D-Central analysis / August 22, 2026 / Canada

CUSMA is still in force. Canadian AI still needs a Plan B.

The agreement has not vanished, expired, or been automatically voided by the tariff dispute. But Canadian businesses have learned that a legal framework is not the same thing as guaranteed operational continuity.

See the Canadian inference pathReview private AI options

What the agreement actually says

CUSMA entered into force on July 1, 2020. Article 34.7 created a joint review on the sixth anniversary. Global Affairs Canada explicitly describes that review as a check-in, not an expiry date, and says the agreement remains in force until 2036. If all three parties agree to extend it, a new 16-year term is added. If they do not all confirm an extension, annual reviews continue through the remainder of the existing term.

Article 34.6 separately provides a withdrawal mechanism. A party may provide written notice, and withdrawal takes effect six months later. The agreement remains in force for the remaining parties. That legal text matters because “CUSMA has been voided” is not a forceful simplification. It is wrong.

Canada has, however, challenged U.S. measures under CUSMA. Global Affairs Canada’s active dispute page records Canadian requests for consultations over 2025 duties on Canadian goods, steel and aluminum, and automobiles. Canada says those measures appear inconsistent with specified U.S. obligations. CUSMA provides consultations and independent panels; if a violation is found and not removed, the prevailing party can suspend benefits of equivalent effect.

That is what a rules-based agreement looks like under stress: the agreement continues, governments disagree about compliance, and formal remedies exist. It is inaccurate to say there is no law. It is equally naive to pretend the dispute has no effect on business confidence while a case works through negotiation or adjudication.

The planning assumption has changed

On August 21, the Prime Minister said Canada had suspended negotiations after last-minute U.S. terms were judged unfair and uneconomic, and that the United States intended to impose a 50% tariff on roughly C$28 billion in Canadian goods at midnight. On August 22, he argued that American trade commitments had become less dependable. Those remarks are the Canadian government’s position, not a court judgment about the entire U.S. legal system.

For an enterprise risk register, the careful wording is “reduced policy predictability” or “policy-continuity risk.” Canadian companies should not declare that U.S. rule of law has ceased to exist. They can observe that tariffs, exemptions, deadlines, and negotiating terms have changed quickly enough to make sole dependence on a U.S.-controlled route harder to justify.

CUSMA remains a valuable legal framework. It should not be confused with a service-level agreement for every cross-border dependency.

Why an AI Plan B belongs in this discussion

There is no official evidence that the current goods tariffs impose a general tariff on AI inference APIs. They do not establish an AI tariff. CUSMA Chapter 19 generally prohibits customs duties on digital products transmitted electronically. The AI risk is not today’s customs rate. It is that Canadian businesses increasingly rely on a small number of foreign platforms for an increasingly essential capability.

A provider can retire a model, change acceptable-use rules, modify output behaviour, adjust price, alter retention, restrict an account, or suffer a service disruption. Governments can impose export controls or other restrictions. Currency and procurement terms can move. A customer may have contractual rights, but a remedy after an outage does not keep a production process running.

OSFI’s third-party risk guidance offers a useful model even outside finance. It asks regulated institutions to consider geography, foreign legal requirements, cloud portability, concentration risk, and exit strategies. The principle is universal: critical services require a path out.

Plan B does not mean copying the frontier provider

The objective is not to recreate every feature of a giant public platform inside a server room. Start with the workloads whose risk and value justify an alternative. A document workflow may need reliable extraction, retrieval, summarization, and bilingual drafting rather than the world’s highest aggregate benchmark. A support classifier may need consistency, low latency, and controlled logs. An internal coding assistant may need repository privacy and predictable availability.

A real evaluation uses those tasks. It records acceptable quality, latency, throughput, context, tool behaviour, citations, failure modes, security boundaries, and operating cost. It then identifies a model and deployment design that meet the requirement. D-Central’s dated open-weight model comparison can support the shortlist, but a leaderboard cannot replace an organization’s own test set.

Five requirements for a credible exit

1. Application portability. Separate business logic from provider-specific prompts, tools, response formats, and identity mechanisms where practical.
2. Data portability. Keep source documents, retrieval indexes, evaluation cases, configuration, and logs in formats the organization can export and rebuild.
3. Model choice. Qualify at least one replacement model for each continuity-critical workflow and review its licence for the intended use.
4. Infrastructure choice. Decide whether the alternative belongs on a Canadian-operated service, dedicated Canadian infrastructure, or the customer’s own premises.
5. An exercised cutover. Test authentication, routing, rollback, recovery time, deletion, and ownership. A clause that has never been exercised is not an operational plan.

Canadian-hosted inference can be the middle path for organizations that want Canadian operation without managing the entire stack. On-premises inference provides a tighter customer-controlled boundary but transfers patching, power, cooling, security, and recovery duties to the organization. A hybrid can preserve a frontier API while giving sensitive or essential work another route.

Use precise sovereignty language

“Hosted in Canada” describes geography. Canada’s AI Sovereign Compute Infrastructure Program uses a stronger definition: Canadian-located and Canadian-governed infrastructure in which data residency, operational control, and decision-making authority remain in Canada. The federal public-cloud white paper also warns that a foreign-controlled provider may remain subject to foreign legal compulsion even when data is physically in Canada.

That does not make every U.S. provider unacceptable. It means a buyer should document the remaining dependency. Who owns the service? Who administers it? Which subprocessors can receive data? Who controls keys? What telemetry leaves the environment? What can be exported? Which law governs the operator? Sovereignty is not achieved by changing a billing address.

Frequently asked questions

Did the United States void CUSMA by imposing tariffs?

No. CUSMA remains legally in force. Canada alleges that specific measures are inconsistent with U.S. obligations and has used the agreement’s dispute process. Breach allegations and treaty termination are different legal questions.

Does the 2026 review mean CUSMA expired?

No. Global Affairs Canada says the joint review is not an expiry date. The current agreement remains in force until 2036 unless a party completes the separate withdrawal process.

Can we say U.S. rule of law is unreliable?

As a legal conclusion about the entire American system, that is too sweeping. As a dependency decision, D-Central’s position is direct: today’s administration has made U.S. market-access commitments too unpredictable to serve as the only foundation for vital Canadian infrastructure. Canadian businesses do not need to wait for a final tribunal decision before building a tested alternative.

Primary sources reviewed August 24, 2026: Global Affairs Canada, CUSMA Joint Review; CUSMA Chapter 34; Active CUSMA disputes; CUSMA state-to-state dispute settlement; Prime Minister of Canada, August 21 statement; OSFI Guideline B-10.