Toronto-headquartered Cohere is reportedly raising at a 10-figure valuation on an eight-figure revenue run rate, as generative AI startups try to turn the buzz about the technology into actual sales. Here’s what you need to know:

The business: Cohere’s large language models (LLMs) can respond to questions, summarize documents and produce more relevant search results. Unlike other AI toolmakers, such as OpenAI and Google, Cohere isn’t building consumer applications. Its business is selling to businesses, who can build the technology into their own products and services. 

The startup has also signed deals with Oracle, an investor, as well as consulting giants McKinsey and Accenture, that will see those firms offer Cohere’s AI tools to their clients. 

As of this month, the startup had an annualized revenue run rate—its current sales scaled to a year-long period—of US$22 million, Reuters reported overnight. That was up from US$13 million in December, a number The Information first reported late Wednesday.

The Information also reported that Cohere has told investors it has over US$300 million worth of potential contracts in the works that could close this year. 

The fundraising: Cohere is well down the road to raising US$500 million at a US$5-billion valuation, Reuters reported. The Information reported that an unidentified pension fund is set to lead the round.

That would put the firm at a valuation multiple of 227 its annualized revenue, higher than most of the other major model builders. Cohere was valued at US$3-billion in a private sale of its stock last August. The startup’s last round was a US$270 million Series C in June 2023, led by Inovia Capital.

Cohere did not respond to The Logic’s request for comment.

Under the hood: Startups looking to sell the technology to power the generative AI boom are balancing the huge costs of training models with the gradual ramp up of the corporate world’s adoption. Amid all the mega-rounds, some in the tech sector are questioning whether selling access to LLMs can bring in the kind of revenues that justify investors’ valuations and expectations, or what new products they might develop to attract clients. 

Cohere sees potential in its latest set of models. Earlier this month, the firm released Command-R, a model specialized for an increasingly popular way that firms use LLMs, known as retrieval augmented generation (RAG).

Most models like Cohere’s are trained on the internet’s worth of words. But they don’t always have the most current information. And companies don’t want their proprietary data mixed into the main system, where competitors could access or benefit from it. RAG lets firms marry a commercial LLM to their internal information, generating more relevant text and giving their employees more useful search results.