A vice-president at a Canadian university called me last month.
Her president had told her the institution needed to be on sovereign AI within the year.
The provincial government had announced a funding program. Other universities were already applying.
She wanted to know what she should be evaluating.
I asked her what sovereign meant in this context.
She paused for a long time. Then she said, "I think it means the data stays in Canada."
I asked what about it staying in Canada specifically mattered.
The physical location of the hardware?
The corporate nationality of the operator?
Which government could compel access? Some combination?
She said she would need to check with her procurement team and call me back.
She did not call back.
I have had some version of that conversation about thirty times in the past six months.
Universities, federal departments, Crown corporations, hospital networks, law firms.
The framing is always the same. Sovereignty matters. We know it matters. We need to do it.
But when I ask why, the answers fall apart.
Prime Minister Carney has said Canada needs sovereign AI. He is right.
The federal government is funding it.
The provinces are funding it.
The universities are organizing around it.
The procurement language is being written.
The money is moving.
And almost nobody can articulate what the word actually means.
The Problem With "Own Your AI"
When I built the product framework for SimpleDirect, we split AI deployment into three levels.
Renting — you call someone else's API.
Owned — you download an open-weight model and run it on a cloud provider you choose.
Sovereign — you run it on infrastructure under Canadian jurisdiction.
The full framework is on the SimpleDirect page.
It works for picking a deployment tier.
Most businesses should be at Owned.
A smaller fraction need Sovereign.
Almost nobody should be at pure Renting for any work that involves real data.
But the product framework doesn't draw out the deeper distinction.
At the product layer you don't need it. At the policy layer, you do.
Owned and Sovereign are not points on the same line.
They are answers to two different questions.
The Two Questions
Ownership is a commercial question.
Do you control the model?
Can the vendor change the price tomorrow?
Can they deprecate it and force you to migrate?
Are the weights yours, or are you renting access to weights they keep on their own servers?
The ownership question matters.
It determines whether you have leverage in your AI infrastructure, whether you can switch providers without rebuilding, whether you control the cost curve.
Most Canadian businesses are getting this question wrong by defaulting to closed API providers when an open-weight model would give them more control.
The SimpleDirect framework is built around solving this.
Sovereignty is a legal question.
Whose government can compel access to your data?
That is the whole question.
Not where the data is. Not who owns the company. Not who pays the bill.
Whose subpoena works. Whose national security letter works.
Whose court order compels the operator to hand over what they're holding.
If the answer is "only the Canadian government," that's Canadian sovereign.
If the answer involves any other government's legal authority, that's something else.
It might still be the right answer for your work.
It might be exactly what you need. But it is not sovereign by the strict definition that policy decisions should be using.
Why The Confusion Is Convenient
Here is the thing about the word "sovereign" being fuzzy.
It is fuzzy because the fuzziness is useful to almost everyone selling an answer.
It is useful to AWS, Azure, and Google Cloud.
They all sell "sovereign" offerings with Canadian data residency.
The data sits in Canada. The operator is US-incorporated.
The parent company is subject to US law including the CLOUD Act, which permits US law enforcement to compel disclosure of data held by US providers regardless of where it physically lives.
The Canadian-residency part is real. The sovereignty part is a marketing claim.
It is useful to Canadian managed service providers running AI workloads on top of US hyperscaler infrastructure.
The provider is Canadian.
The underlying cloud is US. The chain of compulsion still runs back through Washington.
It is useful to federal funding programs that need to distribute money without picking winners.
A loose definition lets the money flow to many credible-sounding offerings. Specifying which ones actually qualify is politically harder than not specifying.
And it is useful to institutional buyers who need to satisfy a procurement requirement without doing the legal due diligence.
A loose definition lets them check the box and move on.
The fuzziness is not an accident. It is a feature for everyone who benefits from it.
It is just not a feature for anyone trying to figure out what sovereignty actually means.
The Compulsion Test
You can disambiguate any offering with one question.
Whose government can compel access to this data?
Apply it.
AWS Sovereign Canada.
Data in Canada. Operator is Amazon Web Services, a US-incorporated subsidiary of Amazon. Amazon is subject to US law including the CLOUD Act.
If you're finding this useful, I send essays like this 2-3x per week.
·No spam
Answer: the US government can compel access. Not Canadian sovereign.
Microsoft Azure Government Canada.
Same structure. Microsoft is US-incorporated. CLOUD Act applies.
Answer: US government can compel access. Not Canadian sovereign.
Google Cloud Canada.
Same structure. Alphabet is US-incorporated.
Same answer.
Telus Sovereign AI Factory in Rimouski.
Operator is Telus, Canadian-incorporated, no US parent. The CLOUD Act does not apply to Telus.
Answer: only the Canadian government can compel access. Canadian sovereign.
OVHcloud Beauharnois.
Operator is OVH, a French-incorporated company. Not subject to US CLOUD Act.
Answer: French government can compel access, plus Canadian government since the infrastructure sits in Quebec.
Not Canadian-only sovereign, but not US-exposed either.
Fine for most cases that need to avoid US jurisdiction specifically.
Cohere.
This one is harder.
Cohere is Canadian-incorporated, Toronto-based, founded by Canadians. Good so far.
But Cohere also has significant US investment from NVIDIA, Oracle, Salesforce, and others.
The inference infrastructure may execute on US-owned cloud.
Whether they qualify as sovereign by the strict definition depends on factors most institutional buyers have not audited — where inference physically executes, the governance structure with US investors, the specific terms of their enterprise data processing agreements.
Worth asking directly rather than assuming.
Your own infrastructure in a Canadian colo with a Canadian operator and no US parent.
Answer: only the Canadian government can compel access. Canadian sovereign.
The test is short. The answers are usually clear.
The reason most buyers don't apply it is that the sales materials are designed to obscure the question rather than answer it.
Why This Matters For Canadian Policy
The PM's instinct is right. The federal AI Compute Access Fund and the provincial programs are real attempts to build Canadian capability. Good.
But policy commits to nothing if the term means everything.
If "sovereign" in the procurement criteria can describe AWS Sovereign Canada, then the money flows to AWS Sovereign Canada.
The Canadian operator with actual jurisdictional control is now competing on price and convenience against a US hyperscaler that gets to claim the same word.
The hyperscaler wins. It is bigger. It is faster. It is more polished. It was already on the procurement officer's vendor list.
The result is that real sovereign capacity does not get built.
The money funds offerings that satisfy the loose definition without addressing the underlying concern.
Five years from now Canada has "sovereign AI infrastructure" on paper and no infrastructure that actually survives the compulsion test.
This is the failure mode the word was supposed to prevent.
The way to prevent it is to define the term precisely in the funding criteria. Whose government can compel access?
That question, asked of every applicant, with the answer documented and verifiable, is the entire policy lever.
It is not a hard fix. It just requires someone with authority deciding to do it before the money is committed.
What Buyers Should Actually Ask
If you are at a university, a federal department, a Crown corporation, or a regulated firm evaluating sovereign AI right now — here is the conversation to have with the vendor.
Who is the legal operator of the infrastructure?
Where is that operator incorporated?
What is the corporate governance structure, including parent companies and significant equity holders in foreign jurisdictions?
Where does inference physically execute?
Under which government's compulsion authority does the operator fall?
If a foreign government issued a legal order to disclose data held in this offering, what is the operator's obligation under that government's law?
These are not trick questions.
They are the questions that determine whether the offering is what it claims to be.
Reputable vendors will answer them directly.
Vendors who deflect, or who describe data residency without addressing corporate jurisdiction, are answering a different question than the one you asked.
Build these questions into your RFP language.
The vendors who can answer them are the ones building real sovereign capability.
The vendors who cannot are building something else.
The Right Conversation
Sovereignty is not an aesthetic.
It is not "Canadian-feeling" infrastructure.
It is not the brand color on the marketing page or the picture of a maple leaf on the dashboard.
It is a legal posture.
A specific answer to a specific question about which government's laws apply to your data.
The PM is right that Canada needs it.
The funding programs are right to support it. The universities and the departments and the firms are right to want it.
But none of that adds up to anything if "sovereign" can describe seven different deployment topologies with seven different jurisdictional postures.
The conversation should start with the compulsion question, not end with it.
Most people skip the question entirely.
That is why most Canadian "sovereign AI" today is not actually sovereign — and why the policy moment will be wasted unless someone with authority defines the term before the money is committed.
It is a one-sentence definition. Canada just hasn't written it down yet.