If Cohere’s deployment strategy demonstrates how to strengthen Control and Data Authority at the inference layer, its compute expansion illustrates the limits of domestic sovereignty in AI training.
On December 6, 2024, the Government of Canada announced an investment of up to C$240 million in Cohere through its Canadian Sovereign AI Compute Strategy.
The investment was finalized on March 20, 2025, as part of Cohere’s C$725 million compute expansion project in Cambridge, Ontario.
The project is being developed in partnership with CoreWeave, the American AI infrastructure provider that will build and operate the facility. Cohere will act as its anchor tenant.
This project is an important component of Canada’s sovereign AI ambitions.
It also illustrates why infrastructure location and infrastructure control must be evaluated separately.
A data centre may be physically located in Canada and support Canadian AI development while still depending on foreign-controlled operational capabilities and internationally concentrated hardware supply chains.
Understanding who builds, operates, and finances the infrastructure
It is important to distinguish the different roles in this project.
The federal government’s project documentation identifies CoreWeave as the infrastructure partner responsible for building and operating the Cohere compute facility.
A separate development in Cambridge has also attracted attention because CPP Investments announced C$225 million in construction financing for a 54 MW hyperscale data-centre expansion in July 2025.
That separate financing arrangement involves a joint venture between Related Digital, TowerBrook Capital Partners, and Ascent, alongside Deutsche Bank financing. The announcement describes an unnamed GPU-focused AI cloud provider as the long-term tenant.
The public financing announcement does not identify Cohere as the borrower or beneficiary. It should therefore not be presented as C$225 million of financing for Cohere’s C$725 million project.
These are distinct infrastructure transactions, even though both relate to Cambridge’s growing data-centre market.
This distinction is important because sovereign AI discussions frequently conflate ownership, construction, operation, financing, and tenancy.
They are different forms of control.
A company may own a model, another may own or finance the building, a third may operate the cloud platform, and a fourth may manufacture the accelerators.
The sovereignty analysis must account for all of them.
1. Training scale versus sensitive inference
Training a competitive foundation model requires large quantities of specialized accelerators, high-speed networking, advanced cooling, distributed storage, and sophisticated software orchestration.
These capabilities are expensive and difficult to establish.
For Canada, building every component domestically before participating in frontier AI development would create an enormous economic and strategic disadvantage.
Partnering with CoreWeave gives Cohere access to advanced computing infrastructure in Canada without requiring the country to independently develop a globally competitive GPU-cloud operator first.
However, the requirements for model training are not identical to those for sensitive inference.
Training workloads may involve public, licensed, or internally curated datasets. Inference workloads may process classified government information, confidential commercial records, health information, and other sensitive institutional data.
The appropriate sovereignty requirements therefore differ.
Canada may reasonably accept foreign infrastructure operators for some training workloads while insisting on substantially stronger domestic operational and jurisdictional control for its most sensitive inference workloads.
2. Contractual versus juridical sovereignty
The Cambridge project represents a form of managed interdependence.
Cohere gains access to computing resources located in Canada, while the facility’s American operator remains subject to the legal and regulatory environment of its home jurisdiction.
Contractual safeguards, encryption, access controls, customer-held keys, and operational procedures can reduce exposure.
But contracts cannot entirely erase the legal obligations of the infrastructure provider.
This is why contractual and juridical sovereignty should be treated as distinct concepts.
For sensitive workloads, Canada needs infrastructure operators capable of providing meaningful domestic legal and operational control.
Companies such as Hypertec, ThinkOn, Aptum, and other Canadian infrastructure providers can contribute to that ecosystem, alongside government-operated infrastructure.
The objective should not be to prohibit international partnerships. It should be to ensure that Canada has credible alternatives when foreign dependencies become unacceptable.
3. Semiconductor and software chokepoints
Even with servers located in Cambridge, the advanced computing ecosystem remains dependent on a global semiconductor supply chain.
NVIDIA’s accelerator architecture, proprietary CUDA software ecosystem, networking technologies, advanced packaging, and manufacturing relationships are central to contemporary AI infrastructure.
The leading-edge semiconductor fabrication industry is also geographically concentrated, with Taiwan’s TSMC occupying a particularly important position.
These dependencies cannot be eliminated simply by establishing Canadian data centres.
Export controls, regulatory restrictions, geopolitical disruptions, or changes in commercial licensing could affect access to critical computing technologies.
The January 2025 U.S. AI Diffusion Rule, followed by the U.S. government’s May 2025 announcement that it would stop enforcing the rule and move to rescind it, demonstrated how quickly the regulatory environment surrounding advanced AI computing can change.
The lesson is not that Canada must manufacture every semiconductor domestically.
It is that Canada must understand its exposure to external chokepoints and establish alternatives wherever those dependencies threaten critical national capabilities.
This is managed interdependence: accepting participation in a global technology ecosystem while ensuring that no single external provider can dictate the country’s strategic options.
The missing ingredient: Continuity
A Canadian data centre is valuable, but infrastructure location alone cannot guarantee continuity.
The real question is whether Canadian organizations can maintain access to critical compute resources when commercial relationships, export rules, or geopolitical conditions change.
This requires several layers of resilience:
- Multiple infrastructure providers.
- Alternative accelerator and software configurations where technically feasible.
- Long-term access and maintenance arrangements.
- Tested workload migration procedures.
- Independent operational expertise.
- Continuity plans for critical model-serving environments.
The country does not need to duplicate the entire global hardware supply chain. It does need to avoid creating critical dependencies without viable contingencies.
Government of Canada — Original C$240 million commitment, December 6, 2024.
Government of Canada — Finalized Cohere investment, March 20, 2025.
4. Open Weights and the Hidden Problem of Substitutability
One of the most strategically important developments in Cohere’s model portfolio is the release of Command A+ on May 20, 2026.
According to Cohere’s technical documentation, Command A+ is a sparse Mixture-of-Experts (MoE) model containing:
- 218 billion total parameters.
- Approximately 25 billion active parameters per token.
- A 128,000-token context window.
- Up to 64,000 output tokens.
- Text and image input capabilities.
- Multilingual support across 48 languages.
- Reasoning and agentic tool-use capabilities.
- An Apache 2.0 licence for its released W4A4 quantized checkpoint.
The model’s architecture illustrates an important principle of modern AI economics: total parameter count does not necessarily determine the computing resources required for every inference operation.
In a sparse MoE model, only selected expert components are activated for a given token. This can make a model with hundreds of billions of parameters more practical to deploy than a similarly sized dense architecture.
Cohere documents W4A4 four-bit weight-and-activation quantization configurations using a single NVIDIA B200 GPU or two H100 GPUs.
This is strategically relevant.
A model that requires an enormous multi-node cluster limits deployment to a small number of wealthy organizations and hyperscale infrastructure providers.
Reducing hardware requirements makes advanced capabilities more accessible to universities, research organizations, enterprises, and public institutions with their own computing infrastructure.
It does not make frontier-grade AI inexpensive, nor does it eliminate hardware dependency. H100 and B200 accelerators remain high-end data-centre components.
But it lowers the infrastructure threshold compared with systems requiring much larger clusters.
Open weights are not the same as fully open-source AI
There is another important distinction.
Command A+ is distributed under the Apache 2.0 licence for the documented W4A4 quantized checkpoint.
This provides users with significant freedom to obtain, deploy, modify, and redistribute that released artifact, subject to the licence’s conditions.
However, open weights do not automatically constitute a fully open-source AI system.
The Open Source Initiative’s Open Source AI Definition identifies three important components required for meaningful modification of an AI system:
- Information about the training data.
- The complete code used to process data, train, and run the system.
- Model parameters, including weights.
Cohere also publishes an EU AI Act public summary of Command A+’s training content, which provides transparency about the data sources and processes used to develop the model.
However, a regulatory training-content summary and the complete information needed to independently reproduce an AI system are not necessarily the same thing.
This distinction matters for governments and enterprises evaluating long-term technological control.
The ability to download a model is not identical to the ability to recreate its training process, maintain its software ecosystem, or independently reproduce its capabilities.
The RAG portability problem: replacing the model is not enough
There is a deeper technical challenge that is frequently overlooked in discussions of AI sovereignty.
Most enterprise AI systems are not simply foundation models. They are interconnected systems built around institutional data, embedding models, vector databases, retrieval engines, reranking systems, orchestration frameworks, and applications.
Consider a typical retrieval-augmented generation architecture.
An organization takes its documents, policies, contracts, research records, and other institutional knowledge and converts them into numerical representations called embeddings.
These embeddings are stored in a vector database, which allows the system to retrieve relevant information when a user submits a question.
A foundation model then uses the retrieved information to generate a response.
The challenge emerges when an organization tries to change one of these components.
Replacing the foundation model does not necessarily require rebuilding the knowledge index. Replacing the embedding model often does.
Embeddings generated by different models generally occupy different vector spaces. Even when two models produce vectors of identical dimensions, those vectors are not automatically interchangeable.
MongoDB’s documentation for its Voyage AI embedding models explains why changing embedding models requires regenerating corpus embeddings to preserve reliable retrieval.
For a large enterprise, this can introduce significant migration costs, operational disruption, and validation requirements.
This means an organization can successfully move away from a foreign foundation-model provider while remaining locked into proprietary retrieval infrastructure.
The problem extends beyond embeddings.
A complete migration may also involve:
- Vector database formats and export capabilities.
- Proprietary reranking models.
- Retrieval and orchestration frameworks.
- Application-specific model integrations.
- Metadata and document-permission structures.
- Evaluation pipelines and quality benchmarks.
The implication is important:
True substitutability must extend across the entire enterprise AI stack, not merely the foundation-model layer.
Cohere’s own retrieval stack: Embed 5 and Rerank 5
This challenge applies to Cohere itself.
In September 2026, Cohere released Embed 5, a new family of embedding models designed for enterprise retrieval across text, images, and complex documents.
The family includes Pro and Fast variants that share an embedding space, allowing an organization to index a corpus using one variant and query it using the other.
Cohere provides Embed 5 through its API, partner platforms, and single-tenant Model Vault deployments.
On October 21, Cohere introduced Rerank 5, its latest enterprise reranking models.
Rerank 5 is designed to improve retrieval quality by reordering candidate documents according to their relevance to a query. Its Context Focus capability can identify relevant passages within documents, helping reduce the amount of unnecessary information passed to the foundation model.
Cohere documents deployment through its API and Model Vault, with private VPC and on-premises deployment available through sales arrangements.
These releases strengthen Cohere’s position beyond the foundation-model layer.
But they also introduce a procurement question that applies to every provider, including Canada’s national champion:
Can customers independently maintain, export, migrate, and replace the retrieval components on which their institutional knowledge depends?
Model Vault, for example, is a Cohere-managed, isolated single-tenant inference environment. It offers stronger isolation than a shared API but should not automatically be confused with customer-operated infrastructure.
Buyers should establish in writing which components they can export, which they can self-host, and what migration support and exit rights they retain.
Sources: Cohere Embed 5 documentation and Cohere Rerank 5 documentation.
Standardization as an exit strategy
There are practical ways to reduce this dependency.
Open inference frameworks such as vLLM provide OpenAI-compatible APIs for serving supported models.
These interfaces can reduce the integration effort involved in moving between compatible model providers.
However, API compatibility alone does not guarantee equivalent model behaviour, performance, tool use, or output quality.
Similarly, standardized data-export formats, portable vector indexes where supported, and documented embedding migration procedures can reduce the costs of changing retrieval providers.
Governments and enterprises should require explicit portability provisions and realistic migration plans in their AI procurement agreements.
For embedding systems, this should include documented model versions, exportable source documents and metadata, tested re-embedding procedures, and independent evaluation of retrieval quality after migration.
A national AI strategy that makes foundation models replaceable but leaves enterprise knowledge infrastructure locked to a single provider has only solved part of the problem.
For Canada, this is where Cohere’s open-weight strategy becomes particularly interesting.
Its long-term sovereignty value will depend not just on whether Command A+ can run outside Cohere’s own infrastructure, but also on whether institutions can integrate it into systems they fully understand, maintain, and replace.