AI access can look simple: choose a plan, open a chat, send a request. The infrastructure underneath is not simple. It is a metered service shaped by product tiers, safety controls, abuse prevention, demand patterns, network conditions, scheduled maintenance, and the physical limits of data centres.

Recent developments make this visible. OpenAI stated on September 10, 2026 that it had temporarily paused new sign-ups and upgrades to its Pro $200 plan while keeping existing subscribers and its Pro $100 tier available. Its API documentation separately describes rate limits at organisation and project levels, overload responses, ramp-rate controls, and reserved-capacity products. Anthropic’s public status record also shows that service incidents remain an operational fact.

These developments do not prove a general AI-capacity crisis. A plan pause can reflect a product decision. A rate limit can protect fair access, security, cost control, or system stability. An incident can be short-lived.

They do establish a more useful principle: essential capability cannot rest on a single login, a single model, a single cloud region, or an assumption that today’s terms will be unchanged tomorrow.

Build the capacity to choose

Canada2080 defines economic sovereignty as the ability to create, own, govern, finance, deploy, repair, and renew capability. In AI, that does not mean Canada must own every model or reject trusted international exchange. It means Canadian institutions can make choices under pressure.

Can an operator move a workflow when an endpoint is constrained? Can a public institution keep a critical service running with a less capable but adequate model? Can it explain who controls the data, encryption keys, model version, budget, and incident response? Can it restore service without waiting for an opaque external decision?

The first response is not to build a monument to compute. It is to understand the workload.

Some work genuinely benefits from a frontier hosted model. Some can run on a smaller specialised model. Some needs no generative model at all. Some sensitive or continuity-critical work may deserve a locally operated path, but only when the institution can maintain it responsibly.

The model landscape is widening. OpenAI’s gpt-oss, Mistral Small, Google’s Gemma family, and other distributions show that more models can now be run outside the issuer’s own hosted service. This increases choice. It does not end dependency.

Weights, licence, hosting, hardware, maintenance, energy, data, and decision rights remain separate questions. Meta’s Llama 4, for example, is distributed under a conditional community licence rather than a licence-free public arrangement. “Open” is not a synonym for sovereign.

The continuity question

Canada needs no panic, no vendor romance, and no false promise of self-sufficiency. It needs the ability to choose, substitute, measure, and explain.

Every public AI deployment should identify its continuity mode before it identifies its preferred model. Every institution should know which workloads require frontier capacity, which can run locally or regionally, what happens during a service interruption, and who is accountable for the fallback.

That is how access becomes capability: not by assuming a service will always be there, but by designing for the moment it is not.