Inside Cohere’s $20 Billion Sovereignty Bet, With Co-Founder Nick Frosst

On a Friday afternoon in mid-June, a letter arrived at Anthropic from the U.S. Commerce Department, and by evening, two of the most capable A.I. models on earth had gone dark. The directive, citing national security, barred every foreign national—including the company’s own non-citizen employees—from touching its newest systems, forcing Anthropic to pull Fable 5 and Mythos 5 offline all users at once. For most of the industry, it landed as a shock. For Nick Frosst, it was a slide he’d been presenting for years, finally rendered in the one medium the technology business respects: a live outage.

“It really solidified—made immediate and tangible—what we’ve been saying for a long time,” Frosst told Observer. “If this technology is foundational, and it is, and it’s not deployed under a country’s jurisdiction, that’s a huge risk. A company or a country could just decide to turn off access one day.”

Commerce lifted the controls on June 30, and Anthropic began restoring access the next day.

In 2019, Frosst, Aidan Gomez and Ivan Zhang co-founded Cohere, the Toronto company now routinely described as Canada’s national A.I. champion. He is the rare frontier-model executive who has spent the boom telling anyone who will listen that the destination everyone is pricing in—artificial general intelligence, the machine that thinks like a person—is not, in his estimation, on the map. Frosst learned the field at the feet of Geoffrey Hinton, who hired him as one of his first researchers at Google Brain in Toronto in 2016. He fronts an indie rock band with millions of monthly listeners and refuses to let a machine write a single lyric.

In September 2025, Cohere was valued at $7 billion. Today, Cohere signed a definitive merger agreement with Germany’s Aleph Alpha that values the combined company at roughly $20 billion, and Schwarz Group, the German retail conglomerate behind Lidl, has committed $600 million to Cohere’s upcoming Series E. The companies are pitching the combination as the first transatlantic sovereign A.I. provider, built for governments and regulated industries that want capable models without ceding control. “No government or enterprise should have to choose between capable A.I. and control over their tech,” Gomez said in announcing the deal. “That belief is exactly why we’re joining forces…Together, we’ll meet the rising global demand for frontier A.I. that’s both powerful and secure.”

A word for the moment

In conversations about A.I., “sovereignty” is having its season—the sudden vogue of a term that, until recently, was nobody’s daily vocabulary. Frosst would like to rescue it from the geopoliticians. To him, the word means autonomy, agency and control.

“It applies to a country, it applies to an organization, and it applies to a person,” he said. For the individual, the test is simple: does the technology empower you, or does it weigh you down? “A lot of people feel like technology isn’t enabling these days. They feel like it’s disempowering—fighting to keep their attention, fighting to keep them engaged. That’s not sovereignty-enabling. That’s sovereignty-removing.” For the enterprise, sovereignty means a model deployed inside the customer’s own walls, connected to the customer’s own data, with no vendor positioned to read it, train on it, or revoke it. For the country, it means the same arrangement, scaled to a jurisdiction, so that no foreign capital can switch off the hospitals and the banks with an afternoon’s notice.

This is not an abstraction Cohere arrived at after the fact. The company sells its models for private, on-premises deployment and has been open-sourcing them under permissive licenses—the better, Frosst argues, for a person or institution to own the thing outright rather than rent it. When the Mythos directive hit, Cohere’s chief A.I. officer, the former Meta research chief Joelle Pineau, reported a “huge number of inbounds” from companies and governments suddenly alert to the cost of dependency.

The protégé who dissents

If there is sentiment in Frosst, it surfaces when the conversation turns to Hinton, Professor Emeritus at the University of Toronto, so-called godfather of A.I. and the man whose name, Frosst notes, “will be written next to this technology for hundreds of years.” Frosst calls him a mentor and a dear friend, “creative and playful and kind and funny and smart,” and credits him with everything he knows about the discipline. He also disagrees with him in public.

Hinton has spent recent years warning that the technology may carry us toward something genuinely human-like, and toward genuine danger. Frosst thinks the warnings are worth having, but the timeline is wrong. “I do not think the technology we have will get us to AGI anytime soon,” he said. Large language models are, in his telling, a transformative tool that is “distinctly not human-like,” and making them better doesn’t push them toward general intelligence so much as toward what they already do well. He points to programming, the showcase application: the tools have automated the easy parts, and yet the industry, his own company included, keeps hiring more engineers, not fewer, because the hard parts (architecture, judgment) still need a person. He expects the broader labor story to rhyme; a technology that “automates about some percentage of everybody’s job, as opposed to a full percentage of anybody’s job”

Toronto has long been an academic epicenter of A.I. research, with the University of Toronto home to some of the field’s top researchers today. This summer, when asked whether Toronto could be the next Silicon Valley, Frosst dismissed the comparison. “Let Silicon Valley be Silicon Valley,” he said. The better question is what Toronto, or Zurich, intends to become on its own terms. Canada has long been brilliant at training minds and then exporting them; last year’s hiring of Pineau, an Ottawa native, poached from Meta’s research lab, is part of a reversal Frosst is happy to narrate. He wants Toronto—”ingenious, inclusive, multicultural, pragmatic, bold”—to commercialize without losing its character, to be more itself rather than a colder imitation of somewhere else.

What the machine doesn’t get to do

For all that, the most revealing answers came when the talk turned to music. Frosst is the frontman of Good Kid, the Toronto band of programmers he started at university. He tours a few weeks a year and works from a room at the back of the bus—”I do some work, play some music with my friends, go to sleep, wake up in another city, do it again.”

He uses A.I. all day for work—drafting, fact-checking, querying the state of his own company—and runs a small open model at home that controls the thermostat and checks the temperature of his kiln, a hobby he cheerfully admits is unnecessary. (“The thermostat works just fine. It’s just enjoyable.”) What he will not do is hand the machine the part that matters. “When I do art, the purpose is to see and be seen, to express myself. As a listener, the purpose is to see yourself reflected in what somebody else is expressing. I’ve never found A.I. to help me do either of those things.”

He has watched the flood of A.I.-generated tracks crest and recede. “Many months ago, there were stories about generated artists trending on Spotify. None of them had any cultural impact—a flash in the pan, no staying power, because ultimately they’re saying nothing. That’s not why we make or listen to art.” Offered, hypothetically, an album conjured from a text prompt, he was courteous and immovable: “I probably won’t listen to your Suno albums. I’m sorry.”

The instinct that runs through him is the refusal to surrender the thing that confers control, whether that thing is your data, your country’s infrastructure or your own voice. The last book that moved him, he said, was The Dawn of Everything, which argues that human societies did not march in a straight line from one stage to the next but adopted and discarded technologies as they found them useful. Frosst seems to be betting his company and his life on the conviction that we still get to choose. In June, Washington needed one Friday afternoon to take that choice away, and 19 days to give it back.