Today, Cerebras is worth about $60 billion 🚀. Its co-founders are billionaires. The IPO was a block buster. They sell inference chips to giants like AWS and OpenAI. It looks smooth now. It feels inevitable.
But in 2019 it was nearly dead.
They were burning $8 million a month. Not for growth. For survival.
“We were spending about $8 million on a technical problem that the industry said was impossible,” Andrew Feldman told TechCruch. He’s the founder and CEO. “We had incinerated nearly $0200 million trying solve it.”
The board meetings were painful. Every few weeks Feldman would walk in with another failure report. The cash was gone. The timeline was broken. But he had no choice to stop. Without the solution Cerebras didn’t exist at all.
The Wafer Delusion
The idea was simple. Simple enough to write on a napkin.
The CPU industry spent 50 years making transistors smaller. Cram them. Dice them. Repeat. Faster chips. Cheaper prices. Good. But AI didn’t want one fast chip. AI wanted a million connected chips talking to each other. Latency kills the magic.
Cerebras flipped the script. What if we don’t cut the wafer? What if the whole thing is the chip? One giant monolithic brain. No interconnect delays. Just pure compute.
Paper is forgiving. Silicon is not.
No one had ever done this. Ever. Putting that many microscopic components on a larger thin surface creates chaos. Compounding engineering errors. The team got the design done. They got TSMC to manufacture the silicon. Then they hit the wall.
Packaging.
Not just sticking the chip to a board. It’s everything after the factory floor. Power delivery. Cooling. Data pipes in. Data pipes out. Feldman said their chips were 58 times larger than the competition. They used 40 times the power.
There were no heat sinks for that size. No vendors who cared. No manufacturing partners willing to risk their reputation. The best engineers in the world had tried this for decades. They all failed.
Bolting the Beast
Cerebras went into trial and error mode. They destroyed an enormous number of wafers. Each one a few million dollars of wasted hope.
The chip was useless until it was packaged. Until it stayed cold. Until data could flow without frying the edges.
They analyzed every single failure. They kept going. Eventually they found the tricks. How to move the data. How to keep the beast cool. In one instance they literally invented a new machine. A custom tool to bolt in 40 screws at the same time. One screw early or loose means the wafer cracks. Crack the wafer lose the company.
Then came July 2019.
The packaged chip went into the computer. They hit power on. The whole founding team just stood there. Stared at it.
Watching a computer run is like watching paint dry. We stared at flashing lights, stunned we’d solved this.
That day felt bigger than any product launch. Feldman called it one of the greatest moments of his life. Which is saying something. This same team sold their first startup, SeaMicro to AMD for $335 million in 2012 💰. They knew money. They knew exit strategy. This felt different.
The OpenAI Connection
Timing matters. The chip worked roughly two years after OpenAI talked to Cerebras about acquiring it.
Feldman confirmed those talks happened. Exactly as the leaked emails showed. The deal died. Squabbling among the founders. Misaligned visions. Messy business.
Today OpenAI isn’t an owner. They’re a partner. A lender.
They loaned Cerebras $1 billion secured by warrants. That gives them about 33 million shares. At Friday’s closing price that’s worth over $9 billion. Big number.
But there’s a catch.
As part of that loan Cerebras agreed not to sell to certain competitors. Everyone knows who the obvious one is. Anthropic 👀.
Feldman didn’t say the name. But he didn’t deny it either.
“The restriction is limited in time,” he said. “It was designed to make sure OpenAI gets the capacity they need.”
Let’s be honest. Cerebras isn’t big enough to feed everyone anyway. Feldman compares it to an all-you-can-eat buffet. You can’t just grab every item off every plate. You stall. You overwhelm yourself. You fail.
“We’re going to work with a part of the buffet first. Get comfortable. Then attack the rest.”
It’s a strategic constraint. A temporary handcuff to secure the future. Does that make you nervous about your competition? Maybe.
The buffet is huge. The lines are moving fast. Cerebras picked a seat at the table. But the food isn’t infinite. And neither is their exclusivity window.
