Cerebras nearly collapsed in its early years despite soaring to a $60B AI chip valuation

Today, Cerebras Systems is a publicly traded company supplying AI inference chips to major technology firms, including OpenAI and Amazon Web Services. The company delivered a major IPO on Thursday, making both of its co-founders billionaires and pushing its valuation to roughly $60 billion by the end of the week.

However, the company’s current success stands in sharp contrast to the crisis it faced during its early years. Back in 2019, just three years after its founding, Cerebras came extremely close to collapse while burning through staggering amounts of capital in pursuit of a technical breakthrough that much of the semiconductor industry believed was impossible.

Founder and CEO Andrew Feldman told TechCrunch that during that period, the company was spending nearly $8 million every month.

“At this point, we had incinerated nearly $200 million trying to solve one technical problem,” Feldman said while describing the company’s struggles at the time.

The pressure on leadership became intense as the failures continued. Every few weeks, Feldman had to appear before the company’s board to explain why more money had been spent without a solution in sight.

Still, according to Feldman, the company had little choice but to continue pushing forward, as failure to solve the issue would have effectively ended Cerebras.

The startup was founded around an idea that sounded straightforward in theory. For more than five decades, the semiconductor industry improved processors by packing increasing numbers of transistors onto silicon wafers and then slicing those wafers into smaller chips. Cerebras believed artificial intelligence workloads demanded a different approach altogether.

The founders argued that instead of connecting many smaller chips for AI computing, the industry could build a single giant chip from a single wafer. They believed that a single massive processor could operate far faster and more efficiently than clusters of smaller, interconnected chips.

The challenge was that nobody had ever successfully turned an entire silicon wafer into a single operational chip at that scale. Attempting to coordinate such a large number of microscopic electronic components on a single thin surface posed enormous engineering challenges.

After Cerebras succeeded in designing and manufacturing the wafer-scale processor in partnership with TSMC, the company encountered an even bigger obstacle.

The real problem became “packaging,” which includes all the work that happens after the silicon itself is produced. That process involved attaching the chip to a motherboard, supplying power, managing heat and cooling systems, and handling the flow of incoming and outgoing data.

Feldman explained that Cerebras’ processors were dramatically larger than anything the industry had previously attempted.

“Our chips were 58 times larger. We were using 40 times as much power as anybody had ever used,” he said.

Because of the chip’s unprecedented scale, there were no ready-made heat sinks, specialised suppliers, or established manufacturing partners capable of supporting the project. According to Feldman, some of the most experienced engineers in the microprocessor industry had spent decades attempting to create extremely large, dense chips and had failed repeatedly.

That left the Cerebras engineering team relying on trial and error. During the process, the company destroyed a huge number of prototype chips while continuing to spend enormous amounts of money.

Without a successful packaging solution, however, the chip itself would have remained unusable regardless of its processing power.

Following repeated failures and extensive analysis of what went wrong each time, the company gradually solved the core challenges involving cooling systems and data movement. At one point, the team even had to build a custom machine capable of tightening 40 screws simultaneously to attach the wafer to a board without cracking the delicate silicon.

Feldman still recalls the moment in July 2019 when the technology finally worked.

After installing the packaged wafer-scale chip into a computer and powering it on, the entire founding team gathered in the lab and watched the system run successfully for the first time.

“We just stood in the lab and stared at it,” Feldman said. “Watching a computer run is about as exciting as watching paint dry. But there we were watching lights flashing on the computer, stunned that we’d solved this.”

He described the breakthrough as one of the greatest moments of his life.

That statement carries added significance because the same founding team had already built and sold an earlier cloud server startup, SeaMicro, to AMD for $334 million in 2012.

Interestingly, the successful launch of Cerebras’ chip also came around two years after OpenAI reportedly explored acquiring the company. Feldman confirmed that acquisition discussions between Cerebras and OpenAI had indeed taken place, aligning with previously revealed emails about the talks.

According to Feldman, those discussions ultimately collapsed amid growing disagreements among OpenAI’s founders. Several OpenAI founders are also angel investors in Cerebras.

Today, the relationship between the two companies has evolved significantly. OpenAI is now both a customer and strategic partner of Cerebras and has reportedly provided the company with a $1 billion loan secured through stock warrants.

According to Cerebras’ S-1 filing, those warrants could eventually grant OpenAI approximately 33 million shares of Cerebras stock. Based on Friday’s closing share price of $279, those shares would currently be valued at more than $9 billion.

The filing also revealed that Cerebras agreed not to sell its technology to certain OpenAI competitors as part of the financing arrangement.

While Feldman declined to confirm directly whether the restriction involved Anthropic, he stated that the limitation is only temporary.

“It’s limited in time, and it was designed to make sure that we could get OpenAI the capacity,” Feldman explained.

Feldman also acknowledged that Cerebras is not yet large enough to fully support multiple rapidly expanding AI model developers simultaneously.

He compared the AI computing market to an all-you-can-eat buffet, explaining that the company is intentionally focusing on serving only a portion of potential customers for now rather than trying to satisfy everyone immediately.

“We’re going to work with part of the buffet only, and we’re going to get comfortable with that, before we attack the rest,” he said.