The Mythos moment showed the dependency plainly: access to frontier AI can be narrowed by nationality, so Canada needs enough domestic data-centre and compute capacity for Canadian users, builders, companies, and public institutions to keep moving when foreign access rules change.
AI sovereignty begins with an operational question. Who can keep working if access changes?
On June 13, 2026, public reporting said Anthropic disabled access to Fable 5 and Mythos 5 after a U.S. government directive restricted foreign use on national-security grounds. Some reports described the order as applying to foreign nationals regardless of location, including foreign employees inside the company.
For Canada, the important point is the power being exercised. A foreign government can draw a nationality line around frontier AI capability. That can reach consumer accounts, enterprise seats, API access, research workflows, internal tools, and products built by Canadian companies.
Most workflows will survive ordinary vendor switching. An app can move from one model endpoint to another. The deeper issue is reliable access to the frontier capability layer itself, especially if Canada also depends on foreign data centres to run the alternatives.
Canada does not need to train the leading frontier model to answer this. It needs enough domestic compute to keep options alive. Canadian model companies should be supported where they can win. Open-weight models should be used where they are good enough. Allied frontier systems should remain part of the stack. The missing piece is capacity Canada can power, site, govern, and procure against.
Canada should own its AI infrastructure layer: powered data centres, domestic inference capacity, operating control, model optionality, and real demand from institutions that need continuity under Canadian law.
Canada can matter by owning enough of the stack that Canadian users, builders, researchers, companies, and public institutions keep moving when foreign access rules change.
I. Why this matters to Canadians
The Mythos episode matters because it put a nationality line on something Canadians have started to treat as always-available infrastructure.
Canadian users and builders quietly assume the frontier models inside their tools will still be there tomorrow. The developer expects the coding assistant to keep improving. The startup expects API access to remain available. The researcher expects the best model to be reachable. Teams inside banks, hospitals, universities, telecoms, utilities, and governments expect pilots and internal tools to keep moving from test into operations.
Those assumptions are reasonable until the capability becomes strategic.
If a U.S. government directive can restrict foreign access to a frontier model, the effect can move through every layer built on top of it: consumer accounts, enterprise subscriptions, API products, research programs, internal tools, and software built by Canadian companies.
Some of those systems can switch models. Many will. A model endpoint is replaceable in a way a power plant or fibre route is not.
The dependency is still real. The best models shape what builders can build, how fast researchers can work, what products Canadian firms can sell, and whether Canadian institutions can adopt AI on equal terms. Falling one model generation behind may not break a service overnight. It can quietly compound across thousands of workflows.
Canada does not need to train the leading frontier model to answer this. It needs enough domestic compute to keep options alive. Canadian model companies should be supported where they can win. Open-weight models should be used where they are good enough. Allied frontier systems should remain part of the stack. The missing piece is capacity Canada can power, site, govern, and procure against.
That is the infrastructure question. Can Canadian users, builders, companies, researchers, and public institutions keep working if access to someone else's frontier model changes? Can sensitive workloads run on Canadian-controlled compute when the answer needs to be yes? Can Canada turn power, land, cooling, fibre, capital, and procurement into useful capacity quickly enough?
The answer should be yes. Canada has the ingredients. The work is to get building.
II. The model race and the infrastructure race
Canada should stay ambitious about models. Canadian IP matters. Canadian research institutions matter. Domestic technical talent matters. Canadian companies should be supported where they have a real path to advantage.
But the leading frontier-training race is now a scale race. It compounds capital, chip supply, research talent, data pipelines, safety infrastructure, distribution, and revenue. Canada can contribute to that race. It should not make the whole sovereignty strategy depend on winning it.
The more realistic Canadian wedge is controlled deployment: specialized systems, domain-tuned models, secure inference, evaluation, routing, model operations, and open-weight deployment for users that need reliability under Canadian control.
Open-weight models matter because they turn compute into optionality. With enough domestic compute, Canada can run capable models inside Canadian-controlled environments, tune them for specific workloads, and avoid total dependence on one foreign frontier endpoint. Without enough compute, model optionality is mostly theoretical.
This is why the infrastructure race matters. Once AI is embedded in public services, hospitals, banks, telecom networks, energy systems, courts, universities, defence-adjacent work, and industrial processes, the question is not just which model scores highest. The question is which system can be procured, audited, governed, routed, updated, secured, and kept online.
Canada can run that race.
It can build powered data centres. It can create domestic inference capacity. It can give regulated users approved routes across allied frontier models, Canadian systems, and open-weight alternatives.
Canada's opportunity is the infrastructure layer that makes model optionality real.
III. Canada's infrastructure advantage
Canada already has the pieces of the layer that matters.
Power. Cold climate. Water discipline. Fibre. Industrial land. Engineering capacity. Public infrastructure institutions. Pension capital. Regulated buyers with real reasons to care about continuity.
Those pieces do not automatically become an AI infrastructure strategy. They have to be assembled into working capacity.
Every Canadian site still has to earn the claim. Canada is not one power market. Ontario is different from Quebec. Alberta is different from British Columbia. The Maritimes are different from Manitoba. Each province has its own grid constraints, tariff politics, interconnection timelines, permitting rules, Indigenous and municipal context, cooling profile, water system, fibre position, and industrial land base.
That specificity is an advantage if Canada does the work.
The best Canadian data-centre story is a portfolio of sites that can deliver powered, cooled, connected, permitted, financeable compute on timelines that matter.
That means getting practical. Which sites can actually receive load? Which utilities can serve it? Which projects have cooling and water plans that survive public scrutiny? Which campuses have fibre, tenants, procurement routes, and operators? Which facilities can support Canadian-controlled inference rather than becoming impressive buildings attached to someone else's control plane?
This is a positive agenda. Canada does not need to talk itself into being an AI superpower. It needs to turn real infrastructure advantages into useful capacity.
Build the data centres. Secure the power. Govern the water and heat. Bring in the tenants. Give Canadian institutions somewhere credible to run the workloads that matter.
IV. What Canada should own
Canada does not need every AI use case to run on Canadian-controlled compute. The target is narrower: public institutions, regulated sectors, critical infrastructure, defence, defence-adjacent industry, sensitive research, and builders whose products depend on reliable frontier access.
Those are the workloads where continuity, auditability, data handling, and operating control matter.
The first layer is physical compute. Powered data-centre capacity in Canada is the base. Real sites. Real interconnection. Real cooling. Real water strategy. Real fibre. Real operations. Real delivery timelines. Without physical compute, the rest is policy language.
The second layer is operating control. Location matters, but it is not enough. A Canadian building can still depend on foreign control of the scheduler, identity system, logs, firmware updates, incident response, administrative keys, and model access. The question is whether the workload can keep running under Canadian authority when the external layer changes.
The third layer is model optionality. Canada should keep access to allied frontier systems. It should also support Canadian models, open-weight models, domain-tuned systems, and routing architectures that let approved workloads move between model families where feasible. Open-weight models are part of the answer, but only if Canada has the compute to deploy them at useful scale.
The fourth layer is demand. Government, health, finance, telecom, defence, critical infrastructure, and sensitive research users need procurement paths that turn domestic compute into actual utilization. Capacity without buyers is idle optionality.
The fifth layer is capital. Canada has the capital pools to own infrastructure if the projects are structured like infrastructure: long-duration contracts, credible tenants, financeable assets, and clear risk allocation. The durable layer is power, data centres, fibre, cooling, and regulated compute capacity.
This is what owning the AI infrastructure layer means: enough physical, operational, model, demand, and capital control that the most important Canadian workloads can keep running.
Conclusion
The Mythos moment should make the opportunity clearer.
The United States is going to treat frontier AI as a strategic capability. That is rational. Canada should keep buying from allies, keep partnering with U.S. labs and hyperscalers, and keep using the best systems available. For the workloads that matter most, it also needs infrastructure it can govern.
Canada's AI opportunity is to build the layer that advanced AI will need everywhere: power, data centres, cooling, fibre, operating control, model optionality, institutional demand, and long-duration capital.
That agenda turns Canada from a customer into a useful jurisdiction. It gives hyperscalers and labs a place to put capacity. It gives Canadian institutions a controlled inference path. It gives pension capital a domestic infrastructure asset class. It gives policymakers a sovereignty strategy concrete enough to build and narrow enough to finance.
Canada should keep using the best allied models available. It should also build enough domestic compute that Canadian institutions are not stuck waiting for someone else's access rules to hold.
The opportunity is practical: build the data centres, secure the power, finance the capacity, and give Canadian users somewhere credible to run the workloads that matter.