Currently, Cerebras Systems operates as a publicly traded entity that produces AI chips for inference, catering to industry leaders such as OpenAI and AWS. It launched an extraordinary IPO on Thursday, with its co-founders both becoming billionaires, and concluded the week valued at around $60 billion.

However, back in 2019, just three years after its inception, it came perilously close to collapse – burning through an astonishing amount of capital. The company was attempting to tackle a technical challenge that many in the semiconductor field believed to be insurmountable. 

“We were expending approximately $8 million monthly,” founder and CEO Andrew Feldman recounted to TechCrunch about that era. “At this juncture, we had effectively eliminated nearly $200 million in the quest to resolve a single technical issue.” 

Every few weeks, Feldman was compelled to make the difficult trek of shame to the board meeting to relay yet another setback and the additional funds consumed. 

But he was left with no alternatives. Without a resolution, Cerebras was doomed regardless.

The company was established with a concept that seemed straightforward on paper. The microprocessor sector had been dedicated for over 50 years to enhancing CPU speed and reducing costs by cramming more transistors onto silicon wafers and slicing those wafers into progressively smaller sections. Yet, AI required such substantial computational power that numerous chips needed to be interconnected and made to communicate effectively. The founders of Cerebras believed that transforming an entire, significantly larger wafer into a single massive, powerful chip would operate more efficiently. 

The challenge was that no one had ever accomplished this successfully before, for any purpose, be it AI or otherwise. Coordinating so many minute electronic components onto a larger, albeit still slim, surface introduced a cascade of engineering challenges. 

Once Cerebras passed the initial hurdle of designing the mega chip and subsequently producing it with TSMC, the team encountered the real impediment. 

They were unable to tackle “packaging.” This encompasses everything that follows the manufacturing of the silicon itself: attaching it to a motherboard, supplying power, managing heating and cooling, along with the channels that would carry and retrieve data, Feldman explained. 

Cerebras’ chips “were 58 times larger. We were utilizing 40 times more power than anyone had ever used,” he stated. There were no pre-fabricated heat sinks. No suppliers. No manufacturing collaborators. The most brilliant minds in microprocessor design had for decades endeavored to create such large yet denser chips, yet had failed. 

The Cerebras group was left with a cycle of trial and error during which “we ruined an incredible number of chips” and a vast amount of funding. But without effective packaging, the chip would be worthless. 

After thorough evaluation of each failure, the team eventually resolved sufficient challenges: how to cool it and circulate data. In one case, they had to innovate their own machine that could fasten 40 screws simultaneously to secure the wafer to a board without causing damage. 

Feldman still recalls the miraculous day in July 2019 when everything came together.

They placed the packaged chip into a computer, powered it on, and the entire founding team (seen below) “simply stood in the lab and stared at it,” he remarked. “Observing a computer operate is about as thrilling as watching paint dry. But here we were, watching lights blinking on the computer, amazed that we had accomplished this.” 

“That was one of the most significant moments in my life,” he expressed. This is noteworthy as this same founding team had previously created and sold a groundbreaking cloud server venture, SeaMicro, to AMD for $334 million in 2012.

The day the chip finally functioned was also roughly two years following discussions between OpenAI and Cerebras regarding a potential acquisition, which Feldman confirmed to TechCrunch occurred as detailed in the publicly disclosed emails. 

Those negotiations fell apart amid rising conflicts among the OpenAI founders, several of whom are angel investors in Cerebras. 

Currently, OpenAI serves as both a customer and a collaborator, having provided Cerebras with a $1 billion loan secured by warrants. These warrants conditionally grant OpenAI approximately 33 million shares of Cerebras’ stock, as revealed in the S-1 filing. (33 million shares equate to over $9 billion at Friday’s closing price of $279.) 

Interestingly, Cerebras also committed to refrain from selling its products to select OpenAI rivals as part of that loan arrangement. Feldman did not confirm that the apparent entity this pertains to is Anthropic. However, he noted that this restriction is temporary. 

“It’s time-limited and was established to ensure that we could provide OpenAI with the necessary capacity,” he stated.

The reality is, Cerebras has yet to scale sufficiently to accommodate multiple rapidly expanding model developers anyway.  He compared selling AI computing capacity to an all-you-can-eat buffet. Rather than attempting to fill itself with all possible clients, “We’re going to focus on a portion of the buffet first, and gain confidence with that before expanding to others,” he said.