Canada lost the AI race it invented: Hinton, Bengio, Cohere

Canada lost the artificial intelligence race in 2013, at an auction, and almost nobody in the country noticed at the time.

The lot was DNNresearch, a University of Toronto spinout with no product, no revenue and three employees: Geoffrey Hinton, Ilya Sutskever and Alex Krizhevsky. A few months earlier the three of them had won the ImageNet competition with AlexNet by a margin so large it ended the argument about whether deep learning worked. Hinton ran the sale as a bidding war. Google won it for roughly $44 million. There was no Canadian bidder at the table, and there was never going to be one.

Everything that has happened since follows from that afternoon. Canada produced the science, sold the asset at a rounding-error price, and then spent thirteen years and several billion dollars subsidising the production of more science for other people to buy. That is not a tragedy of circumstance. It is a policy, and it is still the policy.

The three people who invented the field, and where they ended up

The modern machine learning field has three universally acknowledged founders, and they shared the 2018 Turing Award for it: Geoffrey Hinton, Yoshua Bengio and Yann LeCun. All three ran through Canada. Only one of them built anything durable here.

Hinton spent decades at the University of Toronto and won the 2024 Nobel Prize in Physics for the work he did there. His lab's commercial output went to Mountain View. LeCun did his postdoc under Hinton in Toronto in the late 1980s, then left for Bell Labs and eventually Meta, where he ran the research organisation that shaped a decade of open model releases. Bengio stayed, built Mila in Montreal into one of the largest deep learning research institutes on earth, and is the only one of the three whose institutional legacy is Canadian.

Then there is Ilya Sutskever, who matters more to this story than the founders do, because he is the clearest case of what Canada builds and does not keep. Sutskever moved to Canada as a teenager and took all three of his degrees at the University of Toronto under Hinton. He co-authored AlexNet. He went to Google with the DNNresearch sale, left in 2015 to co-found OpenAI, served as its chief scientist through the run that produced GPT-4, and departed in 2024 to start Safe Superintelligence Inc.

SSI is registered as an Israeli-American company with offices in Palo Alto and Tel Aviv. A man who received his entire technical education in Toronto founded one of the most consequential AI labs of the decade, and it did not occur to anyone that it might be headquartered in the city that trained him. Nobody made a mistake. Canada simply had nothing to offer that the alternatives did not offer more of.

The gap, in one number

The scale of the leakage is now documented rather than anecdotal. Canada hosts roughly 10 percent of the world's top AI researchers and captures under 2 percent of global AI venture investment, which means the country wins about one venture dollar for every five its research standing should command. That five-to-one gap is the entire Canadian AI problem expressed as a ratio.

The supporting figures are worse than the headline. Analysis published by Mila and Bain in January 2026 found that two-thirds of high-potential Canadian-led startups raising more than $1 million did so from headquarters outside Canada. A study of STEM graduates from Toronto, Waterloo and UBC found one in four working abroad, rising to 66 percent among software engineering graduates. In January, Y Combinator briefly told founders it would stop investing in Canadian-incorporated companies at all, and while it reversed the position within days, the reversal did not change the underlying fact that the Canadian YC alumni who became unicorns had almost all already flipped to Delaware.

Two-thirds of the good companies and two-thirds of the good engineers do not leave because of the weather. They leave because the country built one half of a machine.

Canada funds the half that does not capture value

Look at what the money actually bought. The Pan-Canadian AI Strategy has directed more than half a billion dollars into the ecosystem since 2017, which produced Mila, the Vector Institute and Amii, and those institutes worked. They are genuinely world class. The $2 billion Canadian Sovereign AI Compute Strategy followed in 2024, and Budget 2025 added roughly $926 million over five years for more sovereign compute.

Every one of those dollars sits upstream of the point where value is captured. Research funding produces papers and graduates, both of which are internationally mobile by design. Compute funding produces racks, which are useful and which any country with electricity can buy. Neither instrument creates a Canadian owner of a Canadian AI company, and ownership is the only thing that converts a research advantage into national wealth.

The missing piece is the commercial middle: the stage where a prototype finds its first paying customer and the founder decides which country's tax and capital regime to live under. Canada has almost no policy there, and the policies it does have are outgunned by a US venture market that will fund the same founder faster, at a higher valuation, with a customer base attached.

The compute strategy is the most revealing part, because compute is the easiest thing on the list to announce and the least differentiating thing to own. Sovereignty framed as data centres is sovereignty defined by the one input that is purchasable. The inputs that are not purchasable, meaning founders who stay and capital that will write a $200 million cheque without demanding relocation, receive a fraction of the attention.

Cohere is not the counterexample, it is the illustration

The obvious objection is Cohere, and it is a real one. Toronto has one of the very few credible foundation model companies outside the United States and China, founded by Aidan Gomez, Ivan Zhang and Nick Frosst out of the same Toronto research community, valued at $7 billion in late 2025 with roughly $240 million in annualised revenue by 2026. That is a genuine achievement and it deserves more credit than Canadian commentary usually gives it.

Then read what happened on April 24, 2026. Cohere announced a merger with Germany's Aleph Alpha at a combined valuation near $20 billion, anchored by a $600 million commitment from the Schwarz Group, and announced it in Berlin. Cohere shareholders keep about 90 percent of the combined company, so this is a Canadian-led deal rather than a takeover, and the strategy is sound: the market for a non-American sovereign AI vendor is in European regulated industry, and that is where the capital was.

That is the point. Canada's single AI champion had to go to a German retail conglomerate to find an industrial partner willing to underwrite it at scale, because no Canadian institution of comparable size was willing. The country's pension funds collectively manage well over a trillion dollars. They were not the ones in Berlin.

The optimistic case, and why most of it does not survive contact

Three counterarguments get made, and one of them is right.

The first is that brain drain is a misreading, and that a research diaspora is an asset: Canadian-trained people run labs everywhere, which buys influence and eventual return migration. Diaspora strategies do work, but they work for countries that hold equity in what the diaspora builds. Canada holds none. Influence without ownership is prestige, and prestige does not fund a hospital.

The second is that the sovereign compute buildout changes the equation. It does not, on its own. Cohere's own capacity expansion runs on CoreWeave, and the domestic buildout involving Bell, Telus and the Canadian data centre consortium is real infrastructure that will lower costs at the margin. Cheaper compute makes Canadian companies slightly more competitive. It does not make Canadian investors braver, and bravery is the binding constraint.

The third argument is the one that holds. The research base was not a waste, and treating it as a sunk failure would be the actual unforced error. Bengio is the proof: he stayed, built Mila, and in June 2025 launched LawZero, a nonprofit with roughly $30 million to pursue non-agentic safe-by-design systems, incubated at Mila and staffed with more than 15 researchers. He could raise that money anywhere on earth and he raised it for a Montreal institution. Safety and evaluation work is a domain where Canada has genuine standing, where the American labs face a structural credibility problem, and where the eventual regulatory market is enormous. That is a real lane, and Canada is early in it rather than late.

What would actually work

Stop funding supply and start funding demand. The single highest-leverage instrument available is procurement: federal and provincial governments buying Canadian AI systems at scale, on contracts long enough to be financeable, which is precisely how the United States built its defence technology base and how Schwarz Group is now underwriting Aleph Alpha. A grant makes a company survive a year. A contract makes it worth owning.

Second, take equity. When public money creates a company, the public should hold a stake in it rather than a press release about it. Third, make the pension funds' domestic technology allocation an explicit policy target rather than a polite suggestion.

Canada did not lose the AI race because it lacked talent, ideas or money. It lost because it decided, repeatedly and at every branch point since 2013, to be the country that trains the people rather than the country that owns the companies. That is a choice, which means it can be unmade, and the fastest way to unmake it is to start buying instead of granting.

Cover photo: Toronto skyline, March 2024. Public domain (CC0) via Wikimedia Commons.