Bell AI Fabric and Cohere: an evaluation note for regulated organizations

Since July 2025, Bell and Cohere have carried the same promise to Canadian enterprises and public bodies: AI operated in Canada, on hardware installed in Canada, under a Canadian contract. In June 2026, a four-party agreement put a number on it. Three years, roughly USD 220 million, 2,304 NVIDIA Grace Blackwell GPUs deployed at Merritt, British Columbia. The structure is a private contract rather than a federal grant, though it addresses the gap Ottawa's Sovereign AI Compute Strategy was written to close.

The schedule reads alongside the agreement. Merritt Phase 2, the Cohere, BUZZ HPC and Hypertec build, is expected in early 2027. The Winnipeg facility is expected in late 2026, the Saskatchewan campus at 300 MW in the first half of 2027. An organization buying today buys a contractual commitment and a roadmap. The Canadian stack being marketed to it is under construction.

The costlier gap sits elsewhere, in what the word sovereignty is taken to cover. Data residency establishes where the bytes sit. It establishes nothing about how the system's outputs are evaluated, or whether the organization can replay them for a board or a regulator. Organizations are being sold a location as though it were a control, and the organization that buys the location believes it has settled a governance question.

In short: Bell AI Fabric and Cohere deliver a real improvement in operational and contractual control: a data centre at Merritt, a Canadian contractual counterparty, reduced dependence on American hyperscalers. The stack is not a legal firewall: under the national contacts doctrine American courts assess jurisdiction against a company's aggregate ties to the United States, and BCE is listed on the NYSE. As of August 24, 2026, Cohere is in production across Canadian public bodies, and almost none of that production runs on the Merritt stack, whose Phase 2 is expected in early 2027. For a regulated organization already inside Microsoft 365, the decision is settled by tiering use cases against consequence and data class, then assigning one platform per tier. Residency says where the bytes sit; evidence of evaluation says whether the outputs replay.

What is in the Bell AI Fabric and Cohere stack?

Four parties, four layers, sold as a single offer. Reading it layer by layer is the first act of evaluation: each layer carries its own owner, its own jurisdiction and its own maturity, and the single offer conceals all three gaps.

Two announcements structure the offer. In July 2025, Bell and Cohere signed a strategic partnership: Cohere's enterprise models and its North agentic platform became available through Bell AI Fabric, and Bell became Cohere's preferred infrastructure partner for sovereign deployments in Canada. In June 2026, the four-party agreement operationalized it with the contract and the GPUs. The layer the client actually owns is the fourth, the one it configures itself, and that is also where permission errors accumulate.

What is operational today, and what is still under construction?

A distinction worth carrying into any client conversation: Cohere is in production across Canadian public bodies today, and almost none of that production runs on the Merritt stack.

Live as of August 24, 2026:

• Shared Services Canada licensed Cohere from November 2025 for CAD 339,000. Command A is integrated into CanChat, with roughly 11,500 federal employees across 20 departments enrolled and a wider rollout in progress.
• ISED is deploying North, positioned by Cohere as a template for federal modernization.
• Quebec signed a partnership with Cohere in June 2026 to examine public-sector integration.
• The University of Toronto signed a multi-year agreement in July 2026, with North as the orchestration layer.
• Bell is its own reference customer, with over 100 identified internal use cases.
• Earlier pilots include RBC, Dell, LG CNS, Ensemble Health Partners and Palantir.

Not yet available: Merritt Phase 2, the Cohere, BUZZ HPC and Hypertec build, expected in early 2027; the Winnipeg facility, expected in late 2026; the Saskatchewan campus at 300 MW, expected in the first half of 2027.

That reading changes the nature of the decision. Buying Cohere today means buying models and an orchestration platform that work, on infrastructure that is not yet the one being sold. The two purchases document differently, and a compliance file that merges them describes infrastructure that does not exist.

What does the sovereignty claim cover once examined?

Canadian legal commentary through 2026 converges on a single formulation: sovereignty is a question of control, determined by legal jurisdiction and corporate structure, and residency alone does not establish it. Under the national contacts doctrine, American courts assess jurisdiction against a company's aggregate ties to the United States. A New York Stock Exchange listing, American institutional investors, or an American subsidiary can be sufficient.

Applied consistently, that test produces three observations that rarely appear in the sales narrative.

• Microsoft. Canadian data residency improves the compliance posture and does not create a legal barrier against extraterritorial requests. Microsoft's own documentation states this. Customer-managed keys, client-side encryption and confidential computing narrow the exposure; they do not close it.
• Bell. BCE is named in Canadian sovereignty analysis alongside Rogers and TELUS as maintaining extensive American connections. BCE is listed on the NYSE. The doctrine that reaches Microsoft reaches Bell.
• Cohere. Following the April 2026 merger with Aleph Alpha, the company is headquartered in both Canada and Germany. The purely Canadian framing requires qualification.

None of this disqualifies the stack. It relocates the argument. The claim that withstands examination is improved operational control, a Canadian contractual counterparty, and reduced dependence on American hyperscalers. The claim that data becomes legally unreachable does not survive the analysis, and an organization that relies on it in a Loi 25 transfer assessment is exposed. That is precisely what the Sovereign property of the iDIA framework measures: actual command of control, scored on evidence, rather than declared location.

Microsoft 365 Copilot or Cohere North: where does the gap sit?

Three conditions earn Cohere its cost, and sovereignty in the marketing sense is not among them. First: the corpus sits outside Microsoft. Graph is Copilot's structural advantage, and it does not reach Epic, SAP, a PLM environment or a claims platform. Second: a regulator requires the model to be documented, and a validation file cannot be written against a black box that changes without notice. Third: the deployment envelope excludes the public internet, and air-gapped operation is something Copilot and Vertex do not offer.

Where Cohere adds nothing. An organization fully inside Microsoft 365, handling ordinarily confidential rather than regulated data, with no regulator demanding model documentation, buys a second invoice, a second governance surface, and a connector project that reproduces retrieval it already has.

What options remain for a regulated organization already inside Microsoft?

This profile removes the air-gap argument, which is Cohere's strongest. The remaining gap in a regulated Microsoft shop concerns version control and audit evidence, and that gap closes without leaving Microsoft.

• Azure AI Foundry, in the organization's own tenant. Data remains in the tenant; the model provider sees neither prompts nor responses. Version pinning, evaluation gates before production, drift detection, blue/green rollback and an immutable audit trail are platform capabilities. Cohere Command R+ sits in the Foundry catalogue alongside Llama, Mistral and Phi, which means Cohere models can run without adopting Cohere as a platform.
• AWS Bedrock or Google Vertex. The same pattern in a different cloud. Relevant primarily where concentration on a single provider is itself a supervisory concern.
• IBM watsonx. Heavier and slower to stand up, built for regulated deployment, with model documentation and governance tooling included rather than assembled.
• Mistral. European sovereign positioning with private deployment. Relevant where European exposure is material.
• Self-hosted open weights. Maximum control, and the organization assumes the full model risk and operational burden. Viable only with genuine internal capacity.
• Cohere direct, through Bell, or through SAP. SAP Canada is integrating North into its sovereign ERP cloud, which is the shortest path for organizations already on SAP.

How should use cases be sorted into tiers?

Selecting a vendor per use case produces platform sprawl: several contracts, several governance surfaces, and no consolidated answer to the question of which AI systems the organization runs. Sprawl costs more than licences. Use cases sort into tiers, and a platform is assigned per tier.

The four sorting questions. Does the output reach a person directly, or does a human review it first? What class of data does it touch? Will a regulator ever require justification of a specific output? How frequently does the underlying model change, and would the organization detect the change? Importance does not enter the sort. Consequence and data class do.

identifiable · AI Insights

Tier drift. Tiers move upward without a decision being taken. A summarization pilot becomes the artifact people act on. Classification therefore requires a review trigger tied to change in use, not an annual calendar date. That is the Supervised property of the iDIA framework, and the reason an inventory of AI agents is held continuously rather than once a year.

How does the tiering read in health, financial services and defence?

Health. The clinical record sits in Epic, Oracle Health, Meditech and PACS, outside Graph. Copilot's structural advantage stops at the clinical system boundary, and what remains is administrative productivity. Cohere's realistic position is documentation and revenue cycle inside the institution's perimeter. Neither vendor resolves the harder question: anything approaching clinical decision support raises a Health Canada software-as-medical-device question and a high-risk classification question, and no procurement decision settles it.

Financial services. The differentiator is validation. Supervisory expectations around model inventory, documented validation and third-party concentration make a versionless assistant difficult to sustain before a chief risk officer. A pinned, dedicated deployment can be frozen, tested and retested, which is the argument a chief risk officer accepts. Cohere co-developed banking functionality with RBC. The expectation grid of the NIST AI RMF and that of ISO/IEC 42001 describe the same requirement for evidence.

Aerospace and defence. The least ambiguous case. Controlled Goods Program obligations, ITAR and EAR, and classified networks make Copilot structurally ineligible rather than merely weaker. Cohere has partnered with Hanwha, Saab, Thales and TKMS. For context on the scale of the existing exposure: the Department of National Defence's Defence 365 platform runs on American-headquartered providers.

What sequence comes before choosing a vendor?

• Inventory the AI use cases in operation, including those adopted without approval.
• Sort each use case against the four questions.
• Name the specific regulatory obligation attached to each tier. Loi 25 requires a documented transfer assessment. Prudential supervision requires validation files. Health and defence carry separate regimes.
• Assign one platform per tier.
• Establish the evaluation baseline before go-live: a fixed test set with expected outputs, so that drift is measured rather than reported anecdotally.
• Set the review trigger on change of use.

Steps one through three constitute the engagement. Most organizations have never classified their AI use cases against consequence, which produces one of two findings: Tier 3 procurement applied to Tier 1 work, or Tier 3 work running on Tier 1 tooling. The AI Index diagnostic situates the organization on its six properties before that sorting begins, and securing the generative AI pipeline installs the evidence along the way.

What would change this reading?

• Merritt Phase 2 entering service on schedule in early 2027, with published performance and pricing.
• A Canada, United States bilateral agreement under the CLOUD Act framework, which as of early 2026 had no completion timeline.
• Federal or provincial procurement rules that make Canadian-controlled infrastructure a condition rather than a preference.
• Microsoft extending a genuinely disconnected deployment mode for Copilot.
• Further consolidation of Cohere's corporate structure outside Canada.

Data residency establishes where the bytes sit. It establishes nothing about whether the system's outputs are evaluated, reproducible and traceable. That distinction is the work.