Cerebras Systems entered the public markets this week with a valuation of about $60 billion after a blockbuster IPO, but the AI chip company nearly failed in 2019 while trying to solve engineering problems that much of the semiconductor industry believed were impossible.
The company, which now supplies AI inference chips to customers including OpenAI and Amazon Web Services, spent years developing a wafer-scale processor architecture that combined an entire silicon wafer into a single chip instead of cutting wafers into many smaller processors.
According to founder and chief executive officer Andrew Feldman, the company was burning through approximately $8 million per month during the critical development period.
“At this point, we had incinerated nearly $200 million trying to solve one technical problem,” Feldman told TechCrunch.
The company had reached a stage where repeated technical failures threatened its survival. Feldman said he regularly attended board meetings to explain additional setbacks and escalating costs.
Without solving the remaining engineering barriers, the company would not have been able to commercialize its chip technology.
Cerebras Attempted To Build A Single Giant AI Processor
Cerebras was founded around the idea that artificial intelligence workloads required far more computing power than traditional semiconductor architectures could efficiently provide.
For decades, the microprocessor industry improved performance by shrinking transistors and dividing silicon wafers into smaller chips.
Cerebras instead pursued the opposite approach by transforming an entire wafer into one massive processor.
The company believed that a single larger chip could process AI workloads more efficiently than connecting many smaller chips together through networking systems.
However, no company had previously succeeded in commercializing a wafer-scale processor of that size.
After designing the chip and manufacturing it with TSMC, Cerebras encountered major difficulties involving packaging and deployment.
Feldman said the company struggled with attaching the chip to motherboards, distributing electrical power, cooling the processor, and building systems capable of moving data efficiently through the hardware.
The chips were substantially larger and more power-intensive than conventional processors.
“Our chips were 58 times larger. We were using 40 times as much power as anybody had ever used,” Feldman said.
Because no existing vendors manufactured components suitable for the system, Cerebras engineers had to develop custom solutions internally.
The company also destroyed numerous chips during testing while attempting to solve packaging and cooling problems.
Engineering Team Eventually Solved Cooling And Packaging Challenges
After repeated failures and analysis, Cerebras engineers eventually developed systems capable of cooling the chip and transferring data reliably.
One challenge required the company to invent a custom machine capable of tightening 40 screws simultaneously to secure the wafer to a circuit board without cracking the silicon.
Feldman said the breakthrough came in July 2019, when the founding team successfully powered on a completed system 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.”
“That was one of the greatest moments of my life,” he added.
Before founding Cerebras, the same leadership team had previously created cloud server startup SeaMicro, which was acquired by AMD for $334 million in 2012.
OpenAI Acquisition Talks Failed Before Partnership Expanded
The successful launch of Cerebras’ chip platform came roughly two years after OpenAI reportedly discussed acquiring the startup.
Feldman confirmed to TechCrunch that acquisition discussions between the companies occurred as previously described in publicly released emails.
According to Feldman, the talks ultimately collapsed amid disagreements among OpenAI’s founders. Several OpenAI figures later became angel investors in Cerebras.
Today, OpenAI is both a customer and financial partner of the company.
According to Cerebras’ S-1 filing, OpenAI loaned the company $1 billion secured by warrants that could conditionally convert into approximately 33 million shares of Cerebras stock.
Based on Friday’s closing share price of $279, those shares would currently be worth more than $9 billion.
The filing also disclosed that Cerebras agreed not to sell products to certain OpenAI competitors for a limited period as part of the financing arrangement.
Feldman declined to confirm whether Anthropic was among the restricted companies.
However, he said the limitation was temporary and designed to ensure OpenAI had access to sufficient computing capacity.
“It’s limited in time, and it was designed to make sure that we could get OpenAI the capacity,” Feldman said.
Company Says It Is Expanding Gradually Despite Growing AI Demand
Feldman said Cerebras currently remains too small to simultaneously support every major AI model developer seeking computing infrastructure.
He compared the company’s approach to limiting customers at an all-you-can-eat buffet.
“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.
Featured image credits: Magnific.com
For more stories like it, click the +Follow button at the top of this page to follow us.