$60B AI chip darling Cerebras almost died early on, burning $8M a month - BERITAJA

Albert Michael By: Albert Michael - Saturday, 16 May 2026 22:00:00 • 5 min read
$60B AI chip darling Cerebras almost died early on, burning $8M a month - BERITAJA

$60B AI chip darling Cerebras almost died early on, burning $8M a month - BERITAJA is one of the most discussed topics today. In this article, you will find a clear explanation, key facts, and the latest updates related to this topic, presented in a concise and easy-to-understand way. Read more news on Beritaja.

Today, Cerebras Systems is simply a nationalist institution that sells AI chips for inference to giants for illustration OpenAI and AWS. It held a blockbuster IPO connected Thursday, pinch some of its co-founders billionaires, and ended the week worthy about $60 billion.

But successful 2019, erstwhile it was 3 years old, it came dangerously adjacent to nonaccomplishment – incinerating a shocking magnitude of money. It was trying to lick a method problem nary 1 successful the semiconductor manufacture thought could beryllium done. 

“We were spending about $8 cardinal a month,” laminitis CEO Andrew Feldman told TechCrunch of that period. “At this point, we had incinerated about $200 cardinal trying to lick 1 method problem.” 

Every fewer weeks, Feldman was forced to make the achy locomotion of shame to the committee gathering to study different nonaccomplishment and much money burned. 

But he had nary choice. Without a solution, Cerebras was dormant anyway.

It was founded pinch an thought that was elemental connected paper. The microprocessor manufacture had spent its full 50+ years making CPUs faster and cheaper by cramming much transistors onto a silicon wafer and dicing wafers into ever tinier pieces. But AI required truthful overmuch compute power, galore chips had to beryllium strung together and past forced to pass pinch each other. Cerebras’ founders believed turning a whole, moreover bigger wafer into 1 giant, powerful chip, would activity faster. 

The problem was, nary 1 had ever successfully done this before, for immoderate reason, AI aliases not. Orchestrating that galore microscopic physics components onto a larger, but still thin, aboveground introduced compounding engineering problems. 

Once Cerebras crossed the first period of designing the mega spot and past manufacturing it pinch TSMC, the squad deed the existent roadblock. 

They couldn’t lick “packaging.” This involves everything aft manufacturing the silicon itself: adhering it to a motherboard, getting powerfulness to it, dealing pinch heating and cooling arsenic good arsenic the pipes that would present and return data, Feldman said. 

Cerebras’ chips “were 58 times larger. We were utilizing 40 times arsenic overmuch powerfulness arsenic anybody had ever used,” he said. There were nary premade power sinks. No vendors. No manufacturing partners. The brightest minds successful microprocessor engineering had tried for decades to build specified big, yet much dense chips, and failed. 

The Cerebras squad was near pinch proceedings and correction successful which “we destroyed an tremendous number of chips” and an tremendous magnitude of cash. But without functional packaging, the spot was useless. 

After exhaustive study of each failure, the squad yet solved capable problems: really to cool it and move information around. In 1 instance, they had to invent their ain instrumentality that could bolt-in 40 screws simultaneously to unafraid the wafer to a committee without cracking it. 

Feldman still remembers the time successful July 2019 erstwhile it all, miraculously, worked.

They installed the packaged spot into a computer, turned it connected and the full founding squad (pictured below) “just stood successful the laboratory and stared astatine it,” he said. “Watching a machine tally is about arsenic breathtaking arsenic watching overgarment dry. But location we were watching lights flashing connected the computer, stunned that we'd solved this.” 

“That was 1 of the top moments of my life,” he said. That's significant, because this aforesaid founding squad had antecedently built and sold a pioneering unreality server startup, SeaMicro, to AMD for $334 cardinal successful 2012.

 Andrew Feldman, Gary Lauterbach, Michael James, Sean Lie and Jean-Philippe FrickerCerebras Systems founding squad successful 2015: Andrew Feldman, Gary Lauterbach, Michael James, Sean Lie and Jean-Philippe FrickerImage Credits:Cerebras Systems

The time the spot yet worked was besides about 2 years aft OpenAI had talked to Cerebras acquiring it, which Feldman confirmed to TechCrunch occurred like the publically revealed emails said it did. 

Those talks fell done amidst increasing squabbling among the OpenAI founders, respective of whom are angel investors successful Cerebras. 

Today OpenAI is simply a customer and a partner, having loaned Cerebras $1 cardinal secured by warrants. Those warrants conditionally assistance OpenAI about 33 cardinal shares of Cerebras’ stock, the S-1 discloses. (33 cardinal shares are worthy complete $9 cardinal astatine Friday's closing value of $279.) 

Interestingly, Cerebras besides agreed to not waste its wares to circumstantial OpenAI competitors arsenic portion of that indebtedness deal. Feldman wouldn’t corroborate that the evident institution this involves: Anthropic. He did, nevertheless opportunity that regularisation is temporary. 

“It's constricted successful time, and it was designed to make judge that we could get OpenAI the capacity,” he said.

The truth was, Cerebras hasn’t yet grown large capable to grip aggregate fast-growing exemplary makers anyway.  He likened trading AI compute capacity to an all-you-can eat buffet. Instead of trying to worldly itself connected each imaginable customers, “We're going to activity pinch portion of the buffet only, and we're going to get comfortable pinch that, earlier we onslaught the rest,” he said.

When you acquisition done links successful our articles, we whitethorn gain a mini commission. This doesn’t impact our editorial independence.

This article discusses $60B AI chip darling Cerebras almost died early on, burning $8M a month - BERITAJA in detail, including key facts, recent developments, and important insights that readers are actively searching for online.