$60B AI chip darling Cerebras nearly died, burning $8M a month

Today, Cerebras Systems is a public company they sell AI chips to explain it for giants like OpenAI and AWS. It caught a blockbuster IPO on Thursdayand both co-founders are billionaires, and it ended the week with a total of about $60 billion.

But in 2019, three years old, it almost failed – costing a lot of money. It was trying to solve a technical problem no one in the semiconductor industry thought could do.

“We spend about $8 million a month,” co-founder Andrew Feldman told TechCrunch at the time. “At this point, we had burned about $200 million trying to solve one technical problem.”

Every few weeks, Feldman was forced to make a painful trip of shame to a board meeting to report another failure and burnout.

But he had no choice. Without an answer, Cerebras was still dead.

It was founded with a simple idea on paper. The microprocessor industry has spent its entire 50+ years making CPUs faster and cheaper by putting more transistors on a silicon wafer and cutting them into smaller pieces. But AI required a lot of computing power, a lot of chips had to be wired together and forced to communicate with each other. The founders of Cerebras believe that turning an entire cross, even a large one into one powerful giant, works faster.

The problem was, no one had done this successfully before, for whatever reason, AI or not. Building a large number of small electronic devices on a large, but thin, surface presented technical challenges.

When Cerebras crossed the threshold of first developing a mega chip and then developing it with TSMC, the team hit the right track.

He could not solve the “installation”. This includes everything after making the silicon itself: gluing it to the motherboard, getting power, handling heat and cooling and the pipes that will bring and return data, Feldman said.

Brain cells “were 58 times bigger. We were using 40 times more energy than anyone else has ever used,” he said. There were no pre-programmed heating changes. There are no dealers. There are no co-producers. The brightest minds in microprocessor engineering have tried for years to make larger, but denser chips, and failed.

The Cerebras team was left with trial and error where we “spent a lot of chips” and a lot of money. But without a working insert, the chip was useless.

After a thorough investigation of each failure, the team solved the problems of the complete: how to cool and move the data around. At one point, they had to build their own machine that could fasten 40 screws at a time to secure a cabinet to a board without tearing.

Feldman still remembers the day in July 2019 when everything miraculously worked.

He put the packaged chip into the computer, turned it on and the entire startup team (pictured below) “just stood in the lab and looked,” he said. “Watching a computer run is as exciting as watching paint dry.” But we were watching the flashing lights on the computer, and we were amazed that we had solved the problem.

He said: “That was the most exciting time of my life. It’s important, because the same team already built and sold the first cloud computing system, SeaMicro, to AMD for $334 million in 2012.

The day the device went live was also two years after OpenAI approached Cerebras about acquiring it, Feldman confirmed to TechCrunch. such as publicly disclosed emails it said.

The talks ended as tensions grew between OpenAI’s founders, many of whom were angel investors at Cerebras.

Today OpenAI is both a customer and a partner, having loaned Cerebras $1 billion secured by documents. The licenses give OpenAI about 33 million shares of Cerebras, the S-1 reveals. (The 33 million shares are worth more than $9 billion at Friday’s closing price of $279.)

Interestingly, Cerebras also agreed not to sell its products to OpenAI competitors as part of the loan. Feldman would not confirm which company this identification includes: Anthropic. He said, however, that the ban is temporary.

“It’s timeless, and it’s designed to make sure we can access OpenAI,” he said.

The truth was, Cerebras was not big enough to produce many fast-growing species. He compared the sale of artificial intelligence to an all-you-can-eat buffet. Instead of trying to serve all the customers themselves, “We will work only with the buffet section, and we will be comfortable with it, before we attack the rest,” he said.