Sovereign Compute, Borrowed Data

Canada released its national AI strategy last week. AI for All. Six pillars, nearly $200 billion in projected productivity, up to 250,000 new jobs, business adoption climbing from roughly 12% to 60% by 2034. It is a serious document, written by serious people, and most of it is right.

I want to talk about one word, because that word is working harder than anyone on the org chart.

The strategy deploys "sovereign" the way the last decade deployed "cloud" — confidently, repeatedly, and at no point burdened by a definition. Pillar 4 is "the foundation for Canadian sovereign AI": sovereign compute infrastructure at scale, resilient, sustainable, and under Canadian governance.

Compute. Governance. Two of the three things sovereignty needs, listed with admirable confidence. The third is not in the sentence. It is also the only one that decides whether the first two were worth typing.

The third is custody of the data.

What "sovereign compute" is actually promising

Strip the language down and "sovereign compute under Canadian governance" makes two claims. One is geographic: the GPUs are on Canadian soil. The other is administrative: a Canadian entity is in charge of them.

Both are true. Neither will save you.

Geography is reversible. A border is a policy decision, and policy decisions have a half-life. Governance is a promise — binding right up until a foreign law walks past it on the way to the operator who made it. "Under Canadian governance" does not impress a US warrant served on a US parent company. The CLOUD Act has never once asked which province the rack lives in. It also does not require anyone to send you a courtesy note afterward.

So admire the thing we are funding. Sovereign Canadian compute, in a sovereign Canadian building, with a sovereign Canadian flag in the lobby, running inference on data the operator can read like a paperback. That is not sovereignty. That is a very expensive way to host someone else's leverage, with excellent insulation.

The layer that didn't make the slide

The foundation pillar names compute. It names talent. It does not name custody of the data those models spend all day reading and writing.

The data was apparently the one that got cut for length.

This matters, because sovereignty was never a question of where the bytes nap at night. It is a question of who can read them and who can compel their surrender. Residency gives you the postal code. On the matter of who holds the keys, residency has nothing to say and never did.

There is precisely one flavour of sovereignty that survives a subpoena you are not told about: the architectural kind. Data encrypted and sharded so that no operator — in any jurisdiction, under any flag, holding any warrant and any number of lawyers — can read it or hand it over in a form anyone can use. Get that right and "Canadian governance" stops being a promise and becomes a property of the system. Enforced by mathematics, which does not take meetings and cannot be deposed.

Not because you trust the operator. Trusting the operator is how you got here. Because the operator was never handed anything worth the trouble of compelling.

This is the gap between trustworthy and verifiable, and it is not a word game. A trustworthy system asks you to believe it will behave itself. A verifiable system spares you the theology. One of those is a posture you adopt for the press release. The other is a result you can prove at 3 a.m. to a regulator who is not in a generous mood.

Own the control plane, or rent it forever

The Council of Canadian Innovators made the commercial version of this point last week, and made it well: we are on track to own the pipes while someone else owns the part that makes the pipes worth owning. Canada has run this play before, with various things we dug out of the ground. We are very good at building the plumbing and watching the margin emigrate.

The sovereignty version of the trap is parked in the next spot over. The pipes — data centres, compute, fibre — are the fundable, photogenic, ribbon-adjacent layer. The layer that decides whether the data inside them is actually sovereign is the control architecture: who can read it, who can be compelled to surrender it, who holds the keys. That layer is intellectual property. It is patentable. It is ownable. At present it is largely owned by people who do not file Canadian taxes.

Fund the racks, import the control plane, and the result is Canadian-flagged plumbing operated under foreign law. A sovereign address. A foreign landlord. Excellent curb appeal.

What I would change

Nothing about the ambition. The ambition is correct and roughly a decade overdue. Treating AI adoption as a national-competitiveness problem instead of a hobby is the right instinct, and I would not touch it.

One thing about the definition.

The strategy says this is a five-to-ten year transition, which is honest. It also means the framing we choose this month is the framing we are stuck explaining at every conference until 2034. So settle it now, while the concrete is wet: does "sovereign" mean hosted here, or does it mean provably beyond foreign reach? Those are two different buildings. Only one of them is still load-bearing after the first warrant nobody mentions to us.

Use the anchor-customer mandate in Pillar 5 to buy the control plane, not merely the compute. Make Canadian-controlled data custody a line item in the procurement, not a value in the preamble. Reward the companies whose sovereignty you can verify over the ones whose sovereignty comes with a firm handshake and a deck.

Do otherwise and we will spend a decade and two hundred billion dollars carefully moving the problem to a nicer neighbourhood — and we will hold a press conference to call the move a strategy.

Sovereign compute is the half you can buy with a purchase order. The other half you have to architect, which is harder, less photogenic, and does not come with a ribbon. So naturally it is the half we are skipping.

You can move a server out of a country. You cannot subpoena an architecture. One of those is for sale this fiscal year. We should be building the other one.

Overview of Canada’s National Artificial Intelligence Strategy: AI for All

Originally published by Ross Norrie, founder of SkyeConnex, on LinkedIn.

Published June 5, 2026 · More from the SkyeConnex blog