Headlines about AI usually focus on chatbots and consumer apps, but a large share of the money is made by selling models and tools to businesses. Cohere, a company focused on that enterprise market, reportedly generated about $240 million in revenue over a year, a figure that sets the stage for talk of a stock market listing. The number deserves context, because a revenue figure alone says little without knowing how it was earned.
What Cohere sells
Unlike companies that built their reputation on consumer chat products, Cohere targets corporations, governments and institutions that need language AI inside their own workflows. Its offerings typically include models for generating and summarizing text, tools for searching company documents, and options to deploy within a customer’s private cloud or data center for privacy and compliance. That emphasis on control appeals to banks, healthcare organizations and public agencies.
How to read a revenue number
- Growth rate. Revenue doubling year on year tells a different story from revenue rising ten percent.
- Recurring versus one-off. Subscription and usage contracts are more valuable than one-time projects.
- Customer concentration. If a few customers account for most revenue, the business is riskier.
- Gross margin. Running models is expensive; what matters is how much is left after computing costs.
- Retention. Do customers renew and spend more over time?
Why an IPO story needs more than revenue
Public investors will ask whether the company can reach profitability, how it competes against giants that can subsidize prices and whether its technology stays ahead as models improve. An AI company must show it can grow without burning cash on endless training runs. It must also explain its relationships with cloud providers and chip suppliers, which can affect costs and independence.
The enterprise AI market
Corporate buyers behave differently from consumers. They run pilots, demand security reviews, negotiate contracts and move slowly, but once a system is embedded in operations, they tend to stay. Success depends on integration with existing software, support quality and demonstrated returns, such as reduced handling time in customer service or faster document review. The market is crowded with offerings from large technology firms, specialist startups and open-source models that customers can run themselves.
What differentiates a model vendor
- Data privacy. Options to keep data inside the customer’s boundary.
- Customization. Fine-tuning for industry language and internal knowledge.
- Language coverage. Strong performance beyond English for global firms.
- Cost efficiency. Smaller, cheaper models that handle common tasks well.
- Trust and governance. Audit trails, safety controls and clear responsibility for outputs.
Risks for investors and customers
Technology can change quickly. A model that leads this year may be matched by an open-source release next year. Pricing pressure is intense, as large providers cut costs to win market share. Customers face lock-in risks if they build deeply on one vendor’s tools, so many choose architectures that can swap models. Regulation adds another layer, with rules on transparency and safety emerging in several regions.
A worked example
Consider a large insurer that wants to speed claims processing. It pilots a model to summarize claim documents and draft responses, with a human adjuster reviewing each output. After a successful trial, it expands to a hundred adjusters and pays a usage-based fee. If handling time falls by twenty percent, the savings can easily exceed the software cost. That kind of measurable return is what turns pilots into the recurring revenue that public investors want to see.
Questions a buyer should ask a model vendor
Ask where your data is processed and whether it is used to train models for other customers. Request clear service levels for uptime and response time, and find out how pricing changes as your usage grows. Ask for a roadmap, but weigh it lightly; what matters is what works today. Finally, run your own side-by-side test with real documents against two or three alternatives, because a vendor demo and a production workload behave very differently.
The takeaway
A revenue year in the low hundreds of millions shows that enterprise AI is a real business, not only a research topic. Whether it becomes a great public company depends on growth quality, margins and competitive position. For customers, the lesson is similar: measure the value, protect your options and favor vendors that are honest about what their models can and cannot do.
