Cohere develops language-processing technology.DADO RUVIC/Reuters
Artificial-intelligence company Cohere Inc. is in advanced talks to raise up to US$250-million from investors in a financing that could value the Toronto-based startup at just over US$2-billion.
Cohere, which develops language-processing technology, has been in discussions with chip maker Nvidia Corp. and investment firms about securing funds, according to two sources familiar with the matter. The round is being led by Inovia Capital, with partner Steven Woods, a former senior director of engineering with Google, steering the investment for the Montreal venture capital firm.
The Globe and Mail is not identifying the sources because they are not authorized to speak publicly about the matter.
The full financing has not yet closed, according to the sources. Other investors involved in the discussions include earlier backers Index Ventures and Radical Ventures, as well as Thomvest Ventures, which makes investments on behalf of Peter J. Thomson. (Mr. Thomson is also co-chairman of Woodbridge Co. Ltd., which owns The Globe and Mail.)
Cohere and Inovia did not respond to a request for comment.
Cohere’s financing, which has been in the works for months, would be done at a higher valuation than its last financing in early 2022, when it secured US$125-million led by Tiger Global Management LLC. Other Canadian tech companies have struggled to raise funds and have had to cut costs to preserve cash amid a tumultuous market climate.
But it is also a sign of the intense investor excitement about AI. About US$1.7-billion poured into generative AI startups in the first quarter, according to data from PitchBook, while another US$10.7-billion in deals was announced but not yet completed.
Some proponents of AI see the advent of ChatGPT, a chatbot from OpenAI, and other generative tools as a massive opportunity not unlike the dawn of the internet. Microsoft Corp. is investing billions into OpenAI, while Google parent company Alphabet Inc. is integrating AI tools into its products and services. Some investors and executives contend generative text applications can upend internet search, customer service, software development and other fields, boosting productivity and saving costs. That could completely upend Silicon Valley’s power structure.
Some critics contend the trend is overhyped, while raising concerns that AI is developing too quickly for adequate safety guidelines and regulations to be put in place. Many experts are warning about potential dangers ranging from a wave of AI-generated misinformation to existential threats to humanity.
On Monday, prominent AI researcher Geoffrey Hinton said he resigned from his post at Google so that he could talk more freely about the potential dangers of the technology. Mr. Hinton is also an early investor in Cohere.
Cohere was founded in 2019 by Aidan Gomez, Nicholas Frosst and Ivan Zhang, all of whom studied at the University of Toronto. Mr. Gomez is one of several co-authors on a 2017 research paper proposing a new method for computers to interpret language. That method now forms the basis of the current wave of generative text tools.
Like OpenAI, Cohere develops large language models, or LLMs, which can be used to generate, summarize and interpret text. Customers pay to access Cohere’s LLMs and build applications, such as a custom search engine or a chatbot.
LLMs require substantial – and costly – computing power and technical infrastructure. Tech giants such as Microsoft, Alphabet and Meta Platforms have deep pockets to fund their AI operations, whereas startups have to tap investors for capital.
Cohere hasn’t attracted the same kinds of headlines as OpenAI but has been recognized as one of the leading crop of companies trying to develop and commercialize generative AI. One independent evaluation of language models, conducted by researchers affiliated with the Stanford Institute for Human-Centered Artificial Intelligence, found that some of Cohere’s models performed better than the competition in key categories, such as accuracy and fairness. Cohere has raised at least US$165-million in two financings to date.
It has also hired top names as senior executives, including president Martin Kon, the former chief financial officer of YouTube, and two former researchers with Google’s DeepMind AI operation, chief scientist Phil Blunsom and Ed Grefenstette, head of machine learning.
Cohere has said its strategy is to make its LLMs accessible to companies of all sizes to build AI-powered applications, arguing the technology has long been the exclusive domain of tech giants. OpenAI is pursuing a similar strategy, and its models already power countless apps, such as legal research and document drafting tools, writing assistants, and customer-service chatbots. Cohere’s customers include Spotify and legal AI company Casetext.
OpenAI isn’t the only competitor. San Francisco startup Anthropic reportedly secured US$300-million from Google this year. Stability AI, which has released a popular AI image generator, unveiled a language model last month. Elon Musk has mused about starting a rival to OpenAI, which he co-founded.
“There’s absolutely room for multiple players here,” John Somorjai, chief corporate development and investments officer at Salesforce, said in an interview Monday. “It’s so early in the development of these LLMs that it’s very unclear who the winners are, or who they will be.”
Salesforce launched a US$250-million generative AI fund in March and has announced investments in both Cohere and Anthropic.
One of the sources said the company is pursuing a “Switzerland” strategy by emphasizing its neutrality. The company may appeal to potential customers because it is not overly affiliated or controlled by data giants, assuaging fears that internal data could be accessed to improve the giants’ training models.
As the company is still generating modest revenue, according to the sources, a challenge for Cohere will be to justify its lofty valuation by solving real-world business issues for paying customers, and then scaling up revenue quickly and substantially.