Cerebra Systems' IPO on Thursday was a huge success, netting the company, its founders and key investors billions of dollars in profits.
Among the big winners is major shareholder Benchmark, which owns 9.5% of the company's stock. Eric Vishlier, one of the company's general partners, has been on Cerebras' board since the AI chip maker was founded in 2016 and co-led the $25 million Series A round.
But these billions of dollars only happened to Benchmark. That's because Vishria met with the startup almost against his will, he told TechCrunch.
“We had five founders and one deck, and it was our first hardware investment in 10 years,” Vishria told TechCrunch about that first meeting. “I was a venture capitalist for about 18 months.” (Before becoming a VC, in 2013, Vishlia sold RockMelt, a social browser startup he co-founded, to Yahoo for a reported $60 million to $70 million.)
Benchmark is notoriously selective and rarely backs hardware companies, so much so that Vishria blames himself for devoting time to Cerebras.
“Why did you come to this meeting?” he kept muttering. At one point, he eavesdropped and sent a message to the assistant managing his calendar, asking, “Why did you put me in this meeting?” Vishlia looks back.
But his sullen demeanor disappeared by the third slide, as co-founder and CEO Andrew Feldman explained Cerebras' grand plans.
“The first deck is the title slide. The second deck is the team. And I thought, 'Wow, that team is really cool.'” And the third slide is something like, “GPUs aren't really useful for deep learning.” It just happens to be 100x better than the CPU. 'And as soon as he said that, a light bulb went off,' Vishlia recalled. “I thought, 'Oh, of course, why wouldn't a graphics processor be perfect for AI?'”
Still, this was years before Google's famous Transformer paper, the 2017 study that laid the foundations for modern AI, that ultimately led to ChatGPT. Cerebras was pitching a new kind of giant chip designed for AI training, but the processor industry wasn't ready for manufacturing.
Vishria was so intrigued that he spoke with several benchmarking partners and was quickly told that they too didn't know much about the hardware. They said if he wanted this deal, he would have to bring in one of the guys who was the founder of Benchmark in the 1990s and had an understanding.
Still, Mr. Bishria arranged a meeting for Mr. Feldman to pitch chip packaging, cooling and other topics to founding partner Bruce Dunleavy.
“Most of those meetings were like a dog watching TV to me,” Vishlia joked. Because I could barely understand it. After the pitch, Dunleavy warned that what Cerebras was attempting would be difficult. Others have tried and failed. But he thought this team had a chance. But he feared there would be no market for the chip.
Although Vishlia didn't fully understand the technology, he believed there would be a market for Cerebras “if we could make AI faster,” and said the team had the makings for success. They had previously sold their startup SeaMicro to AMD.
“The benefit of having a previous successful exit is that it removes some of the uncertainty in the minds of venture capitalists,” Cerebras CEO Andrew Feldman told TechCrunch. “We didn't just fall off the back of a Cub truck. We were an experienced team.”
hardware is difficult
After that, Cerebras struggled for eight and a half years with struggle after struggle to build the product.
Feldman and Cerebras co-founder and chief technology officer Sean Lee needed to invent a new cooling method to prevent a chip of that size from burning up when powered. They needed to invent a machine that could drill 40 screws at the same time without breaking the wafer. and so on.
Benchmark investors repeatedly wondered, “What are we doing?”
Additionally, the hardware is expensive. The company's chips were still in development when it raised $500 million from a long list of investors. The 2022 VC bear market meant they needed to raise again.
“The company still doesn't have a lot of traction, so yeah, that was the real challenge,” Vishria recalls.
But about 18 months ago, everything changed. Cerebras' chips were designed for training and successfully manufactured by TSMC, the world's largest contract chip manufacturer, but they turned out to be even better at inference, running AI models and generating responses, rather than first teaching them. Just as that realization came, so did the world of AI's insatiable thirst for that kind of computing. We had big customers and revenue.
Cerebras attempted to go public in 2024 instead of another private round, but became embroiled in U.S. government scrutiny over national security concerns after a huge investment by its only major customer, Abu Dhabi-based cloud provider G42. General investors were also concerned about dependence on G42 and huge losses.
The delay was actually a blessing. Currently, OpenAI and AWS are also large customers. Cerebras doubled its revenue and posted a profit last year.
Vishlia is providing all the support to the Cerebras team, requiring “tenacity, ingenuity and adaptability,” he says.
But it's also very difficult for investors who find winners far outside the company's normal comfort zone. Benchmark held 17,602,983 shares worth $3.3 billion at its IPO opening price of $185, and would have held more than $5.3 billion if the price remained above $300 on the first day of trading. The shares cannot be sold until the six-month lock-up ends. This is a standard restriction that prevents insiders from selling immediately after a company goes public.
The company bought about 80% of those shares in an earlier round for about $18 million, various disclosures show, Vishria confirmed to TechCrunch. Cerebras bought the rest in a later, more expensive round, costing about $250 million, the company disclosed in its S-1.
After all, the venerable venture capital firm likely spent $270 million on a stock that could be worth billions or more, depending on stock price trends.
Employees of venture capital firms can receive bonuses if their investments yield significant profits. So what about Mr. Vishlia's assistant who made him sad by oking the first meeting? He laughed and said, “I think she'll do well, very well.”
If you buy through links in our articles, we may earn a small commission. This does not affect editorial independence.