls — neither establishes Canadian jurisdiction. Network paths matter too: ISED's resilience review notes much domestic Canadian traffic routes through the United States; local exchanges (CIRA's IXP overview) can keep some Canadian-to-Canadian traffic in Canada. Buyers of critical services should demand documented routes, peering and failover behaviour.

Hardware access is a geopolitical dependency

Canada does not design or fabricate the leading accelerators; dependence becomes a problem only when there is no workable alternative, no supply assurance and no plan for changes in prices, export rules or vendor support. U.S. policy shows the volatility: the AI Diffusion Rule (January 2025) was rescinded before its compliance date that May, but the EAR's advanced-computing, end-use and Entity List controls remain; May 2026 BIS guidance requires licences for specified items destined for entities headquartered in Country Group D:5 or Macau — a screening trigger when recipients have those ownership links, not a blanket restriction on Canadian public-sector workloads. Before committing capacity: know the chip classification, ownership chain, intended use, destination and transfer rules (EAR Part 744, counter-diversion guidance).

State capacity is part of the infrastructure

Sovereignty is also a delivery-capacity problem. The AI Strategy for the Federal Public Service 2025–2027 reports a 30% vacancy rate for digital roles — a government that cannot retain technical staff cannot verify supplier claims. Canada's research ecosystem (Vector, Mila, Amii) must connect to operational roles: people who run clusters, secure networks and assess vendors, within federal security-screening requirements. Process friction is real: SSC's evaluation found standard cloud brokerage fast, but requests needing additional enabling services averaged 278 days in its 2023 snapshot. Responsibilities span ISED, Treasury Board Secretariat, SSC, PSPC, departments and the provinces — one option is an empowered delivery and investment unit plus a formal federal–provincial–territorial forum for shared security tiers and pooled procurement.

Provincial choices shape the market

Ontario

The 2026 budget set out a provincial AI strategy and greater access to sovereign compute; an August consultation draft proposes assessing data centres on economic value, data security and sovereignty, community benefit and environmental impact — with projects bearing the full cost of electricity.

Alberta

Its data-centre strategy emphasizes attracting investment through power capacity, cooling and streamlined development while protecting grid affordability and reliability.

British Columbia

Its 2026 call allocates up to 300 MW for AI projects and 100 MW for other data centres over two years, and announced more than C$30M for research infrastructure at SFU's Cedar and UVic's Arbutus national host sites.

Different tools — research compute, private campuses, utility reservations, community criteria — none alone creates a national sovereign service. Together they define where power, funding and provincial demand exist. Québec's distinct private-sector privacy rules add another legal variation (below).

Public procurement, adoption and demand pooling

Statistics Canada reported 19.2% of Canadian businesses used AI to produce goods or deliver services in 2026, up seven points year-over-year. An IBM-commissioned September 2025 survey found 79% of full-time office workers using AI at work while only 25% relied on enterprise-grade tools — a shadow-AI gap that prohibition alone leaves invisible. The Buy Canadian Procurement Policy Framework (effective 16 December 2025) creates an opening for domestic capabilities through public purchasing. Demand pooling — inventories of repeatable workloads, common security tiers, portability tests — can make domestic services viable for small agencies and provinces, while respecting that federal Protected B does not automatically map to provincial health-data tiers. There is no universal 15–20% sovereign-cloud premium: compare full lifecycle costs against each workload's sensitivity.

Trade rules and policy durability

CUSMA's Chapter 19 (Digital Trade) expressly excludes government procurement and, with a limited exception, government-held information; covered procurements fall under Chapter 13 and its schedules, and the Chapter 32 National Security Exception is not a general shortcut. The CUSMA joint review took place 1 July 2026 — the government says the agreement remains in force through 2036 — while Canada repealed its Digital Services Tax on 26 March 2026, showing digital-policy instruments can be exposed to trade pressure. Durability comes from tying requirements to published security, continuity and public-interest objectives.

Data authority includes language and community

Language

A 2025 benchmark study found results on common benchmarks did not reliably predict performance on Québec French; the CEREALES dataset adds research depth. Evaluate models with the language varieties they will actually serve, and publish error rates. The Translation Bureau's GCtranslate is being deployed for content up to Protected B; CanLII generates legal summaries in both official languages — useful applications that still need review routes and safeguards.

Indigenous governance & Québec

Indigenous data requires its own governance — OCAP® (ownership, control, access, possession) — not a generic residency rule; Canadian jurisdiction and Indigenous data sovereignty overlap but one does not deliver the other. Under Québec's private-sector privacy law (Law 25 overview), transferring personal information outside Québec requires a privacy impact assessment and written agreement, and exclusively automated decisions carry notice and review rights. The federal Bill C-27 died with that Parliament; new privacy legislation arrived in June 2026. Buyers should define concrete duties per service rather than wait for one comprehensive AI statute.

Data centres also affect the places that host them: the federal Responsible Data Centre Development Principles call for local benefits, ratepayer protection and transparency — credibility depends on how they show up in provincial and municipal approvals.

Partnerships should widen Canada's options

The answer is not cutting off the United States, which provides investment, hardware, cloud and a large market — it is avoiding any essential service depending on one provider, one border crossing or one policy assumption. Canada and Germany launched a Sovereign Technology Alliance in February 2026 and in September invested in Montréal-based LawZero. The planned Cohere–Aleph Alpha combination — definitive agreement signed September 2026, still subject to regulatory approvals as of 23 October — would create a transatlantic supplier with dual headquarters in Toronto and Berlin, and means Cohere should no longer be treated as a simple proxy for an exclusively Canadian model layer. Good agreements specify who can use capacity, where data moves, who owns resulting research, how incidents are handled and what happens if priorities change.

Applying the five-part test to Canada

  1. Control — who governs the workload across its lifecycle?Name the legal provider, operators, administrators and data custodians; test who can access or change the system, which laws apply, who holds keys and logs, and who can suspend service.
  2. Continuity — what still works during disruption?Rehearse provider outages, network failures, export restrictions, power interruptions and supplier exit; measure recovery time and degraded service.
  3. Substitutability — can the system move in practice?Run migration tests covering data, model weights, MLOps tools, identity, vector indexes, logs and code; record time, cost and performance change.
  4. Data authority — who decides how data and outputs are used?Document legal authority, consent, purpose limits, retention, appeal and benefit-sharing; test English and French performance; follow OCAP® for Indigenous-led work.
  5. Economic viability — can the service be sustained, and does value remain?Calculate full lifecycle cost including exit; publish who can access shared compute; track utilization, Canadian technical jobs, research access and IP.
Two cross-cutting checks apply to every answer: independent evaluation and red-teaming before and during deployment, and linguistic/community adequacy tested with the people a service will affect. Funding, a facility, a Canadian model or an export clause can create potential — sovereignty depends on usable access, continuity, legal authority, alternatives and durable value.

Build a Canadian option, not a Canadian bubble

Canada's opportunity is to connect strong research, capable firms, clean power, public demand and trusted partners into systems with room to manoeuvre — with trained people, practical procurement, clear legal models, local network routes, realistic schedules, community trust and repeated evidence that services can be operated and moved. The country does not need to reproduce every layer of the global AI industry; it needs alternatives at the points where a single dependency could leave Canadians without control of an essential workload. A public information tool needs different safeguards from a health system, a classified service or an Indigenous language archive.

The measure: researchers can get compute, companies can build durable products, governments can govern their suppliers, communities can shape projects that affect them, and essential services keep working when a provider or geopolitical condition changes. Canada does not need to become self-sufficient in AI. It needs to be difficult to corner.


Further reading