Nick Frosst on AI
From Sovereignty Stack — Cohere, Kepler, Dominion Dynamics, Build Canada · · Toronto Tech Week
“For me, the right way is like open. It's one in which the models are released for people to use and to run locally. It's one in which it can be deployed privately and on prem so that people can run it on real data and it can be controlled by that organization or person that's using it and it can't be shut off by whoever made the model.”
On , Nick Frosst, Cofounder and President at Cohere, spoke about AI during Sovereignty Stack — Cohere, Kepler, Dominion Dynamics, Build Canada on Toronto Tech Week.
About Nick Frosst
Nick Frosst, co-founder and president of Cohere, has been speaking publicly about AI sovereignty, enterprise AI, and the company’s strategic bets. In a May 2026 panel on sovereignty, Frosst described sovereignty as “agency and autonomy” that can apply at multiple scales—individual, organizational, and national. He argued that buying Canadian technology is necessary to generate incentive systems for world-class products, and expressed support for mandating that Canadian pension funds invest in Canadian companies. Frosst also noted that Cohere has raised the majority of its capital from outside Canada, but has worked with Canadian investors such as PSP, Radical, and Anovia. He characterized Cohere’s decision to release an Apache 2.0 open-weights model as a bet that the company’s value lies in its deployment, agentic framework, and search stack rather than the model alone. In a series of podcast appearances from March and April 2026, Frosst emphasized Cohere’s focus on enterprise customers and pragmatic AI, contrasting it with companies that he said are “banking your finances on the creation of AGI.” He stated that Cohere is “building ROI, not AGI” and that the company has trained competitive models with “1/100th of the compute of our competitors.” Frosst described Cohere as unique in its singular focus on enterprise, its non-American and non-Chinese origin, and its approach to providing sovereignty for users. He predicted that enterprise AI will increasingly involve workers directing language models to perform tasks, rather than writing code themselves.