Last updated on April 28th, 2026 at 09:06 am
Canada’s AI compute strategy keeps shipping announcements. The strategy itself remains in draft.
On April 15, the federal government opened applications for the AI Sovereign Compute Infrastructure Program (SCIP), the centrepiece of the $2.4 billion Canadian Sovereign AI Compute Strategy first announced in Budget 2024 and reinforced in Budget 2025. Applications close June 1. The release describes the program as a foundation for Canada’s “digital backbone” and a guarantor of “technological sovereignty.”
It is worth taking that claim seriously, because the country has been waiting on a promised refreshed national AI strategy for a long time, and the gap between what gets announced and what gets shipped has become its own data point. Four gaps – no model, no builder, compromised compute counterparties, incumbent foreign software vendor – are the likely reasons why the refreshed AI strategy hasn’t shipped.
The SCIP announcement is the third major communication of the same money. Budget 2024 announced $2 billion. December 2024 launched the strategy that allocates it. February 2025 issued the call for statements of interest. April 2026 opens the actual application window. A program selected from these applications will then need to negotiate, contract, site, power, provision, commission, and operate a frontier-class cluster. A realistic in-service date is late 2027, more plausibly 2028. By then the hardware generation referenced in the application will be at least one cycle old, and the global compute frontier will have moved to whatever comes after Blackwell. Canada is provisioning, in 2026, for a frontier that will not exist by the time the racks are hot.
That is a sovereignty problem, in addition to a procurement problem, because sovereignty at the compute layer is a function of timing as much as ownership.
The deeper issue is that the word “sovereign” appears throughout the strategy without a layer-specific definition. Sovereign at which layer? Hardware? The chips will be NVIDIA, fabbed in Taiwan, subject to US export controls that have already been tightened twice in 24 months. Cloud control plane? Operator nationality? Foundation model? Workload priority during a geopolitical squeeze? The strategy does not answer these, and the SCIP application materials defer them to negotiation. “All funding amounts are notional and subject to negotiation” is the actual line on the program page.
The Cohere announcement from earlier in the strategy makes the structural problem visible. Cohere received a commitment of up to $240 million to build a multibillion-dollar Canadian data centre in partnership with CoreWeave. CoreWeave is a US-listed neocloud whose business model depends on a small number of hyperscaler tenants, whose unit economics are inverted at current GPU depreciation rates, and whose acquisition by a hyperscaler is a recurring market rumour. Anchoring Canadian sovereign compute on a CoreWeave partnership means Canada’s sovereignty posture inherits CoreWeave’s counterparty risk. If CoreWeave is acquired or restructured, the control plane sitting on top of Canada’s flagship sovereign AI investment changes hands in a transaction Canada has no vote in. This is a scenario any Layer 5 sovereignty audit produces in its first iteration.
The Palantir issue sits in the same room and is rarely named. Palantir Canada is a pre-approved federal AI vendor with contracts that extended into 2028, with an executive bench drawn from former senior Canadian officials, and with a parent company whose US federal contract base nearly doubled to roughly $970 million in 2025. A national AI strategy that builds public compute capacity while leaving the public sector’s AI software stack open to a vendor whose business model is structural integration into government workflows has not solved the sovereignty problem. Canada cannot build sovereign AI on top of a federal software layer that is already owned, operated, and updated from Denver.
The SCIP announcement does not engage with Bill C-22, the Lawful Access Act, tabled March 12, 2026, which is moving through the House in parallel with this procurement. C-22 compels telecom and electronic service providers designated as “core providers” to build and maintain technical capabilities for government access to subscriber and transmission data, and authorizes regulations requiring metadata retention for up to one year. A sovereign compute build operated by a Canadian entity, sitting on Canadian soil, under a statutory regime that compels it to maintain government access pathways and retain metadata on its users, is a different sovereignty posture than the one the SCIP release implies. The strategy defines sovereignty as Canadian ownership and Canadian location. C-22 defines Canadian infrastructure as lawful-access-ready by design. Those two definitions are in the same policy environment and produce different compute. The strategy does not say which one the winning SCIP applicant will be built to.
Canada does not have a frontier model. We do not currently have the capital stack, the talent retention, or the compute lead time to build one. Cohere is the closest candidate and is operating in a financing environment where every independent foundation model lab outside the hyperscaler perimeter is under pressure. Cohere’s models are suitable for specific enterprise and government tasks like document search, summarization, retrieval-augmented workflows, and internal knowledge tools, but they are not at the capability level required for general-purpose public-facing assistants, agentic systems, or the kind of reasoning-heavy workloads that define frontier use, meaning a sovereign compute build anchored on Cohere gives Canada a credible enterprise stack and a foundation for classified or sensitive internal government workflows, but it does not give Canada a domestic answer to ChatGPT, Claude, or Gemini for citizen-facing services, and any general public deployment at that tier still routes through American models.
Building national supercomputing capacity without a credible national model layer to run on it produces an asset whose anchor tenant has to be either a US hyperscaler reselling capacity, a US-headquartered model lab buying inference, or a Canadian research consortium too small to amortize the build. None of those outcomes is sovereignty. Two of them are subsidy.
This is the coherence (no pun intended, but present) problem the refreshed strategy keeps colliding with. You cannot ship a sovereign AI strategy without resolving, in writing, the relationship between the compute layer, the model layer, the application layer, and the procurement layer. The current approach treats compute as if it were the foundation that makes the other layers possible, when in practice it is the layer most exposed to every dependency the strategy is supposed to address.
A serious version of the announcement would name a compute target in exaFLOPs, a power envelope, a delivery date tied to a specified hardware generation, an anchor research or model-layer tenant, an allocation policy that prioritizes Canadian-controlled workloads, a position on US export-control exposure, and a clear answer to whether US hyperscalers can be paying customers of the resulting infrastructure. It would also name the model-layer commitment the compute is meant to serve, because compute without a model strategy is a real estate decision dressed in industrial policy language.
None of this is an argument against building. Canada needs the capacity. The argument is that calling the build “sovereign” without the supporting commitments turns the word into a marketing skin over a procurement decision, and procurement decisions made under that label tend to lock in the very dependencies they were announced to address.
The country has been promised a refreshed AI strategy for two years. SCIP is not that strategy. It is a funding mechanism with a sovereignty label attached, shipped on the calendar of an application deadline rather than the calendar of a coherent national position. The harder document, the one that names the model layer, the procurement perimeter, the export-control posture, and the relationship between public compute and the private vendors already embedded in federal workflows, is still waiting to be written.
It is the document that would actually be worth opening applications for.
Jennifer Evans is the principal of Pattern Pulse AI and a cofounder of Tech Reset Canada