Sovereign AI Funding: Canada's Cohere Stake, UK Budget, Buyer Steps
Canada offers up to $1.6 billion for Cohere and the UK adds £140m for sovereign AI. What buyers should do.
Sovereign AI is a country's or an organisation's ability to build, run and control its own AI, without depending on a foreign supplier for the parts that matter. In the last two weeks of October 2026, two governments put money behind it. On 29 October Canada said it will make an equity investment of up to approximately $1.6 billion in Cohere. On 28 October the UK Budget added £140 million to its Sovereign AI Unit. Neither buys a national model. Both buy domestic suppliers, compute and security capability, and that changes what an enterprise buyer should ask for.
We read the Canadian government's release and the UK Treasury's Budget document. Where a figure comes only from press coverage, we label it as reported.
What is sovereign AI?
Sovereign AI means control over the AI you depend on, and data location is only one part of it. A vendor can host your data in your country and still be run by a foreign parent, depend on foreign chips, or be able to switch you off. Our post on seven kinds of sovereignty breaks control into parts you can score. Our post on sovereign AI vs private cloud explains why isolation alone does not meet the bar.
Governments use the word more broadly than enterprises do. For a government, it often means domestic capability: companies, compute and skills inside its borders. For a buyer, it usually means something narrower and more testable. Where does the model run, who can access it, what law applies, and can you leave?
Keep both meanings in view, because the two announcements below are about the first meaning. They change the supply of suppliers, which affects the second meaning later.
Why is Canada investing in Cohere?
Canada says it wants a sovereign AI supplier that stays Canadian. The release, dated 29 October 2026 in Montréal, says Prime Minister Mark Carney announced a commitment to make an equity investment of up to approximately $1.6 billion in Cohere. The release calls it a commercial investment and an equity stake.
What the release says, as we read it:
- Amount and currency. "Up to approximately $1.6 billion". The release does not name a currency. It uses the dollar sign only. Because the source is the Canadian government, Canadian dollars is the likely reading, but we cannot confirm it, so we write "$1.6 billion" and no more.
- Staging. The money "will be made in stages, with regular reporting on progress against the company's commitments". The release gives no stage sizes, milestones or timing.
- Co-investors. Private investors "including NVIDIA, Bell, and Schwarz Group". The Quick facts section names only NVIDIA and Schwarz Group, for "this financing round".
- What the money is for. More advanced models and products, expanded computing infrastructure and a larger Canadian workforce. The security framing is control over where sensitive data is stored, how it is used and who can access it.
- What Cohere commits to. At least 1,000 new technical jobs in Canada by 2030. At least $1.7 billion of further computing infrastructure in Canada by 2030, again with no currency stated. Staying incorporated and headquartered in Canada. Research partnerships and continued work with the Canadian AI Safety Institute.
- Corporate context. The release says Cohere recently merged with Aleph Alpha, combining Canadian and German expertise.
Two things the release does not say. It does not say Cohere will be a national model that government bodies must use. And it does not say who sets the price or terms for public-sector customers. Our inference: the clearest signal is that a government chose an equity stake, which gives it a seat in the supplier's future, instead of a procurement contract alone. Whether that changes the product is unknown.
What did the UK's 2026 Budget fund for AI?
The Budget document says an additional £140 million for the Sovereign AI Unit, "to back promising UK AI companies to scale and succeed here". The Budget was delivered on 28 October 2026. We read the HM Treasury publication (HC 726), which settles a split that press coverage got wrong.
The Treasury text splits the £140 million into two lines:
- £60 million for the Sovereign AI Fund, from 2027-28 to 2029-30, to invest directly in promising UK AI companies. The Budget text describes this as additional to the existing £500 million Sovereign AI Fund, which has a new focus area of physical and industrial AI.
- £80 million for the Sovereign AI procurement scheme, a challenge-led R&D procurement scheme for innovative AI companies to work on government problems, including health and care, AI security and assurance, and defence.
Our arithmetic: £60 million plus £80 million is £140 million, which matches the headline. Press reports differed on this split, so we use the primary document: £60 million for the fund and £80 million for procurement.
Other AI lines in the same document:
| Item | Amount | Notes (from the Budget text) |
|---|---|---|
| AI Security Institute | Additional £120 million | Over 2026-27 to 2029-30, for frontier AI evaluation and testing |
| National AI Biosecurity Initiative | £165 million | New |
| National Agentic Security Unit | £40 million | To detect and respond to rogue or misaligned agentic activity |
| Cyber defences | £154 million | As stated in the Budget overview |
| Sovereign AI Unit | Additional £140 million | £60 million fund plus £80 million procurement |
| AI Pathfinder Fund | £60 million | Cross-government interventions |
| Local AI partnerships | £30 million in one chapter, £15 million in another | The document gives both figures. We could not reconcile them |
| AI Economics Institute | Additional £23 million | Research on AI's economic effects |
The Treasury states that the Budget has invested over £700 million in AI security, adoption, infrastructure and hardware (paragraph 4.128). Individual allocations should not be added without checking for overlap. The local partnerships figures remain inconsistent between chapters, and that discrepancy is the Treasury's, not ours.
The Budget also says the Ministry of Defence is assessing opportunities to increase sovereign compute and data centre capacity through its secure estate. It sets out a new Data Centre Covenant, which sets tests that strategic data centres are expected to meet. The document describes both as intentions, not funded facilities.
What are governments actually buying?
Three things, none of them a model for you to use.
- Capability inside the border. Equity in Cohere and the UK's Fund both back companies, not products. The aim is that a capable supplier exists locally.
- Compute. Cohere's pledge to spend at least $1.7 billion on computing infrastructure in Canada by 2030 is the release's compute commitment. The UK's covenant is about the conditions that data centres should meet.
- Security units. The UK lines for the AI Security Institute, the National Agentic Security Unit and cyber defence are about testing and responding, not building models.
For an enterprise, that suggests the market will gain more local suppliers and more local capacity over several years. It does not suggest a ready-made sovereign product today. Spending commitments run to 2030, and the Canadian money arrives in stages that are not yet described. Our inference.
Does sovereign AI mean using a national model?
No. Neither announcement creates a national model. Cohere is a commercial company, and the UK programme funds many companies. "Sovereign" describes control properties, and a national flag on the vendor is not one of them.
A Canadian or British supplier can still depend on foreign cloud regions, foreign GPUs, or a foreign parent. A foreign supplier can offer a deployment where you hold the keys and the data stays on your premises. The question to ask is what the arrangement lets someone else do to you, not where the company was founded. Our post on what in-region inference must mean in the EU walks the request path hop by hop, which is the right way to test any claim.
Open-weight models change the picture, because you can run them yourself. They come with licence terms that vary a lot, which we compare in our open-weight licences post.
What should buyers ask a "sovereign" vendor?
Use this as a procurement checklist. Each question has an answer you can ask for in writing.
| Question | Why it matters | Good answer looks like |
|---|---|---|
| Who owns and controls the company? | Ownership decides which laws can compel it | Named parent entities, and any foreign state or investor rights, disclosed |
| Where does inference run, and on whose hardware? | In-region hosting can still use a foreign operator | Named regions, named operators, and a diagram of the request path |
| Who can access customer data, and under what legal process? | Access rights are the core of control | A written list of roles and the legal grounds |
| Who holds the encryption keys? | Whoever holds keys can read the data | Customer-held or customer-managed keys, with the product status stated |
| Can we run the model ourselves? | An exit path is part of sovereignty | Open weights or a deployment package with clear licence terms |
| Can we switch models without rewriting our app? | Portability protects you from a supplier change | A standard API and no vendor-only endpoints in core logic |
| What happens to the service if the vendor is acquired? | Funding and ownership change | Contract terms on change of control and on data return |
| Which parts depend on foreign suppliers? | Chips, cloud and models are separate dependencies | A named dependency list |
Add a data-residency question that covers logs, backups and support access, not only primary storage. Ask for evidence rather than claims. We make no recommendation of any vendor, Cohere included. The two announcements do not tell you how any supplier performs.
What should you do now?
Our view, labelled as such: do not wait for the national programmes. Change what you ask for today.
- Write the control properties into procurement. Use the checklist above in every AI tender, whichever country the vendor is from.
- Separate your application from the model. If your code calls one vendor's proprietary endpoint, a future supplier change is a rewrite. Use a standard interface and keep model IDs configurable.
- Keep an evaluation set. You can only move to a new supplier if you can show it passes your tests.
- Re-check in a year. The Canadian stages and the UK's 2027-28 fund allocations will show whether these programmes produce usable products.
What could go wrong with these programmes?
Three risks are worth naming, all of them our inference and none stated by either government.
Concentration. An equity stake in one company can tilt public-sector buying towards it. If you sell to government, or buy from a supplier that does, watch whether tenders begin to prefer the funded vendor. The Canadian release does not say they will.
Slow delivery. The commitments run to 2030 and the Canadian payments are staged. A buyer with a decision this year cannot wait for a capability that is promised for later.
Label inflation. When money flows to "sovereign" suppliers, more vendors will use the word. The checklist above is the defence: a claim you cannot test is marketing.
None of this argues against the programmes. It argues for treating them as a slow change in who is available to buy from, and for keeping your own architecture flexible meanwhile.
Where Swfte sits
Swfte-hosted customer data is held in AWS eu-west-1 (Ireland). Zero-data-retention agreements with model providers are in progress, which is a status and not a guarantee. For customers who need to run Swfte on their own infrastructure, the enterprise deployment generator produces Terraform, Helm and Docker Compose bundles, and generating one requires a licence key. A dedicated deployment is available on request. Other data locations are agreed per dedicated deployment, and customer-managed keys are designed for dedicated deployments but are not generally available today.
Swfte Connect offers one API across many providers. You can bring your own keys for any provider, and Connect can fall back to another provider in the chain when one is out of funds, rate-limited or down. That supports the portability question above: your application can keep a stable interface while the model behind it changes.
We make no claim about legal sovereignty, and nothing here says Swfte meets any national standard. Use the checklist on us as you would on any vendor.
Related: Seven kinds of sovereignty, Sovereign AI vs private cloud, Sovereign inference in the EU, Open-weight licences compared.
Sources: Prime Minister of Canada, Canada announces landmark investment in Cohere · HM Treasury, Budget 2026 (HC 726) · PublicTechnology, Budget's £700m AI splurge · accessed 9 November 2026.
Method note: we read the sources above on 9 November 2026. The Canadian release names no currency. The Treasury states an AI investment total of over £700 million in paragraph 4.128. The checklist, the inferences and the arithmetic are ours.