Cohere

Cohere builds large language models and sells them to enterprises, which sounds like a crowded position until you look at how it has chosen to differentiate. Rather than competing primarily on consumer-facing capability, the company has concentrated on deployment: models that can run inside a customer’s own cloud environment or on their premises, for organisations whose regulator, contracts or risk appetite will not permit sending data to an external service.

That constraint is not a niche. Banks, insurers, hospitals, defence contractors and government departments all operate under data-residency and confidentiality requirements that make an ordinary API integration a non-starter, and collectively they represent a very large share of the enterprise software market. Building for that segment means accepting harder engineering and longer sales cycles in exchange for a defensible position.

The company raised roughly US$600 million at a valuation near US$7 billion in 2025 and has published customer relationships including RBC, Bell, Dell, Thales, SAP and LG. It maintains a Toronto headquarters alongside offices in San Francisco, Palo Alto, London and New York, which is the pragmatic arrangement for a Canadian company competing for talent and customers in this category.

Its strategic significance to Canada exceeds its size. A country that consumes artificial intelligence built elsewhere is a customer; one that builds foundational models retains the research capacity, the senior engineering talent and the policy leverage that come with domestic capability. Cohere is the clearest instance of the latter, which is why its trajectory attracts attention beyond its commercial results.

The competitive position is demanding and worth stating plainly. Cohere is competing against companies with vastly larger capital bases and compute budgets, in a field where scale has historically translated into capability. Its response has been to compete on deployment, efficiency and enterprise fit rather than on raw model size, building models that are practical to run inside a customer’s own infrastructure rather than models that top public benchmarks. Whether that positioning holds depends on whether enterprise buyers continue to value control over frontier capability.

Retrieval-augmented generation has been central to that strategy. Rather than relying on what a model absorbed during training, the approach grounds responses in a customer’s own documents retrieved at query time. For enterprises the appeal is direct: answers can be traced to a source, the underlying knowledge updates without retraining, and confidential material never has to enter a training set. That is a considerably easier proposition to put in front of a risk committee.

The economics of the sector remain unresolved for everyone in it, and Cohere is no exception. Training and serving large models is expensive, competition has pushed inference prices down sharply, and the eventual distribution of margin between model developers, cloud providers and application builders is genuinely unclear. Nobody in this market has proven a durable model, and confident predictions about which layer captures the value should be treated with caution.

For Canadian technology policy the company is a live test of whether a domestic frontier AI capability can be sustained without the state intervention seen in other countries. The outcome will shape how Canada thinks about sovereignty in this category for a long time, which is a heavy burden for one company and not one it asked for.

The enterprise sale in this category is slower and more demanding than consumer adoption, and the difference is structural rather than cultural. A bank deploying a language model has to satisfy model risk management requirements, demonstrate that outputs can be explained and audited, establish what happens when the system is wrong, and obtain approval from a risk function whose job is to say no. That process takes quarters. It also produces contracts that do not churn, because an organisation that has been through it has no appetite to repeat the exercise with a competitor.

Multilingual capability has been a genuine area of differentiation rather than a checkbox. Much of the enterprise market sits outside English-speaking countries, and models trained predominantly on English perform noticeably worse elsewhere. Building strong performance across many languages opens markets that are underserved precisely because the largest competitors have optimised for their home market first.

The talent question is the one that determines whether the Canadian position holds. Senior machine learning researchers are among the most contested employees in the world, and compensation at the largest American laboratories is very difficult to match. Retaining that calibre of person in Toronto requires offering something other than the highest number, usually the ability to work on a distinctive problem with genuine ownership of the outcome. That is a real advantage while a company is ascendant and a fragile one if momentum falters.

For readers assessing the company, the useful indicators are contract renewals and expansions with named enterprise customers rather than benchmark results or valuation headlines. Benchmarks measure capability in the abstract. Renewals measure whether the technology survived contact with an organisation’s actual work, which is the only test that determines commercial durability.

All startups →Latest news →