Canada Is Renting Its Sovereignty

In the last piece I wrote about the anchor:

Canada designated Cohere as its national champion at the foundation model layer — $240 million, a 24-year term, an MOU that called the company "the only sovereign, cloud-agnostic large language model operating in Canada," and a fund Cohere itself proposed.

That designation is settled for a generation.

This piece is the opposite kind of writing.

The last one was analysis of public records, and I tried to stay out of it.

This one carries a point of view.

I should be upfront that I operate in Canadian AI infrastructure, so read it as an operator's argument, not a neutral survey.

Here's the argument, and it's sharper than the last one. Canada has a sovereignty strategy on paper and an American operating reality in practice.

The champion is in the press release. The production system is rented.

What the government actually runs

Start with the most boring, most revealing document in the pile:

A Treasury Board policy notice telling federal institutions how to turn on Microsoft Copilot for Work.

Treasury Board reserves the right to audit which departments enable it.

This isn't new ground for Microsoft in Ottawa.

Azure has held a Government of Canada cloud framework agreement since 2019 and is certified to host Protected B data.

Meaning sensitive federal workloads have run on a US-incorporated company's cloud for more than half a decade, well before the word "sovereignty" was fashionable.

Copilot is the AI layer on top of that same stack.

Ontario went further and faster on the productivity side.

Internal materials reportedly put more than fifteen thousand provincial civil servants on Copilot every week — the highest adoption in the country.

The paperwork is public. The texture behind it — what's actually live, in which departments — I've checked against conversations with people working in and around these deployments.

It matches. Nobody is hiding this. It simply hasn't been named for what it is.

Now the consequential layer, and it's closer to home than any cloud contract.

Immigration, Refugees and Citizenship Canada has used advanced analytics for years to triage applications — machine-learning systems that sort temporary resident visas, study permits, and sponsorships, auto-streaming the low-risk and flagging the rest.

The automated triage of low-risk visa applications goes back to a 2017 pilot.

Immigration lawyers have spent the past two years raising procedural-fairness concerns about refusal letters that don't appear to reflect meaningful human review.

This is AI making consequential calls about real people's lives, in production, today.

Immigration cuts against my own thesis in an instructive way, so I'll be straight about it: IRCC's triage was largely built in-house, on government systems.

It is, in that narrow sense, sovereign. And it is still contested in Federal Court.

So the immigration story isn't really about who owns the infrastructure.

It's about whether the government can explain and govern the AI it deploys on citizens — even when it built that AI itself.

Hold that thought. The deeper problem isn't just where the models run.

It's that Canada is improvising its AI posture in real time: building in a few narrow places, renting almost everywhere else, and governing both unevenly.

Then the part everyone assumes is the most sovereign, and isn't.

The Department of National Defence's own AI Strategy commits the military to being "AI-enabled by 2030" — and then admits, in writing:

Neither DND nor the Canadian Armed Forces is "positioned to adopt and take advantage of AI," describing the current approach as "fragmented."

The strategy states plainly that the most advanced AI now lives outside defence, in industry and academia and open source, and that DND will therefore have to procure and co-develop with partners.

For corporate functions, defence sits inside the same Copilot guidance as the rest of government.

What the public record does not show is which models run on which clouds inside defence operations.

That's the document I'd want to see — the one an Access to Information request might surface — and I won't assert what I can't yet show.

But the shape of the record is already clear enough: a military that admits it isn't ready, in a country whose sovereign champion it does not appear to actually run on, buying capability from whoever can deliver.

The case for renting

Before this starts to sound like a scandal, I should make the other side's argument, because it's strong.

Renting is the rational choice right now. Azure is mature, FedRAMP- and CCCS-certified, supported, and already wired into how departments work.

A from-scratch sovereign alternative for everyday knowledge work would realistically be slower, worse, and more expensive for years before it caught up — if it ever did.

A deputy minister who picks Copilot over a sovereign tool that doesn't exist yet isn't being careless.

They're being responsible with a budget and a mandate to deliver service now.

And no serious person thinks sovereignty means total self-sufficiency.

Middle powers share and borrow capability constantly, from NORAD to Five Eyes to NATO. Some dependence is just how a country Canada's size operates.

So this isn't a story about foolish procurement. The decision to rent is defensible.

The problem is narrower and slower-moving: temporary dependence has a way of hardening into permanent architecture.

Every workflow built on a rented foundation raises the cost of ever owning one. The convenient choice today quietly forecloses the sovereign choice tomorrow.

That's not a scandal. It's structural drift — and structural drift is the kind of thing that stays invisible until it's irreversible.

The scale of the drift is worth a number.

Canada budgeted roughly $2 billion for its Sovereign AI Compute Strategy. Over the same window, Microsoft alone committed about C$19 billion to expanding Canadian cloud and AI capacity — including, with no apparent irony, something it calls a "Sovereign AI Landing Zone."

The sovereignty strategy is being outspent nearly ten to one by a single American vendor deepening the very dependence the strategy was meant to reduce.

That's not a fair fight, and pretending it is would be the real naïveté.

Why renting isn't free

This is where the CLOUD Act matters — not as a headline, but as the reason renting carries a cost that doesn't show up on the invoice.

The 2018 US statute lets American authorities compel a US-incorporated company to produce data it controls, regardless of where that data physically sits.

The detail people fixate on — which model answers a given prompt, OpenAI's or Microsoft's own — matters less than the fact underneath it: Microsoft is a US-incorporated company, and that is the part US law reaches.

Microsoft's enterprise terms promise not to train on tenant data, and those promises are real and beside the point.

When a valid US order arrives, the company complies.

Jurisdiction follows incorporation, not geography.

A Canadian data centre operated by an American company is data residency, not data sovereignty.

The government's own data-sovereignty white paper concedes the point: a provider operating in Canada but subject to foreign law leaves Canada without full sovereignty over the data.

Even the designated champion doesn't escape it.

Cohere is hosted, and its named compute partner, CoreWeave, is American-incorporated — so the sovereign champion's own stack carries the exact exposure the strategy was meant to solve.

But that's almost beside the point now, because the more revealing fact is simpler: the government doesn't run on its champion. It runs on Copilot.

The sovereignty is the designation. The dependency is the deployment.

If you run a regulated Canadian organization and someone sells you "sovereign AI," there's one question that cuts through the whole pitch.

Don't ask where your data is stored. Ask who can be compelled to hand it over.

The first question has a comforting answer and no meaning. The second has meaning — and, for most of what's deployed in this country, an uncomfortable one.

Where the opportunity actually is

So where does that leave an operator like me.

Not at the foundation model layer — that's filled for twenty-four years.

The opportunity is everywhere the sovereign option loses by default, which is to say almost everywhere the actual work happens.

The federal public service didn't choose American rails because it stopped caring about sovereignty.

It chose them because a deployable sovereign alternative for everyday work didn't exist, and convenience always wins against a thing that isn't built yet.

That's not a values problem. It's a supply problem — and supply problems are the ones operators get to solve.

I have to be straight about the field, because the lazy version of this ends with "and it's all wide open." It isn't.

There are managed service providers selling Canadian-owned hosting, boutiques auditing which of your tools are exposed to foreign jurisdiction, law firms publishing a CLOUD Act explainer every other week.

If you picture empty territory, you're picturing it wrong.

What's scarce isn't another vendor. It's the operator who connects three things almost nobody connects: what the policy actually says — down to the documents most people never read — what the procurement reality actually is, and what a managing partner or a deputy minister actually has to decide on Monday.

That takes primary-source credibility and enough reach to be heard over a field all shouting the same word.

It's a narrower and more defensible position than being the forty-fifth sovereign cloud.

And it's worth being honest about what this very piece is.

I'm not describing the explainer layer from the outside.

Reading the ATIP, the Treasury Board notice, and the defence strategy, and saying out loud what they add up to — that is the work, and doing it in public is the position.

The rest, I'll have more to say about soon.

What's settled, and what isn't

The anchor is set for a generation; arguing with it is wasted breath.

What isn't settled is the gap between what Canada says about sovereignty and what Canada runs.

Right now the country has written itself a sovereignty strategy and logged into an American one.

That can be the defensible choice today and the wrong architecture for the decade — both at once.

The question isn't whether Canada should ever rent. It's whether it notices when renting has quietly become the only thing it knows how to do.

The work, for whoever wants it, is building the thing worth owning — at the layers where the renting actually happens, before the drift sets.

Sources

The factual claims here are public and verifiable.

Verify exact figures and quotes against the originals before publishing — especially the white-paper wording, which is paraphrased.

• Microsoft Copilot for Work in the federal government, and Treasury Board's audit reservation — TBS, Microsoft Copilot for Work Policy Implementation Notice (Canada.ca, 2025); Responsible use of artificial intelligence in government (Canada.ca)
• Azure Government of Canada framework agreement since 2019 and Protected B certification — Microsoft, Canada Protected B – Azure Compliance; Microsoft News Centre Canada (Dec 2019)
• Ontario Public Service Copilot adoption (~15,000 weekly) — Global News reporting, as summarized Nov 2025
• Microsoft's ~C$19B Canadian cloud/AI investment (2023–2027) and the "Sovereign AI Landing Zone" — Microsoft investment announcement, Dec 2025
• IRCC advanced analytics, 2017 triage pilot, and procedural-fairness concerns — CBC, Immigration lawyers concerned IRCC's use of processing technology (Sept 2025); IMI Daily (Nov 2025); Meurrens on Immigration
• DND/CAF AI Strategy: "AI-enabled by 2030," "fragmented," not "positioned to adopt," procurement from outside defence — DND/CAF Artificial Intelligence Strategy (Canada.ca / D2-633/2024E-PDF); CBC, New DND strategy warns Canadian military's approach to AI 'fragmented'
• US CLOUD Act, residency vs. sovereignty, and the GoC white-paper concession — BLG (April 2026); Osler (Nov 2025); CanSpace (2026)
• The $2B Sovereign AI Compute Strategy, $240M SIF contribution to Cohere, CoreWeave partner, and MOU "sovereign, cloud-agnostic" language — the ATIP records and announcements detailed in the companion piece