Canada's AI Access Problem: Public Options, Fallback Capacity, and Sovereignty Under Model-Access Risk

 Advanced AI systems are increasingly becoming part of the operating infrastructure of governments, universities, firms, and regulated sectors. Yet access to these systems is not the same as ownership or control. Organizations may retain staff, data, software, cloud contracts, and compute resources while still losing operational capability if access to a critical model layer becomes restricted, degraded, repriced, or withdrawn.

This working paper examines what governments can realistically do once model access is recognized as a form of infrastructure dependence. It argues that Canada cannot guarantee permanent access to externally controlled frontier AI models. However, it does not follow that Canada is powerless. The practical policy challenge is not to eliminate dependence, but to govern it before it becomes invisible, brittle, and expensive to unwind.

The paper develops the concept of model-access resilience and argues that governments possess a range of instruments that can reduce vulnerability without claiming complete technological sovereignty. These include dependency visibility, procurement conditions, interoperability requirements, multi-provider resilience, continuity planning, sovereign-compute support, public fallback capacity, critical-sector protections, and strategic bargaining power through public spending.

Drawing on Canadian AI policy, sovereign-compute initiatives, digital-sovereignty discussions, procurement mechanisms, and operational-resilience frameworks, the paper distinguishes between what governments can control, what they can influence, what they can negotiate, what they can test, and what they can only monitor. The central claim is that Canada has instruments, not guarantees.

The contribution of the paper is to move the discussion of AI sovereignty beyond a simple choice between complete dependence and complete self-sufficiency. Instead, it proposes a practical policy architecture for managing model-access risk under conditions of deep uncertainty.

This Zenodo research object includes the publication manuscript, rendered PDF, source and claim registers, review materials, transcript supplements, metadata files, manifest hashes, and provenance materials documenting the development of the paper.