The CEO of OpenAI rival Cohere shakes off the haters: ‘We’re still sort of the underdog’

You may not be familiar with Cohere, but the Toronto-based startup is among OpenAI’s top competitors in the race to build the most capable large language models (LLMs). Cohere, along with Big Tech behemoths like Google and Meta, as well as other startups like Anthropic and Mistral, are all chasing OpenAI in the frenzied competition for AI dominance that began with the release of ChatGPT in November 2022.

It’s a pursuit that requires deep pockets: The cost of training large LLMs is staggering and necessitates a constant stream of new funding. But in many ways, Cohere is running its own race. For one thing, it is targeting big business customers and has not developed a consumer chatbot. Cohere’s rivals include Big Tech companies with their own cloud computing arms; startups closely partnered with them; or those giving open-source models away for others to build upon. Cohere, meanwhile, has sought to maintain its financial independence from any single cloud ecosystem (though it does have a partnership with Oracle, the fourth largest cloud provider).

However, Cohere, which was founded in 2019 by three alumni of Google Brain, is now under pressure to prove its models, and its go-to-market strategy, can deliver world-class performance to businesses—while at the same time raising enough money to pay for the computing power it needs and generating enough revenue to satisfy investors. Industry-watchers are tracking the company closely to see if it can make the grade.

At a time when the conventional wisdom is increasingly that the generative AI boom is only further consolidating the power and market dominance of America’s largest technology companies, the go-it-alone effort by the Canadian upstart represents a stark counterpoint. If Cohere can succeed, it will show that the power of Big Tech can be challenged by nimble, well-funded startups—including those whose fates are not determined by checks written by Big Tech partners. If it fails, it will be another proof point that the money and computing resources needed to succeed in the AI race mean that only the largest tech giants can compete. Which narrative proves correct will have important implications for how AI and Big Tech are regulated going forward.

Competing in a crowded space

So far, the startup’s report card is mixed. On the plus side, it has succeeded in shipping several well-received new models over the past two months. Its Command family of models, focused on techniques that reduce hallucinations and offering more affordable pricing, can also operate in more languages than the competition. The company lately boasts new revenue channels including, since the start of this month, the opportunity for customers to access Cohere’s new Command R+ model on Microsoft’s Azure cloud, where the majority of generative AI workloads reside. Cohere emphasizes it now offers the only proprietary models that are available on every cloud—AWS, Google Cloud, Oracle Cloud Infrastructure, and Microsoft Azure.

But Cohere has had a tougher time placating critics who question its stratospheric valuation, which is set to more than double to $5 billion after a $500 million round that is rumored to be in the works. An article last month in The Information called Cohere’s 2023 revenue “tepid.” However, a source briefed on Cohere said the company’s annualized revenue at the end of Q1 had risen to about $35 million, up from what was reported by The Information to be just $13 million in annualized revenue at the end of last year. Meanwhile, The Information has also reported Anthropic’s projection that it will generate more than $850 million in annualized revenue by the end of 2024, while it said some leaders believe OpenAI can reach an annualized revenue of $5 billion by that time from consumers and businesses.

There are other challenges: The marketplace is jam-packed with others releasing new models and competing for mindshare and market share. Not a week seems to go by without a major new release of a model family: Just last week, Meta’s release of Llama 3 set the tech world buzzing. In addition, the enterprise market has become increasingly competitive. As Fortune has reported, OpenAI remains the big winner with corporate customers, at least for now, and every LLM provider is working to elbow its way into the enterprise space with offerings that enable organizations to fine-tune AI models with their own proprietary data for specific use cases.

The fact is, billions may still be pouring into generative AI startups, but there have already been casualties in the LLM space—including Stability AI, which has lost its CEO, shed jobs, and reportedly suffered cash shortfalls; and Inflection AI, which, after raising $1 billion, lost its CEO, Mustafa Suleyman, and much of its staff to Microsoft.

‘We have to prove ourselves’

On a Zoom call, Cohere CEO Aidan Gomez, who logged on from London, where he now spends about half his time, showed no hint of the pressure facing him and his team. He seemed relaxed and said he is “stoked” by the latest positive feedback from developers and outreach from enterprises. “We’re still sort of the underdog,” he explained. “We have to prove ourselves … I think in the past couple of months we’ve really done that.”

When asked about Cohere’s new round of funding, which would mean the company will have raised more money than any other LLM startup except for OpenAI and Anthropic, Gomez simply said that every LLM company must raise money to pay for compute. “Cohere has to raise money to build the supercomputers that build our models,” he said, adding that the company has seen “very considerable growth” in the first quarter of 2024. “In the first 10 weeks of the year, we grew over 60%, and that’s continuing to accelerate.” But he declined to disclose specific financial figures.

As a Canadian company, Cohere is also about to enjoy subsidized access to some additional computing power: Two weeks ago, the Canadian government announced it will invest $2 billion to “build and provide access to computing capabilities and technological infrastructure for Canada’s world-leading AI researchers, startups, and scale-ups.” Experts say Cohere is set to benefit considerably. In addition, Canadian pension investment manager PSP Investments is reportedly set to co-lead Cohere’s next funding round.

As for questions about Cohere’s revenue, Gomez calls them a “mis-framing” at a moment when companies are just starting to put AI use cases into production. Now, he says, Cohere is finally ready to put the pedal to the metal: Last year companies were experimenting with generative AI, but those workloads are small, he explained. “As we’re seeing those POCs [proofs of concept] succeed and start to flip into real heavy-duty production workloads, we’re really well positioned to capture that,” he noted, “because of our focus on building models that are efficient and scalable, which is something that our competitors have neglected relative to us.”

Finally, Gomez insists that direct comparisons with OpenAI and Anthropic are misleading. Rather than offering its models only through an API (application programming interface) and charging for each token an LLM serves, Cohere also offers private deployments like an SaaS (software as a service) product: The customer has their compute on one of the big clouds or on premise, and Cohere charges a fee on top of that.

“That leads to much higher margins for us, because we’re not paying for that compute,” Gomez explained, though Cohere declined to discuss specific financials. “I can say it’s extremely unique in the market; if you look at the product offerings from our competitors, they tend to look like basically dollar per token, which has the compute baked in, which requires a lot of capital.”

Gomez also argued that Cohere’s decision to forgo building a consumer chatbot, with freemium users using lots of compute, means the company does not have the same inference costs (the expense of actually running fully trained AI models in deployment) as OpenAI and Anthropic do. “We’re starting to hit the inflection point now where spending on inference compute is becoming higher than spending on training, which is indicative of market maturity,” he said.

Challenges and opportunities

Gartner analyst Arun Chandrasekaran says the challenge for a company like Cohere is in its go-to-market product strategy. Cohere may well value its independence and presence on multiple clouds, he explained, but model providers that are tightly partnered with hyperscale cloud providers may have lower customer acquisition costs because their products and services are prioritized on their partner’s cloud.

“The cloud providers are increasingly starting to make available what they call a model ‘garden,’ where they make available both first-party models, as well as a number of third-party models,” he told Fortune. “But it’s up to you as a model provider to create the go-to-market motions to prioritize your model and your model integration with the cloud provider. The cloud companies are going to, in some sense, prioritize their first-party models.”

Daniel Newman, CEO and chief analyst of the Futurum Group, agreed, saying that any of the Big Tech companies “will lead with where they have the biggest ROI opportunity.” Microsoft will lead with its own models, he explained, or those of its partners, OpenAI and Mistral. Google will lead with Gemini. And Amazon will lead with its own Titan models or those from Anthropic, in which it has made a sizable investment.

That said, both Newman and Chandrasekaran said there are paths for Cohere to succeed long term.

Cohere’s SaaS-based approach “has legs” when it comes to monetization, Newman said. He noted that Cohere’s sole focus on large corporate customers was also likely to give it an edge over startups trying to appeal to both businesses and consumers, though he added the company could benefit from more specialization—in highly regulated industries that have their AI workloads on-premise rather than in the cloud, for example.

“I think if Cohere really carved out its niche they could grow really handsomely into a market that’s in the earliest innings,” he explained.

Chandrasekaran agreed, saying the demand for more domain-specific and task-specific models offers potential opportunities for Cohere. “The challenge is you have to find a niche that is large enough to demonstrate a really good degree of accuracy,” he said. Another option is delivering value through ease of use—providing something out-of-the-box that makes it simple for enterprise companies to quickly deploy AI models. “That could be another way to tell enterprises that time is money, and you can charge a little bit of a premium,” he said.

In addition, he pointed out that Cohere does have an opportunity to even more closely partner with Oracle—both in Oracle’s cloud arm, OCI, as well as being the underlying model for powering Oracle apps. “Oracle is obviously trying to compete with Microsoft and Salesforce and others in terms of embedding AI models into SaaS applications,” he said.

But Newman emphasizes that there is still “a pretty big gap right now” between OpenAI (as well as Anthropic and Meta) and Cohere in terms of “where they’re seen as the most viable for long-term success and growth.” He compared the LLM landscape to that of the main cloud providers—where AWS remains the leader with Microsoft Azure and Google Cloud playing catch-up, but Oracle’s OCI is in a distant fourth place. “I think Cohere can do well,” he explained, “I just think they are unfortunately kind of the fourth cloud right now.”

Time to deliver 

When asked to respond to Cohere’s detractors, Gomez refused to get outwardly negative or defensive. “I love feedback,” he says. “It‘s what makes our products better. So we want to build great, useful technology, and criticism is a core piece of that.”

Often Gomez seems just happy to be in the AI race at all. It’s a long way from his childhood in rural Ontario, where he said he grew up on a 100-acre forest property with maple trees that the family would tap to make maple syrup.

On the other hand, he gained industry fame as the youngest of eight Google Brain authors of the seminal paper that introduced transformers, a breakthrough neural network architecture that became the underpinnings of OpenAI’s GPT-4 and ChatGPT. After the transformers paper was released in 2017, deep learning pioneer Geoffrey Hinton invited Gomez to join his Toronto lab, where Gomez met Nick Frosst; the pair would go on to found Cohere with fellow Google alum Ivan Zhang.

Gomez said he has been gratified by outreach and excitement from the developer community and existing customers in response to Cohere’s new models—as well as companies that are taking a fresh look at Cohere after trying other options, including open-source models. But he knows the most important thing is to deliver in order to meet the moment.

“It’s now time for production; workloads are becoming real now—it’s no longer testing,” he said. “That’s what I’ve been waiting for, for over half a decade. So I can’t wait.”

This story was originally featured on Fortune.com