This article appears in the 2026 BetaKit Most Ambitious issue. Read more stories of the Canadian tech innovators strengthening our autonomy, security, and prosperity here.
. . .
Nick Frosst thinks there are only a handful of countries in the world at the frontier of the AI race. The United States. China. France, where Mistral holds the flag alone. And Canada, which has chosen Cohere as its global champion. You can count them all without using your thumb.
“It’s a wildly complicated engineering problem,” said Frosst, one of Cohere’s three co-founders, of the technical capacity required to build large language models from scratch. “I think it’s also downplayed. This is really, really challenging. And you’ve seen massive companies spend huge amounts of money on it and just not…” He lets the sentence hang. The graveyard of failed attempts finishes it for him.
The Toronto-based company doesn’t make a consumer chatbot. What it builds are enterprise-grade language models designed to be deployed privately on a customer’s infrastructure. Most of the genuinely useful data in the economy is regulated and sensitive—the kind that drives decisions in financial services, telecom, or energy. The kind that can’t be sent over the wire to a data centre in Virginia. “LLMs are really only as useful as the data they have access to,” Frosst said. “Most of the real, useful data has regulations around it. It can’t just be sent to some random data centre somewhere.”
Cohere’s answer is on-premise deployment and Model Vault, a walled-off set of GPUs running inside the customer’s environment. CEO Aidan Gomez put it plainly at Davos: Cohere has no ability to observe data flowing through its models or shut them off. “We build it completely from scratch,” he explained. “We have no external dependencies other than chips.”
‘Protectionism in a nicer font’
The company’s Command A models run on just two GPUs, hitting a sweet spot between performance and what enterprises actually have access to. Frosst said the company has spent “orders of magnitude less” on training than some competitors. North, Cohere’s agentic platform, wraps it all into a workspace where knowledge workers hand off the tedious parts of their day to a model running entirely within their employer’s infrastructure.
“‘Sovereign AI’ gets thrown around a lot, and it’s easy to hear it as protectionism in a nicer font.” Frosst resets the term by reaching for an analogy that’s tied to, but distinct from, AI. “I often compare this to energy sovereignty,” he said. “Canada is a pretty energy sovereign nation. Ontario generates roughly 55 percent of its power from nuclear. That’s a sovereign capability. But we have to collaborate with other people to keep that running.”
Nuclear plants, wind turbines, and hydro all require international collaboration across everything from fuel to engineering to financing. But the finished product is sovereign: the lights stay on because of decisions made in Ontario, not because of the goodwill of a foreign government.
The pitch lands differently when it comes from the customer rather than the vendor. Michel Richer, president of Bell AI (a key Cohere partner and customer), says Bell’s understanding of data sovereignty has evolved dramatically over the past 14 months as the geopolitical landscape shifted. It now encompasses not just where data sits, but who controls the compute it runs on. “We want to ensure that no one from an external jurisdiction has access to the data, but also no one from an external jurisdiction can decide what runs and what doesn’t,” Richer said. Bell’s CEO has a line for it: “We’ll never put the off switch for Canada somewhere else.”
“A necessary component of working at Cohere is that you can’t think the technology is magic.”
Nick Frosst, Cohere
“Sovereignty doesn’t mean isolation,” Frosst said, crediting the line to AI minister Evan Solomon. “When I talk about sovereignty, really what I mean is autonomy and agency … None of that means walling yourself off.” A country doesn’t build its own power grid because it hates its neighbours: it builds one because when the sun goes down, you want the light switch to be in your own building.
The argument would be academic if nobody were buying. Gomez has said publicly that last year, Cohere’s revenue was roughly 80 percent US-concentrated; by early 2026, that had dropped to less than half, even as the business overall more than doubled. The growth is coming from everywhere else.
Bell signed its strategic partnership with Cohere in July 2025; nine weeks later, North was rolling out to Bell’s management team, with more than 100 use cases already identified. The company’s engineers are forward-deployed alongside Bell’s teams, working on custom models; it’s a capability Richer says most other LLM providers don’t offer enterprise customers.
In Europe, Cohere expanded its SAP partnership to deploy North on SAP’s Sovereign Cloud. In Japan, Fujitsu built Takane, a custom Japanese-language model, on Cohere’s platform. In South Korea, LG is doing the same in Korean. On the defence side, Cohere signed an MOU with Sweden’s Saab to embed AI into GlobalEye surveillance jets, partnered with Thales Canada, and is working with both ThyssenKrupp and Hanwha Oceans, the two shortlisted bidders for Canada’s next submarine fleet. By partnering with both, Cohere plays a role regardless of who wins the bid.
Gomez framed these target sectors (telco, financial services, energy, public sector) explicitly as national security concerns. The power grid, the communications grid, and every financial transaction are the load-bearing walls of the economy. And there’s a reason a Canadian company is helping stand up those walls for so many international allies.
“The Canadian flag is a good brand these days,” Frosst said. “People are excited to work with us.”
Cohere gave its strongest signal yet of that geographic rebalancing act in late April, announcing plans to merge with Aleph Alpha, the German enterprise AI company, in a deal valuing the combined entity at around $20 billion USD. The deal has direct support from both the Canadian and German governments and a clear strategy: Cohere would absorb its smaller counterpart and fast-track its presence in Europe’s largest economy and the EU as a whole.
Not AGI but ROI
Beyond its home country, Cohere is unique among global competitors in that its co-founders appear to believe something genuinely different about the technology. Frosst thinks OpenAI CEO Sam Altman’s rhetoric about approaching AGI was “academically disingenuous” and did “a disservice” to the field. His mantra is four words: not AGI but ROI. At a Vector Institute event this spring, he asked a room of AI researchers how many believed large language models posed an existential threat. Two hands went up. “They have fundamental limitations,” Frosst said. “I think those limitations are drastically underestimated.”
He’s not dismissing the technology; he’s arguing it’s being misapplied. Cohere trains its models on synthetic environments—fake companies, fake emails, fake internal APIs—all designed to teach a model how to be useful inside a business. “We’re not training it to be an amazing conversationalist,” he’s said previously. Cohere customers don’t ask the model to do abstract math reasoning; they ask it to file expenses and summarize contracts.
“A necessary component of working at Cohere is that you can’t think the technology is magic.”
Nick Frosst’s first job was making burgers. He and his co-founders still go to McDonald’s after every fundraising round, including the $500-million USD raise that brought Cohere’s valuation to $6.8 billion. He orders two Junior Chickens. He fronts a Juno-nominated band called Good Kid; its debut album just dropped. These are not the biographical details of someone interested in building, or claiming to be building, a digital god.
“We are a product of Canada,” Frosst said, and he meant it literally. “I went to public school. I went to university. I’ve had health care my whole life. We are a product of the government and the country that we grew up in. And so it’s a deep honour to be able to help in any way we can.” He pauses. “We build ourselves up by building each other up.”
A kid who grew up on Canadian public infrastructure is now trying to build his country’s AI infrastructure and, in turn, help other nations do the same. The scarcity is real, the opportunity is real, and the flag on the door still matters.
READ BETAKIT MOST AMBITIOUS 2026
In the 21st century, there is no sovereignty without technology.
Meet the Canadian innovators strengthening our nation’s autonomy, security, and prosperity in BetaKit Most Ambitious.
BetaKit is the official media partner of Toronto Tech Week.
Feature image courtesy Studio Wyse