Cerebras Challenges Nvidia's AI Chip Dominance as Competition Heats Up

The CEO of Cerebras, which could go public soon, takes on Nvidia in a David and Goliath battle for AI chip supremacy

This story is available exclusively to Business Insider subscribers. Become an Insider and start reading now. Have an account? Log in. Feldman founded his company in 2015, 16 years after Nvidia first debuted the graphics processing unit, which later turned out to be an unlock for AI. Feldman and his cofounders saw an opportunity for chips designed for AI -- not repurposed. The chip they designed is often described as the size of a dinner plate. The size essentially makes it both fast and tricky to execute. Some have even called it technically "radical." Cerebras shares the rest of the AI chip market with a handful of other startups like Groq, Positron AI, Rivos, and Etched, and Big Tech staples like Intel, Amazon, Google, and AMD. Still, Nvidia gets most of the hype -- and the revenue. Cerebras has raised more than $700 million in funding, according to PitchBook. The public markets may soon have the opportunity to endorse or cast aside Feldman's vision. Though the company will not confirm it, The Information reported last month that Cerebras has confidentially filed for an IPO. The David and Goliath comparison is inevitable. "I'm a professional David," Feldman said. He's adamant that the future of AI will turn the industry toward his chips. Cerebras already has a customer base that includes GSK, AstraZeneca, and TotalEnergies. Cerebras has attracted customers from industries with heavy-duty computing needs, such as pharmaceutical drug discovery, oil and gas, and military and defense applications. Ample memory and the large wafer reduce the requirement for networking and speed up calculations. The draw for machine learning power users is also financial, Feldman said. These industries don't just have hefty computing workloads; they also have business models that can benefit from small improvements in computing speed or results. A slightly better drug discovery model could mean millions in profits. More consumer-facing models don't necessarily have that level of nuance. Still, Feldman expects that even consumer-directed AI models will grow larger, and larger models need faster chips. "What OpenAI showed was that this class of model would get better as it got bigger, approximately without bound," Feldman said."As we go bigger, we will need more compute," he added. Right now, most of that computing happens on Nvidia chips. And as a competitor, Nvidia is "extremely aggressive," Feldman said. For example, he says Nvidia uses its market strength with GPUs to advance its networking business. "Is that anti-competitive? Tough to say. But it's what bullies do," Feldman said. Tactics aside, Cerebras faces a perhaps more powerful force. Some call it convention or prevailing wisdom. Feldman called it "mindshare." Tech procurement has an old phrase: "Nobody gets fired for buying IBM." Today's equivalent is, "Nobody gets fired for buying Nvidia." Feldman's response: "Nobody gets promoted for using it either." "You can't be faster than your competitors because they're using it too," he said. Competing with that mindshare is an uphill battle as Nvidia faces more competition and goes on defense. Early Cerebras investor Lior Susan, founding partner of Eclipse Ventures, said Cerebras's capabilities are not necessarily jostling for the existing market but creating new markets. He also told BI that even getting 10% of Nvidia's more than $3 trillion market capitalization would make him "a happy man." Still, Susan said if he could snap his fingers and change one thing to make the path easier for Cerebras, he would wipe out Nvidia's massive head start -- the decade in which the company built that mindshare and the myriad AI tools to make the company synonymous with that work. An Nvidia spokesperson declined to comment for this story.

AI startup founders discuss what it's like to be a part of Nvidia's sprawling startup empire

Investing through its own corporate development team and its in-house venture fund called NVentures, Nvidia grew its startup portfolio from $300 million in the beginning of 2023 to more than $1.5 billion today -- making more than 24 investments last year alone. The deals have included foundation model developers like Cohere and Mistral -- Europe's open-source competitor to OpenAI, AI cloud providers like Coreweave, picks and shovels giant Databricks, generative AI darling Runway, data provider Scale AI, model sharing hub Hugging Face, and search engine challenger Perplexity. The company also runs Nvidia Inception, an incubator and venture network that counts more than 20,000 early-stage companies among its membership. The Inception program doesn't come with investment; it's more of a support network and a pathway for startups to access prized computing power. Though Chris Brown, a partner manager for the program, said at an event last month in New York that the program isn't trying to sell the startups anything. "We're helping build an ecosystem, and we understand our role in that ecosystem is very foundational," Brown said. Investors often need help seeing their way through a melee of AI tools and startups that can seem to overlap and stack on top of each other, he added, so the Inception team plays that role too. Though much of Nvidia's startup empire's day-to-day functioning is overseen by Head of Corporate Development Vishal Bhagwati and NVentures Head Sid Siddeek, CEO Jensen Huang personally approves every investment, according to multiple founders who have been backed by Nvidia as well as people familiar with the matter. "He is the company," said the CEO and founder of a startup that has had Nvidia on its cap table since 2020, who spoke on condition of anonymity. "There's no difference between Jensen and Nvidia in my mind; the big decisions go through him." "I think it's crazy," said another founder whose company received investment from NVentures, who spoke on condition of anonymity. "I mean sure if Satya [Nadella] is making a $10 billion investment in OpenAI, of course he'll review it," the founder said. "But Nvidia is supporting companies that are seed level, and Jensen is reviewing all of those deals." This founder met with Huang personally before he signed off on the Series A investment. He says Huang was intimately familiar with the company's technical reports and up to date on all the latest research papers. Huang also sometimes personally steers other investors. When Umesh Padval, a semiconductor-industry veteran and managing director of VC firm Thomvest, was considering an investment in Cohere's Series C funding round, Padval spoke to Huang about it and was, in part, persuaded by Nvidia's presence in the round. "I wanted Jensen and Nvidia to be in Cohere because then there is no issue about supply," Padval told BI regarding the 2023 deal, which came through at the height of the GPU shortage, the frenzy of which has since somewhat dissipated today. It's unusual for a public company CEO like Huang to take such a hands-on approach to venture investing, but the startup founder who spoke anonymously sees it as a feature, not a bug, allowing the company to remain nimble despite being worth two Facebooks, an Uber, and a Salesforce. "He has built the biggest startup in the world," said the founder who was backed by Nvidia in 2020, "they're a 3 trillion dollar startup, in the sense that they don't have that hierarchy, they don't have all of those rules. Jensen says they don't even have an org chart." Nvidia, once seen as a maker of niche hardware for hardcore gamers, has, over the past two years, transformed into one of the most important (and most valuable) companies in the world. Its high-powered graphics cards are the brains upon which advanced AI models are trained. Any company trying to do anything major in AI has spent the past two years buying them by the truckload. And many, if not most, of the startups the company has funded are also Nvidia customers. But Nvidia's first-mover advantage in chips can't last forever. Google, Amazon, and AMD (run by Huang's cousin Lisa Su) are all racing to develop their own chips to compete with Nvidia's GPUs and a handful of smaller outfits are coming too. Seeding a startup field with an affinity for Nvidia tech is one strategy to stay on top. But, like an oil-rich nation seeing the writing on the wall, Huang is racing to diversify before it's too late. "There is no way Nvidia stays at this value," said the CEO of a company that is both a customer of Nvidia and a recipient of its investment, "it's more valuable than Amazon right now, that makes no sense. It is a great company, it is ahead of the competition, but how long can it hang on to that? Is it six months or is it just two more hours, who knows?" The CEO says he's already noticed a change in his negotiations with GPU providers, who are being more generous with their pricing as the chip market becomes more saturated. He's also been the recipient of Huang's famous hands-on approach. He says he's received personal calls from Huang in which the Nvidia CEO weighs in on company decisions or expresses disapproval of his public statements. He says that Huang sometimes seems to know more about what's going on inside his company than he does. In every conversation, he says, there's an elephant in the room. "He's very obsessed with what the competition is doing," he said. The founder who was first backed by Nvidia in 2020 sees Huang's granular focus on the companies he invests in as part and parcel of his broader vision for Nvidia. "They want to see into the future," he said, "through their investments, through their startups that they're collaborating with, they want to understand where the world is going so that they can get ahead of it."