Cohere Brings Model Vault to Canada for Enterprises Needing Data Sovereignty
Cohere's single-tenant inference platform now runs on Canadian infrastructure, giving domestic enterprises a way to keep model traffic inside national borders.
- Cohere's Model Vault is now available in Canada, keeping inference traffic inside the country.
- Vault is a single-tenant, fully managed inference platform for Cohere models with auto-scaling and no rate limits.
- Zero Data Retention policy means prompts and outputs are not stored after processing.
- Recent confidential computing support hides data from Cohere and cloud providers during inference.
- Integrates with North, letting agents run in your VPC while Cohere handles GPU inference.
- Expansion follows Cohere's pending merger with Aleph Alpha, doubling down on sovereign AI positioning.
Cohere brings dedicated Model Vault inference to Canada
Cohere has made Model Vault available in a Canadian region, allowing enterprises to run its models on dedicated, single-tenant infrastructure in Canada. The launch gives teams with data-residency and isolation requirements a managed alternative to operating their own GPU clusters.
Applications continue to call Cohere models through an API, while Cohere provisions, patches, scales, and monitors the serving infrastructure. Each vault receives dedicated capacity, avoiding the resource contention associated with shared inference services.
What runs inside the vault
| Area | What Cohere provides |
|---|---|
| Isolation | Model-serving infrastructure dedicated to one customer. |
| Capacity | Autoscaling and reserved resources designed to prevent shared-service contention and related rate limits. |
| Models | Supported Cohere generation, embedding, and reranking models. |
| Retention | A Zero Data Retention policy under which Cohere says request data is not stored after processing. |
| Operations | Managed provisioning, scaling, maintenance, and model serving. |
| Access | Direct API calls or integration with Cohere’s agent platform. |
Cohere says administrators can provision a vault through its dashboard within minutes. Model availability may vary by region, so teams should confirm that each required model and version is supported in Canada before committing an application architecture.
Where the security boundary sits
Canadian hosting controls where model inference runs, Zero Data Retention governs persistence after processing, and confidential computing limits access while a request is in memory. These controls address different stages of the request lifecycle.
During ordinary inference, a model server must access prompt data in memory to perform calculations. Cohere says its confidential-computing support processes requests inside hardware-backed protected environments intended to restrict access by cloud operators and Cohere personnel. Buyers should confirm that this feature covers their selected Canadian deployment, models, accelerators, and software versions.
Residency claims also require a review of the surrounding data paths. Security teams should establish where Cohere processes or stores prompts, outputs, embeddings, logs, metrics, backups, support records, and account metadata. Disaster recovery and failover regions require the same scrutiny.
How North connects
Model Vault can provide the inference layer for Cohere’s North agent platform. In that configuration, the application and control plane run in the customer’s environment, including agents, tools, orchestration, and connections to internal systems. North sends model requests to the Canadian vault endpoint, while Cohere operates the underlying serving infrastructure.
This division lets platform teams retain control over application state, permissions, tools, and systems of record while outsourcing model operations. Teams should still document which information enters each inference request and apply access controls before data reaches the endpoint.
Why a Canadian region matters
Cohere is based in Toronto and sells to banks, telecommunications providers, public agencies, and other organizations that often impose contractual or regulatory controls on data location. Canadian inference can simplify architecture and risk reviews for those customers by removing cross-border model processing from the request path.
Compliance still depends on the full system, including client applications, retrieval pipelines, observability tools, identity services, backups, and support workflows. A Canadian endpoint supplies one part of that architecture rather than certifying the complete application.
The rollout also fits the sovereignty strategy behind Cohere’s definitive merger agreement with Germany’s Aleph Alpha, described as a transaction worth roughly $20 billion. The proposed company would maintain headquarters in Toronto and Berlin, with region-specific infrastructure supporting its focus on governed enterprise AI.
Questions to settle before production
- Which prompts, outputs, embeddings, logs, metrics, backups, and support artifacts remain in Canada?
- Does confidential computing cover every required model and hardware configuration, and what attestation evidence is available?
- Do failover and disaster-recovery systems stay within Canada?
- What throughput, concurrency, latency, scaling, and availability commitments appear in the service agreement?
- Which authentication, private-connectivity, encryption, and audit-log options are supported?
- How does Cohere handle model upgrades, deprecations, version pinning, and rollbacks?
- How is Zero Data Retention enforced, audited, and reflected in the contract?
- What pricing applies to reserved capacity, traffic, model choice, and additional environments?
Administrators can create a vault in the Cohere dashboard, select the Canadian region during provisioning, and connect API clients or North. Production rollout should follow capacity testing, failure testing, security review, and confirmation of the applicable residency and service terms.
Comments
