Canada committed $2 billion to owning its own AI compute. The hard part is what runs on it

Canada has a long-standing and genuinely unusual position in artificial intelligence. The foundational work on deep learning was done substantially in Canadian universities, sustained by public research funding through a period when the field was unfashionable, and the country built three national institutes around that work: Mila in Montreal, the Vector Institute in Toronto and Amii in Edmonton. By the ordinary measures of research output, Canada punches far above a country of its size.

It also, for most of the past decade, had almost nowhere to run the resulting models. Training a large model requires concentrated access to specialised accelerators, and the overwhelming majority of that hardware sits in facilities owned by American hyperscalers. Canadian researchers and Canadian companies have largely rented it. That arrangement worked while capacity was plentiful and cheap. It stopped working when it was neither.

The federal response is the Canadian Sovereign AI Compute Strategy, announced in December 2024 with up to $2 billion attached. It is worth reading as three separate programs rather than one, because they operate on completely different timescales and carry completely different risks.

The first is an AI Compute Access Fund, in the range of $300 million, which subsidises access to existing compute for Canadian companies. This is the only component that does anything immediately. It does not build capacity, it buys it, and it is essentially a bridge for firms that need cycles now and cannot wait for construction. The criticism of it is obvious and partly fair: public money flowing to foreign cloud providers is not sovereignty in any strict sense. The defence is also fair, which is that a Canadian company priced out of compute in 2026 will not exist to use a domestic facility in 2029.

The second is the Canadian AI Sovereign Compute Infrastructure Program, roughly $700 million, aimed at helping the private sector build AI data centres on Canadian soil. This is the component that most resembles industrial policy in the conventional sense, and it is where the strategy touches the physical constraints that dominate coverage in Alberta. A data centre needs power, transmission, cooling and land, and it needs them at a scale and on a timeline that provincial systems have to accommodate. The federal government can fund a facility. It cannot connect one.

The third is approximately $1 billion for public supercomputing infrastructure serving researchers. This is the least commercially visible and arguably the most durable. Academic compute is what produced the Canadian AI research position in the first place, and it has been persistently underfunded relative to the ambition attached to it. Researchers who cannot get allocations do smaller experiments, or they take positions at institutions abroad that can offer them, and the second outcome is how a research advantage quietly ends.

The word doing the most work in the strategy is sovereign, and it is worth being precise about what it can and cannot mean. It plainly does not mean a domestic supply chain: the accelerators are designed in the United States and fabricated in Taiwan and South Korea, and no Canadian program changes that. It does not mean domestic ownership either, since the infrastructure program funds private builders who may well be foreign-owned.

What it can mean is jurisdiction. A facility on Canadian soil operating under Canadian law puts the data, the workloads and the legal process governing both inside the country. For health data, government workloads, defence-adjacent research and anything touching personal information under Canadian privacy law, that distinction is real and has practical consequences. It is a narrower claim than the branding suggests, and it is the claim the program can actually deliver.

The risk that deserves the most scrutiny is not that the facilities fail to get built. It is that they get built and sit underused. Compute infrastructure is only valuable in proportion to the demand for it, and demand comes from companies with models worth training and budgets to train them. Canada has a well-documented weakness in exactly that layer: strong research, capable early-stage companies, and a thin population of firms large enough to consume industrial-scale compute. A country can fund capacity into existence. It cannot fund demand into existence the same way.

There is a related question about what kind of compute is being built. Training runs and inference workloads have different economics, different hardware profiles and different geographic logic. Training is bursty, tolerant of latency and drawn to cheap power, which is an argument for Alberta and for Quebec. Inference is continuous, latency-sensitive and drawn to population centres. A strategy that funds the wrong mix ends up with facilities that are technically impressive and commercially awkward, and the published material has been considerably clearer about dollars than about which of these it is buying.

Energy is where the federal strategy and provincial reality meet most directly, and the meeting is not always smooth. Ottawa can announce a program. Whether a specific facility connects to the grid is decided by a provincial system operator under provincial rules, on a queue that in Alberta is currently oversubscribed by more than fifteen to one. Quebec, which has surplus hydro and has courted these projects, has its own constraints and has been increasingly selective about which loads it accepts. A national compute strategy that does not resolve who supplies the electricity is a national compute strategy with an unspecified dependency in the middle of it.

For Alberta specifically the strategy is an opportunity with a caveat attached. The province has the generation, the land, the cold climate and an explicit policy of courting the sector, which is a genuinely strong combination. It also has a grid queue that has forced developers to bring their own generation, which raises the capital cost of entry and favours very large well-capitalised builders. Federal money that lowers that barrier could shift which projects are viable. Federal money that arrives without the interconnection being solved simply moves the bottleneck one step earlier.

The measures worth tracking are unglamorous. How much of the access fund is actually drawn down, and by how many distinct companies rather than a handful of large ones. Whether infrastructure-program facilities reach financial close rather than announcement. Whether the academic allocation increases the number of researchers who can run large experiments in Canada, which is measurable and rarely reported. And whether any of the new capacity has a named anchor tenant, because a data centre without a committed buyer for its compute is a building, and the AI investment cycle has already produced a good number of those.