Cohere told investors last week that annualized revenue crossed $210 million in July, roughly triple the figure it disclosed a year ago, and that more than 80 percent of it comes from deployments running on customer-controlled infrastructure rather than through its hosted API. The Toronto company spent 2023 and 2024 in an unwinnable position, building models that were competitive and not leading, sold through an API against competitors with far more capital. Its response was to stop competing on that axis. Cohere now sells complete deployments, model plus tooling plus support, installed inside a customer's data center or a national cloud, priced as an enterprise software contract.
The strategy in detail
A Cohere deployment includes its North platform, the Command model family, embedding and reranking models, and a set of connectors for enterprise data sources, delivered as a package that runs on the customer's own GPUs. Contracts are annual with a floor that starts around $1.2 million according to two customers, and pricing scales with deployed capacity rather than with tokens. That structure is familiar to any enterprise that has bought database or middleware software and completely unfamiliar to anyone who thinks of model providers as API companies.
The customers are institutions that cannot use a hosted service for regulatory or political reasons. Cohere named several: the government of Canada, a Japanese banking group, and Fujitsu as both a customer and a distribution partner. Others are unnamed and include, according to people familiar with the arrangements, a Gulf state sovereign fund, two European defense suppliers, and a large Canadian insurer.
Why this segment was available
OpenAI and Anthropic both sell primarily through APIs and through hyperscaler marketplaces, and both have been reluctant to ship weights to customer premises for reasons that combine commercial protection with genuine safety concerns. Microsoft offers Azure Government and sovereign arrangements that keep data in-country and still run on Microsoft-operated infrastructure. For a defense ministry that requires the model to run on hardware it owns in a building it controls, the options have been open weights plus internal engineering, or Cohere.
Meta's Llama and the Chinese open weight models occupy the first option and require the customer to build everything around them. That is entirely feasible for an organization with a strong engineering team and impossible for most government departments. Cohere's product is the assembly, the support contract, and the accountability, which is exactly what enterprise software has always sold.
The story is rarely the launch. It is what breaks, what ships, and who owns the mess at 2 a.m.
The model quality question
Cohere's Command models are not competitive with the frontier on general reasoning benchmarks and the company has stopped pretending otherwise. Its argument is that the tasks its customers run, retrieval-augmented question answering over internal documents, summarization, classification, and structured extraction, do not require frontier capability, and that a smaller model that runs on eight GPUs instead of sixty-four is the better engineering answer.
That argument holds for the workloads described and it constrains what customers can eventually do. An organization that deploys Cohere for document search and later wants agentic workflows requiring strong reasoning will find the model insufficient. Cohere's answer is a larger model shipping this autumn and a partnership structure that lets customers route specific requests to a hosted frontier model when policy permits, which several customers said they will never enable.
The economics compared to the API players
Enterprise software revenue is less capital-intensive than API revenue, because the customer buys the GPUs. Cohere does not need to build data centers, does not carry inference cost against revenue, and has gross margins that a person familiar with its finances described as above 70 percent, which is a different business from the API providers running at margins reportedly well below that on inference.
It is also slower and lumpier. Enterprise deals take nine to eighteen months, revenue recognizes over contract terms, and growth requires a field sales organization that costs money before it produces. Cohere has roughly 180 people in sales and solutions engineering against about 220 in research and engineering, a ratio that looks like an enterprise software company and not like a research lab.
Whether it lasts
The threat is that OpenAI or Anthropic decides this segment is worth serving. Both have made gestures in that direction, and Anthropic has an arrangement with Palantir that reaches classified environments. If either shipped a genuinely deployable on-premises product with support, Cohere's advantage would compress quickly to price and to the sovereignty argument that a Canadian company is not an American one.
That last point is doing more work than it did two years ago. Several European and Asian customers told us that a supplier outside US jurisdiction is now a procurement requirement rather than a preference, driven by uncertainty about American export policy and data access law. Cohere did not create that condition and it is the largest single factor in the company's improved position, which makes its future partly dependent on geopolitics rather than on engineering.
The other durability question is talent. Research staff join labs to work on frontier models, and a company that has explicitly decided not to compete at the frontier has a harder recruiting story regardless of compensation. Cohere lost several senior researchers to American labs in 2024 and 2025, and its current hiring emphasizes systems and deployment engineering rather than research. That is the correct staffing for the business it now runs. It also means the gap between Cohere's models and the frontier will widen rather than narrow, and the strategy has to keep working with that widening priced in.
Skarvonix will keep following this beat with reporting grounded in how systems behave outside the launch keynote.
- LLMs
