Cohere co-founder Nick Frosst at the company's Toronto office on July 22.Christopher Katsarov /The Globe and Mail
Canadian artificial-intelligence startup Cohere Inc. has raised US$500-million in a new financing round to compete with OpenAI, Anthropic, Google GOOGL-Q and other well-funded rivals seeking to develop and profit from generative AI.
The latest funding values Cohere at US$5.5-billion.
The round was led by the Public Sector Pension Investment Board, which manages pensions for the Canadian federal public service, and includes new investors Cisco Systems Inc. CSCO-Q, AMD Ventures, Fujitsu Ltd. FJTSY, U.S. investment manager Magnetar Capital and Export Development Canada.
Cohere, founded in 2019 in Toronto, is one of the few companies in the world that builds the large language models that underlie chatbots and other generative AI applications, which can produce and analyze text and other media. The company is distinguishing itself from competitors by focusing on building models that can be tailored for business use, such as answering questions, summarizing information and automating processes, rather than making general-purpose chatbots, such as OpenAI’s ChatGPT, or text-to-image generators.
“It’s that philosophical difference of getting this stuff to be useful,” Cohere co-founder Nick Frosst said in an interview, “and not chasing [artificial general intelligence] but focusing on solving enterprise problems.”
Cohere’s valuation has more than doubled since it raised US$270-million in June of last year in a round led by Inovia Capital in Montreal. That’s a sign not only of the investor enthusiasm for Cohere, but of the bullish – some skeptics might argue bubbly – sentiment toward generative AI more broadly. Cohere has attracted a slew of investors over the past few years, including Nvidia Corp. NVDA-Q, Oracle Corp., Salesforce Ventures CRM-N and Toronto’s Radical Ventures, which wrote the company its very first cheque.
Cohere was founded by Mr. Frosst, Aidan Gomez and Ivan Zhang. In 2017, as a Google intern, Mr. Gomez co-authored a research paper that outlined a new method for AI models to produce text, which is now the basis for how today’s LLMs work.
Building LLMs is an expensive endeavour that requires a lot of dedicated computing horsepower from graphic processing units, or GPUs, which are in high demand around the world. Cohere will use the money it has raised to pay for computing resources and to bring on more employees, among other initiatives. The company employed 250 people at the start of the year, including in San Francisco, New York and London, and it aims to double that number by the end of 2024.
Serious questions remain about the costs of building and running generative AI models, and about their reliability. Generative AI applications can make factual errors and invent information, a quirk known as hallucinating.
But Mr. Frosst said that enterprise users have been impressed with Cohere’s latest models, which were released earlier this year under the name Command R. “We’ve seen a lot more traction,” he said. “That was a real turning point for us.”
Guillermo Freire, senior vice-president, mid-market group at EDC, acknowledged the hype and uncertainty around generative AI, but said that Cohere has all of the ingredients to be a lasting international player, in part because of the multilingual capabilities of its models. “These technologies are still emerging,” he said, “but we’re confident that Cohere is among the few that will be successful.”
Some Cohere users have moved beyond pilot projects and are putting generative AI applications in front of customers and employees, including to answer legal questions in multiple languages and automate some human-resource functions. Cohere has also partnered with consulting firms McKinsey and Accenture to bring the technology to clients.
Toronto-Dominion Bank is testing Cohere’s models, though the effort is still in the early stages. “Over the next several months, we’ll begin experimenting with Cohere’s models to determine if they can help accelerate the testing, rollout and reliability of our generative AI use cases,” said Maksims Volkovs, the bank’s chief AI scientist.
Cohere’s focus on practical uses for generative AI has led it to invest in a technique called retrieval augmented generation, or RAG, which it says reduces hallucinations. With RAG, an AI application can access information outside its training data, allowing a customer-service chatbot to pull from a corporation’s policy documents to provide accurate information, for example.
The drive for AI companies to generate revenue has taken on more importance as valuations have soared and skepticism lingers about the technology’s usefulness in real-world business settings. Mr. Frosst declined to comment on Cohere’s finances, but Reuters and The Globe and Mail have previously reported that the company expects annualized revenue to hit more than US$300-million by the end of the year.
In June, a Goldman Sachs report questioned whether the returns of generative AI justify the cost. “My main concern is that the substantial cost to develop and run AI technology means that AI applications must solve extremely complex and important problems,” Jim Covello, head of global equity research, said in the report.
So far, generative AI has made some processes more efficient, such as computer coding, but it’s still more expensive than other methods, according to Mr. Covello. A lot of people “substantially overestimate” what generative AI can do today, he added. “The technology is nowhere near where it needs to be in order to be useful,” he said.
Cohere is well aware of the expenses of building and deploying generative AI, according to Mr. Frosst, which is why the company takes cost efficiency into account. “We’re not doing science experiments. We’re building a product,” he said. “We make models that we know if you go into production, it’s not going to break the bank. It’s going to be a reasonable decision for your business.”