Canadian AI sovereignty: follow the workload, not the map

A Canadian address answers only one of the questions

A data centre can sit in Canada while an organization still depends on a foreign supplier for software, account access or essential hardware. Conversely, a Canadian organization can gain useful capacity through a partner without controlling every layer itself. “Sovereign AI” becomes meaningful only when the dependency being discussed is named.

Canada’s AI for All strategy treats sovereignty as a foundation involving compute, data and talent. It describes a build, partner and buy approach rather than total self-sufficiency. Our reading focuses on what a researcher, business or public-service buyer would need to verify before treating a sovereignty claim as useful evidence. This is a purchasing and accountability framework, not a certification of any provider.

Separate the five kinds of control

Ask for a separate answer to each row. A provider may meet one requirement and leave another unresolved. That is more informative than awarding a single “Canadian” badge. The documents you need also differ: a facility address cannot establish contractual rights, and a contract cannot show that capacity is available on the day you need it.

Scroll the table sideways to see every column.

Read the infrastructure promises as milestones

The strategy describes a public supercomputer as a goal by 2031 and an expansion of sovereign compute and cloud infrastructure. Its wording also distinguishes partnerships being finalized from capacity proposed for later years. An announcement, a construction commitment and usable compute are different milestones. We have not independently verified the delivery of the announced infrastructure.

For each project you follow, keep four dates: announcement, agreement, operational opening and first service available to your user group. Leave a date blank until there is evidence. This avoids converting “planned” into “available” as the same announcement passes through news stories and presentations.

A published megawatt figure describes a kind of infrastructure scale. It does not tell a small research team how many accelerators it can book, which software is supported or whether its application is eligible. Ask for the service catalogue and allocation process before translating an infrastructure headline into practical access.

A research team and a small business need different answers

Consider two hypothetical users. A university group wants to train a model over several weeks; a small manufacturer wants a reliable daily inference service. The first may care most about a large temporary allocation and the ability to move checkpoints. The second may care more about response time, support and predictable operating cost.

Neither need is resolved by the phrase “domestic capacity.” For the university group, request scheduling limits and storage-transfer conditions. For the manufacturer, request a service agreement, supported deployment options and a recovery procedure. If an access program subsidizes one workload but excludes the other, report that boundary rather than calling the program universally available.

This distinction also changes how to judge success. A facility opening is an infrastructure result. Reduced waiting time for eligible researchers is an access result. A reliable service at an affordable cost is an operating result. Keeping them separate makes progress reports harder to inflate.

AI Compute Access Fund cost planSeparate a capacity announcement from a funding assumption in your budget.

The rest of the strategy still matters

The six pillars connect protection, skills, adoption, sovereign foundations, Canadian companies and international partnerships. Compute alone cannot establish whether workers receive useful training or whether public services improve. Our analysis is that infrastructure reporting should be joined to evidence about the people expected to use it.

For a training commitment, ask who finished and what they could do afterward. For business adoption, ask what task changed and whether the claimed improvement includes correction and integration costs. For a public-service deployment, ask how errors are challenged. These are proposed measures of usefulness, not outcomes we claim the strategy has achieved.

Use a claim log instead of a sovereignty ranking

When comparing announcements, record the exact claim, its owner, the supporting document and the missing proof. Do not rank suppliers from branding or infer legal protection from geography alone. The relevant legal and security review depends on the workload and organization.

The next useful story is often the unanswered row: who can use the new capacity, on what terms, and when? That question gives a Canadian reader something more concrete than another list of investment announcements.

Claim: __. Workload and intended users: __. Canadian processing locations: __. Operational control: __. Eligibility and capacity evidence: __. Delivery milestone actually reached: __. Exit test or unresolved dependency: __. Last source check: __.