Cohere Is Canada’s Biggest AI Hope. Why Is It So American?

With more than 400 employees and valued at $6.8 billion, Cohere is one of Canada’s leading artificial intelligence companies. At first glance, it seems perfectly poised to fulfill the federal government’s goals of keeping up with the global AI race and for achieving tech sovereignty from the United States.

Unlike its competitors—OpenAI, Microsoft, and Anthropic—Cohere doesn’t offer a flagship consumer chatbot like ChatGPT; instead, it builds large language models and AI programs for businesses and governments. Over the summer, Cohere signed agreements with the governments of Canada and the United Kingdom to incorporate AI into the public sector. It also announced a partnership with Bell Canada, whereby Cohere’s AI infrastructure will be available to Bell’s customers, and Bell will become Cohere’s “preferred Canadian AI infrastructure provider.”

Launched in 2019, the company was built by homegrown talent: co-founder and chief executive officer Aiden Gomez, who completed his undergrad at the University of Toronto, previously interned for AI pioneer Geoffrey Hinton and Lukasz Kaiser at Google Brain, and in 2017, co-authored an influential paper that helped pave the way for the creation of LLMs. And for now, the company seems intent on staying Canadian. At a conference in June, Gomez—along with leaders from Wealthsimple and Shopify—insisted that, in order for the country’s tech sector to grow, its entrepreneurs had to resist the lure of US acquisition.

But look beyond those qualifications, and Cohere starts to lose its sheen. For one, the company leans heavily on American tech. In 2024, Cohere received $240 million in federal funding to support building a $725 million data centre in Cambridge, Ontario. The funding is part of the $2 billion Canadian Sovereign AI Compute Strategy aimed at increasing the country’s national capacity to develop AI products and support AI infrastructure. Cohere will reportedly use the funding to partner with CoreWeave, a US AI infrastructure company, to build the data centre. CoreWeave is a popular choice for US companies, including OpenAI and Microsoft, as it provides the much-needed data processing for AI to work. Geoff Gordon, CEO of Denvr Dataworks, an AI cloud platform based in Calgary, told The Logic earlier this year that he was profoundly disappointed that the federal funding would “flow to a US company, which is just ridiculous.” Cohere co-founder Nick Frosst told the publication he hoped to see Canadian companies offering the same type of infrastructure someday.

The government contract didn’t require Cohere to choose a Canadian supplier, though there are options. ThinkOn, an AI company that stores Canadian federal data, owns eleven data centres in the country and eighteen other centres in the US, UK, and Australia. Another alternative, Cerio, works with companies to create energy-efficient data centres. Yet neither of these companies use the same technology as CoreWeave, which has access to the popular graphic processing units used to power AI data centres through its partnership with US chip manufacturer Nvidia. CoreWeave’s infrastructure underpins some of the world’s largest AI operations—including those of OpenAI, Microsoft, and Meta—supported by multi-billion-dollar agreements. Likely faced with a choice between investing in Canadian companies with hopes of long-term payoff or using American tech to get quick results, Cohere seems to have bet on the latter.

Some of Cohere’s other associations can’t be explained away as easily. Cohere has a relationship with Palantir, a US multi-billion-dollar data analytics company with a history of partnering with US military and intelligence agencies. Working with Palantir would give Cohere access to advanced AI software and US government clients. At the 2024 DevCon1 conference, a Cohere representative described how the company serves Palantir customers by creating custom AI models for them. He didn’t specify who the clients are.

For critics, this is a concern given Palantir’s infamy. It was founded by Peter Thiel, an avid supporter of US president Donald Trump, and a former mentor to Vice President J. D. Vance. The company has been criticized for, among other things, creating a program to help the US Immigration and Customs Enforcement identify personal information to target immigrants for deportation. In 2018, more than 200 Palantir employees signed a petition asking for the ICE program to be suspended but were unsuccessful. Palantir signed another $30 million contract this year to track people who overstay their visas or leave voluntarily.

These actions don’t directly implicate Cohere. But neither Palantir’s roots nor its current activity seem to bother Cohere’s leadership. Frosst commented on the relationship in The Logic, mentioning that he’s unfamiliar with Palantir’s work, but that he’s comfortable with any company using Cohere’s tech as long as they follow its terms of use. The terms include prohibiting deployment of the tech to store or distribute harmful material or to create weapons.

And yet those terms of use could mean little considering that Cohere’s tech itself may have been built using questionable practices. This year, a group of fourteen publishers, that includes the Toronto Star and US-based publishers Condé Nast and Forbes, sued Cohere for using their articles without permission to train AI systems. (Microsoft and OpenAI are facing similar copyright infringement lawsuits.) The media companies allege that Cohere’s systems created hallucinations: fake articles that were attributed to the publishers. In a statement in TechCrunch, Cohere director of communications Josh Gartner has said the lawsuit is “misguided and frivolous,” and in May, Cohere filed for the case to be dismissed. A US court denied the motion in a ruling last Thursday.

Either outcome would have been worrisome. If the case proceeds, it could cast doubt on the integrity of the federal government’s chosen AI provider. If it doesn’t, that would raise the question of how AI darlings, like Cohere, might be held accountable in the future. Because as of this writing, the government has yet to deliver national AI regulations that would, in part, protect Canadians’ data.

To avoid a potentially problematic bind, the federal government could take its cue from the European Union’s AI Act, which creates copyright and privacy protection for user data. Otherwise, in forging ahead with its Cohere partnership for the sake of prioritizing AI adoption, the government could be risking the very data sovereignty it’s trying to defend.