The Short Answer
A government stake in an AI company does not give any agency control over the AI it runs. Canada's up to C$1.6 billion investment in Cohere buys equity and company commitments, not an agency's code, keys or exit. Sovereign AI is a property of each deployment. With ibl.ai you own all the code and the data, on infrastructure you control, with any model.
The two ideas answer different questions, and the Cohere announcement is a useful moment to separate them. National industrial policy asks whether a country has AI companies. Agency procurement asks what happens to this system on the day the vendor relationship changes.
What did Canada actually buy when it invested in Cohere?
Canada bought a minority equity position in a company, with conditions on that company. The Prime Minister's October 29, 2026 announcement commits up to C$1.6 billion, in stages, and describes it as a commercial investment.
Cohere's published commitments are to create at least 1,000 Canadian technical jobs by 2030, put at least C$1.7 billion into Canadian compute by 2030, and keep its Canadian incorporation and headquarters.
The announcement states neither a valuation nor the stake's size. BetaKit reports a Series E of up to C$7 billion at about C$36 billion post-money, led by Germany's Schwarz Group, with the federal money coming through the Canada Development Investment Corporation.
Minister Evan Solomon declined to disclose the commercial terms, BetaKit adds.
The Hub estimates the stake at about 4.5%. That is a commentator's estimate in a piece critical of the deal, not an official figure.
What does a government stake in an AI company leave unanswered for an agency?
A stake leaves every deployment question open, because an agency's control comes from its contract and its architecture rather than from its government's shareholding.
The table compares what the published Cohere terms address with what an individual agency has to settle itself.
| Question | Published in the stake's terms? | Where an agency settles it |
|---|---|---|
| Where the company is headquartered | Yes | Not needed |
| Compute built in the country | Yes, at least C$1.7 billion by 2030 | Not needed, though the agency still picks where it runs |
| Who holds the source code of this deployment | No | Licence and escrow terms |
| Who holds the encryption keys and identity provider | No | Architecture |
| Whether the model can be swapped | No | Architecture and contract |
| What keeps running after the contract ends | No | Licence terms |
The announcement does speak to the goal. It frames the investment as giving Canadian businesses and governments more control over where sensitive data is stored and who can access it.
That control is available, and Cohere itself sells it: TechCrunch reported that Cohere's North platform can run on-premise, in hybrid clouds, in VPCs or air-gapped.
A private deployment answers the location question. It does not by itself answer the code and exit questions.
Could a government stake tilt AI procurement toward the investee?
A stake could tilt procurement, but nothing published shows that it has, so treat this as a conflict to manage rather than a finding.
The October announcement states no procurement commitment, and it calls the investment commercial. Yet Canada was already a Cohere customer before it became a shareholder.
In August 2025, Canada's AI and procurement ministers signed a memorandum of understanding with Cohere to explore AI across federal operations.
The Logic has since reported a $339,000, 12-month Shared Services Canada contract putting Command A into CANChat, and a deployment of North for up to 1,400 ISED users, with terms undisclosed.
Those are ordinary, small steps, and they all predate the stake. But shareholder, partner and buyer are three different roles. An officer in any country in that position needs the evaluation criteria written first, so a later award can be defended on its merits.
What four properties make an AI deployment sovereign?
Four properties make a deployment sovereign, and each can be checked in a contract or an architecture diagram.
The code. The agency holds the source code under a licence that survives the vendor. If the vendor is acquired, changes direction or fails, the system keeps running and can be maintained by someone else.
The model choice. The agency can run commercial or open-weight models and switch between them without rebuilding integrations. A pricing change at one provider then becomes a configuration decision, not a migration.
The perimeter. The system runs on the agency's cloud, on-premise, in GovCloud or air-gapped, with the agency's keys and identity provider. For regulated data, that is enforced by infrastructure rather than promised in terms.
The economics. Cost follows use. As an illustration, 10,000 employees at $30 per seat per month is $3.6 million a year whether 500 of them use the tool or all of them do. Per-seat is the wrong shape for a workforce with uneven use.
Does sovereign AI require a model you can reproduce?
Sovereign AI does not always require a reproducible model, but reproducibility is what lets an agency answer an auditor's question about what is inside one. For most workloads, the ability to switch models matters more than the ability to retrain one.
K2 Horizon, released on September 3 by MBZUAI's Institute of Foundation Models, shows the far end of that range. It is six Apache-2.0 models, with data and training code shipped for the smaller sizes at release and promised for the largest.
Where a mission does need that level of transparency, an architecture that treats the model as replaceable can adopt such a release as configuration. We covered that case in why agencies need model-agnostic AI.
What should a government agency ask before buying sovereign AI?
An agency should ask five questions, and each has a yes-or-no answer a vendor can put in writing, whichever country that vendor comes from.
- Do we receive the full source code under a perpetual licence, and does it cover everything we would need to keep running?
- Can we change the model without changing our integrations, and who has done that in production?
- Can it run on our infrastructure, with our keys and our identity provider, including air-gapped if we need it?
- Does cost follow usage, or a seat count?
- Were our evaluation criteria published before any vendor with a government relationship was considered?
Owning code also creates a staffing question: someone has to operate it. BCG's latest research shows how few AI-tagged staff at services firms do applied work, so ask who runs the system after handover.
How does ibl.ai deliver sovereign AI for government agencies?
ibl.ai delivers sovereign AI by handing the agency the platform itself. With ibl.ai you own all the code and the data, under a perpetual licence, and you run it on your own infrastructure.
The platform is model-agnostic across any LLM, with no per-seat pricing, and you can deploy anywhere: your own cloud, on-premise, GovCloud or a fully air-gapped network. Details are on our government solutions page.
ibl.ai is family-owned and operated from New York, NY.
Sources: terms of the investment are from the Prime Minister's announcement. Round size, valuation and lead investor are as reported by BetaKit. The federal release prints a bare dollar sign; we read its figures as Canadian dollars.