- Aidan Gomez is impressed by DeepSeek but doesn’t think its AI is enterprise-ready.
- In an interview with Business Insider, the Cohere CEO said companies are looking for custom models.
- Gomez, an ex-Google researcher, said DeepSeek has validated his view that AI can be cheaper.
Aidan Gomez, the CEO of Cohere, felt vindicated in his belief that powerful AI didn’t need to be so expensive when DeepSeek released a model that would go on to blow a $1 trillion hole in the US stock market.
“I think it validated Cohere’s strategy that we’ve been pursuing for a while now,” the 28-year-old computer scientist said in an interview with Business Insider. “Spending billions of dollars a year isn’t necessary to produce top-tier tech that’s competitive.”
But while Gomez, an ex Google researcher, considers DeepSeek’s R1 “a really impressive release,” he’s not convinced it should be a serious option for businesses.
He said organizations are looking for customized AI models rather than something off the shelf — and are cautious about giving AI tools access to sensitive data.
“We don’t see the enterprises that we sell to relying on R1 to power their systems,” Gomez said. “We don’t see it as a competitor on our side.”
Why DeepSeek isn’t enterprise-ready
As the leader of a $5.5 billion company building AI for enterprises, Gomez has a clear business reason to make this case. But, as one of the eight Google Brain researchers who co-authored the 2017 seminal “Attention is All You Need” paper that sparked the generative AI boom, his position carries weight.
Nor is he alone, with some US firms trying to adopt DeepSeek running into a host of troubles, BI previously reported.
For Gomez, DeepSeek isn’t a quick win for businesses — regardless of how impressive its tech might be.
“What we’re seeing from enterprises is that they don’t just want to buy a model,” he said. “You’re going to have to build something with that model, you’re going to have to deploy a lot of technical resources to see value, and it will take time.”
To unlock “a new tier of value,” he thinks enterprises must carefully consider how they customize core AI technology with their proprietary data.
In November, amid an industry-wide debate over whether AI performance gains had hit a wall, leaders cited private and synthetic data as key resources that organizations must tap into to maintain a competitive edge.
It’s a point echoed by Gomez. And, as concerns rumble over DeepSeek being “back-ended by servers in China” — US lawmakers are seeking to ban the startup’s software from government devices — the Cohere CEO said enterprises must put privacy first if models are to touch “more and more sensitive data.”
“That’s something that will unlock usage in enterprises because right now, they’re hesitant to build systems that touch sensitive data,” he said. “Our competitors treat it in a way that’s less secure.”
DeepSeek has not returned Business Insider’s requests for comment about its data privacy policies.
All about AI agents
While Gomez thinks Deepseek’s R1 is impressive, he believes the real value will come from transforming a base model into a tool that’s proving to be another hot area for the industry this year: agentic AI.
Software programs that can perform tasks autonomously have been high on the agenda of business leaders this year. Agentic AI was a hot topic at Davos, while Nvidia CEO Jensen Huang said at the Consumer Electronics Show that 2025 will be the year it takes off.
Toronto-based Cohere, one of a handful of companies competing with AI rivals like OpenAI, Google, and Anthropic, is focusing on bringing AI agents to enterprises.
Last month, Cohere introduced its early-access program for North, its own agentic AI that’s designed to meet specific workloads.
Gomez sees it as another way for Cohere to gain an edge over companies that just want to deploy a base model like DeepSeek’s R1.
Gomez declined to say how much it can cost Cohere to adapt a platform like North to the needs of specific enterprises. He said after an “upfront investment,” agents can “operate quite autonomously” once they’re plugged in and then allow enterprises to “start reaping the value.”
Getty Images
DeepSeek the disruptor
Despite his concerns about DeepSeek, Gomez views the startup spun out of a Chinese hedge fund as a positive disruptive force for the AI industry.
“The fact that they published their training efficiency numbers let people see that it doesn’t need to be so capital-intensive to publish fantastic models,” he said.
AI leaders continue to scrutinize DeepSeek’s claims that it produced AI on par with the performance of Silicon Valley’s best models at a fraction of the cost. In the meantime, investors are questioning whether large AI infrastructure spending is still justified.
Addressing the implications of Sam Altman’s $500 billion Stargate project, Gomez said that “spending more and more” on infrastructure for training AI models, rather than “inference,” is a mistake. Inference refers to an AI model making predictions or decisions about new data, while training is the process of building a model’s capabilities.
“I think DeepSeek’s a big proof point of that,” he said.
The other implication of DeepSeek’s emergence out of seemingly nowhere is the validation of an open-source approach.
While there is debate over whether DeepSeek’s AI is truly open-source — it has secured an MIT license and made its model weights open, but hasn’t disclosed the data it used to train it — Gomez sees “great technology” coming from both open-source and closed-source players.
Still, enterprises will need more than an impressive Chinese model to build powerful AI into their operations. As Gomez put it, “It’s not just enough to download a model.”
