Who Owns Useful Intelligence? A Canadian Plan for Sovereign AI

Carney’s October 29 sovereign-AI announcement parallels the case I presented two weeks earlier: Canada needs the ability to shape large models at the frontier and carry their capabilities into local life.

Carney has been repeating the sovereign-AI message. On October 29, he posted about secure AI built in Canada, and his government announced up to C$1.6 billion in equity investment in Cohere. His post and the Cohere announcement put a national spotlight on a question I addressed in my October 15 presentation: who can build and refine the models Canadians will rely on?

In “AI After Peak Energy: From the data-centre bubble to local intelligence”, I argued that Canada needs the ability to build and refine large models at the frontier, then carry that capability into regional and local systems. My starting point was energy. As net energy tightens, every sector has to seek efficiency: the same useful result with less fuel, power, material, time, and waste. Coordinated software can search records, compare options, and help people plan the next step. People still choose and carry out the work.

That can mean a resident finding the rule behind a permit, a clerk searching years of meeting records, a student getting instruction far from a university, or a grower using local weather and soil information. AI can make knowledge easier to reach and use. The public should be able to understand who shaped that knowledge and what rules guide the system.

Security is a core reason to build our own models

Models carry the influence of their training records, safety rules, and deployment context. In my presentation, I argued that open models developed in China and cloud models developed in the United States can reflect different national and commercial settings. Canada has its own laws, histories, records, public institutions, and security needs. Our models should be shaped and evaluated with those realities in view.

Security tests I discussed showed that a model’s security behaviour changes depending on what country and institution it believes it is serving. In one test, DeepSeek-R1 produced vulnerable code in 19% of baseline cases and 27.2% in a political and industrial-control scenario; I explained that the model had been told it was operating in the United States. A separate Booz Allen test found more code flaws from Qwen under a U.S.-government persona. These results make model provenance and country-specific evaluation central security concerns. Canadian ownership gives us the ability to set Canadian requirements, use Canadian records in evaluation, and inspect how models behave in our systems.

Security also depends on how we use a model. While Canadian models are being developed and evaluated, independent model families can check one another. Software scanners, sandbox runs, and integration tests can provide repeatable checks. People should approve changes that affect sensitive systems, public rights, or major consequences. A model can draft and retrieve; accountable people verify and decide.

Build and refine models at every scale

The sovereignty plan I described reaches all the way to the frontier. Canada needs the ability to build and refine large, general-purpose models that shape the highest level of AI capability. The purpose of the largest data centres can be to make and improve models. Once trained, useful capabilities can reach people through smaller regional and local systems.

Regional labs can refine large models and distill versions for provincial law, history, and services. Local facilities can adapt models to municipal bylaws, records, and procedures. A larger teacher model can provide examples, corrections, and tests for a smaller model designed for a particular task. The result can run on local equipment and support work such as public information, records search, meeting transcription, education, health information, repairs, and agriculture.

These layers have different physical needs. My presentation mapped occasional frontier campaigns to power-rich Canadian corridors, regional labs to places with fibre and technical staff, and local refinement to existing buildings and available connections. It identified Grey County and Owen Sound as possible sites for local work, with hydroelectric corridors such as Québec City–Lévis, Winnipeg, and Prince George as potential locations for frontier-scale projects. Those locations and capacities are planning proposals, not approved facilities.

Put the infrastructure to useful work

Large-scale model development requires substantial electricity, storage, cooling, water, equipment, and capital. Grid connections and construction capacity set real limits on how quickly new facilities can operate. Canada can plan around those limits by matching each kind of work to the scale and site it needs.

The data-centre boom may also leave behind useful infrastructure. Railways and fibre networks continued to serve people after investment cycles changed ownership and lowered prices. A correction in AI infrastructure could similarly make mature models, server racks, fibre, and power connections available to new owners. Local communities could then own machines for useful work instead of relying on a distant provider for every task.

This connects the frontier and local parts of the plan. Canada needs to shape the large models; regional teams can adapt them; local organizations can put them to work. A capable model should help people make better use of their time, knowledge, tools, and energy. The result is practical intelligence closer to the people who need it.

Sources