Canada's "AI for All" strategy names Ottawa the anchor customer for sovereign AI — while quietly running $46.8M of Palantir contracts. The procurement gap, explained.
Canada's federal government released "AI for All" in June 2026 and named itself the "strategic anchor customer" for Canadian artificial intelligence. The inconvenient part, as Caseway founder Al Vigier argued in a widely discussed op-ed in The Line, is that Ottawa is already a serious AI customer — it just buys American, and it buys quietly. The piece hit the Hacker News front page and reopened a debate every country pursuing sovereign AI now faces: what does "sovereign" mean when your defence department runs on Palantir?
It's the same tension we mapped in the Europe AI landscape and in Switzerland's answer, the Apertus open foundation model: sovereign-AI strategies are loud about building domestic capability and quiet about the foreign systems already running inside government.
TL;DR: The Questions People Are Asking
Question
Answer
What did Canada announce?
"AI for All" — Ottawa as anchor customer for Canadian AI; 60% business adoption target by 2034 (from ~12% today)
What's the Palantir issue?
Undisclosed DND contract from March 2020: $14.4M → ~$44.4M value, ~$46.8M spent; Ontario Provincial Police on Gotham since 2015
What's actually in the strategy?
$500M for equity stakes, $700M for compute, Trusted AI Certification, a health-first missions program — notably, no purchase orders
Is Canada alone in this?
No — France, Germany, Spain, and the UK are all distancing themselves from Palantir
Can Canada build its own?
Contested — Cohere exists, but talent flight to US salaries and closed procurement are real obstacles
The Core Contradiction: Buying Quietly While Promising Sovereignty
Vigier's numbers are the sharpest part of the argument. Only about 12 per cent of Canadian businesses use AI, and barely eight per cent of small firms. The strategy wants 60 per cent by 2034, with government purchasing supposedly leading the way.
But the purchasing already happens — behind a wall of "not for public disclosure":
The Department of National Defence signed a contract with Palantir's Canadian arm in March 2020, starting at $14.4 million. It was never disclosed. More than a dozen amendments later, it reached about $44.4 million in value, with roughly $46.8 million actually spent by October 2025.
A separate $3.7 million Defence contract surfaced only after a Conservative MP pressed the government on its AI spending.
The Ontario Provincial Police have run Palantir's Gotham platform since 2015 — precisely the data-fusion and decision-support category the strategy says Canada must own.
The op-ed's test for an anchor customer is binary: did it buy, in the open, from Canadian firms? Nothing in "AI for All" forces that answer to be yes.
What the Strategy Offers Instead of Purchase Orders
Look at the instruments: $500 million to take equity stakes in promising businesses, $700 million for compute, a Trusted AI Certification program, and a missions program starting in health. Vigier's observation: not one of those is a purchase order.
His argument is that governments reach for equity and certification schemes precisely when buying is hard — grants and equity are what you offer when your procurement machinery cannot write a clean contract to a small vendor. Even the one procurement promise routes through a new Office of Digital Transformation: an office, not a mandate to write cheques.
Two of his warnings deserve attention from founders anywhere:
Government equity is a mixed blessing. A state that takes a stake in your company takes a seat at your table and "plants a doubt in the mind of every allied investor." Government-built national champions have a long record of becoming wards of the state. This concern applies well beyond Canada — the same critique gets leveled at Nvidia's ecosystem investments, which we broke down in Nvidia's revenue-share program for AI startups.
Starting missions in health is starting where you'll stall. Health is the most privacy-bound, slowest-procuring corner of the state. If you wanted to prove public AI buying works, you'd start where the law already demands an audit trail: courts, public safety, defence.
The Hacker News Debate: Three Real Disagreements
The HN thread surfaced substantive disagreements worth separating:
1. Is Palantir specifically the problem, or foreign vendors generally?
One camp argued "America or not" is a red herring — the issue is Palantir's governance and alignment with the current US administration, which has repeatedly threatened Canadian sovereignty. Commenters catalogued the European exodus: France, Germany, Spain, and Britain have moved or are moving to disassociate from the company, with German intelligence offices snubbing it outright. Another camp countered that a home-grown Palantir would do the same things "with a Canadian Alex Karp" — the fix is on-premises deployment and oversight, not flag-swapping.
2. Can Canada actually build this?
The pessimist case: only the US and China can run domestic products of this kind; Canadian engineers flee to triple their pay; Cohere hasn't shipped a top-20 model recently. The optimist case: Palantir's stack "is not really that complex" — its famously scrappy forward-deployed engineer model is a services motion, not deep technical moat — and govtech is hard to enter mainly because procurement is closed. Canada does have real assets: Cohere's open-weight Command A+ line and its North Mini coding models are genuine sovereign-adjacent capability.
3. Is "sovereign AI" mostly theater anyway?
The bleakest thread: Canadian critical infrastructure — grid operations, banking, cybersecurity — already runs on US cloud products end to end. On that view, "sovereign AI" buzzwords are a way to funnel tax money to programs without touching the actual dependency. It's the same critique leveled at European initiatives, and the same reason China's approach — free models and cheap domestic compute — is studied so closely: it attacks the dependency itself, not the branding.
Why Equity Stakes and Certifications Are the Easier Path (Not the Better One)
It's worth pressing further on why "AI for All" reaches for equity and certification instead of purchase orders, because the mechanism explains a lot about how modern governments actually behave under fiscal and political constraint.
A purchase order creates an enforceable obligation with a deliverable and a date. If Ottawa signs a contract to buy a Canadian data-fusion platform for a named department by a named quarter, failure is visible: the platform either shipped or it didn't. That kind of commitment is politically risky for civil servants, because a failed procurement is a headline, while a failed grant program quietly disappears into next year's budget line.
An equity stake or a certification scheme defers the deliverable indefinitely. Money moves, press releases go out, and "supporting the ecosystem" becomes the measurable unit rather than "delivered working system." This is a well-documented pattern in industrial policy generally — not unique to Canada or AI. Compare it to how difficult it has been for the US or EU to convert semiconductor subsidy programs (CHIPS Act, EU Chips Act) into actual fabs on actual timelines, versus how straightforward it is to announce the subsidy itself.
The genuinely hard part is qualifying small vendors under existing security and liability frameworks. Palantir got into the Canadian government not because it lobbied harder than domestic alternatives (though it likely did), but because it could clear security questionnaires, indemnification requirements, and integration timelines that a 40-person Canadian startup usually cannot absorb without a government-scale legal and compliance team. This is the same "govtech is notoriously difficult to break into" problem raised on Hacker News — and it's a solvable procurement design problem, not a talent or technology gap. Streamlining vendor qualification for smaller firms, as the op-ed suggests, addresses the actual bottleneck; equity stakes do not.
The Deeper Pattern: Sovereignty Claims Without Sovereignty Audits
There's a broader principle worth naming, because it recurs across every country's AI strategy we've covered, not just Canada's: a sovereignty claim that hasn't been tested against an actual dependency audit is a marketing claim.
Compare three approaches:
Canada, as described here: announces anchor-customer ambitions while running undisclosed multi-year contracts with a US vendor inside Defence and provincial policing — the two domains where sovereignty arguably matters most.
Switzerland's Apertus, covered in our sovereign foundation model piece, took the harder route: publishing training data provenance and reproducibility so regulated users can actually answer compliance questions, rather than asserting sovereignty as a brand attribute.
The EU AI Act regime, in our Europe AI landscape guide, backs sovereignty rhetoric with binding enforcement dates and audit obligations — imperfectly, but with teeth.
The distinguishing test isn't whether a country announces a sovereign AI strategy — nearly every government has by mid-2026. It's whether the strategy comes with a mechanism that would surface a Palantir-style undisclosed contract before an opposition MP has to force it into the open. Canada's "AI for All," as written, does not.
Our Read: Procurement Is the Whole Game
Strip away the Canada-specific details and this is a general lesson for every sovereign-AI push, from the EU AI Act regime to national compute funds:
Demand beats subsidy. A government that buys a product creates a company; a government that grants money creates a grant-writing department. The proposed fixes — source lists usable by sub-50-person companies, a fixed floor of departmental AI budgets for domestically controlled products, quarterly published receipts — are procurement plumbing, not vision statements. That's why they'd work.
Transparency is the sovereignty test. The scandal isn't that Canada bought Palantir; it's that the contracts were undisclosed until forced out. Sovereign systems bought in secret reproduce the accountability problem they claim to solve.
Declare the conflict. Vigier runs a company that builds exactly what the strategy claims to want, and says so in the piece — "weigh my motives accordingly." Notably, he asks for open purchasing rules, not funding. That's the right ask even from a self-interested source.
Contract figures and policy details reflect the op-ed and public reporting as of July 6, 2026. Canadian procurement records may be updated; verify against the Government of Canada's proactive disclosure databases.
