Cerebras Waited Ten Years for a Market That Finally Needed Speed

On No Priors, Sarah Guo and Elad Gil sit down with Andrew Feldman, co-founder and CEO of Cerebras, not long after the company listed. The hosts put the market value around $63 billion, with Guo citing about $60 billion on the day of the taping. Cerebras, founded in the mid-2010s around new machine-learning workloads, now sells AI computers optimized for those jobs. The product that matters in this conversation is inference: using a trained model, not training it. Feldman says they are not a little faster than GPUs. They are 15, 18, even 20 times faster.

That lead, he says, is across the board: big models and small ones, American models and Chinese models, trillion-parameter networks and billion-parameter ones. Starting around 2025, models got smart enough to be useful. People stopped pointing at AI and started folding it into the workday, which is when speed stopped being a novelty. Cerebras was crushed with demand. By the end of that year they had signed a deal with OpenAI that Feldman calls one of the biggest in Silicon Valley, north of $20 billion. In March they signed with AWS to put machines in Amazon’s data centers. The year and a half since has been a chase to meet supply.

Guo first met Feldman around 2016, when saying “AI” still sounded a little strange. The models of that moment were convolutional nets and RNNs, with GANs just emerging, and the party trick was telling a chair from a cat. Feldman laughs at the distance traveled. The company that was trying to classify furniture is now a tens-of-billions public name. The years in between, this episode argues, were not a straight ramp. They were a long stretch of being right too early, then a sudden flood toward the people who already had the fastest machine.

Plenty of people, Feldman admits, said the architecture was weird and that Cerebras had called it wrong. His reply is the design rule he wants remembered. You will not get 15 or 20 times better than a GPU with a minor modification of their architecture. If you aspire to a radical improvement, the design has to be different. From day one they chose wafer scale: a 46,000-square-millimeter chip, the size of a dinner plate, while everyone else built postage stamps. The industry listed reasons it was impossible. In 2019 they proved it was possible, started delivering, and kept improving. They were fast while AI was still a novelty. When something is a novelty, nobody cares if you are fast, because nobody is using it.

The conviction to build against an empty market, he says, was a mix of vision, the right co-founders, a little arrogance, and a little luck. As computer architects they treated a new workload as an opening. Entering the x86 world, where nothing new had happened for generations, looked hopeless. Graphics produced the discrete GPU and Nvidia. Mobile compute produced ARM. Intel and AMD, the companies that looked best positioned, took no share. So the bet was that AI would eat a lot of compute, would need a dedicated architecture, and that architecture could not be a derivative of what already existed. Those bets were 100 percent contrarian. They also, he says, turned out to be dead right.

There were years when it did not feel that way. Wafer-scale computing had been attempted across the seventy-year history of the industry and had failed, including by Gene Amdahl, one of the names Feldman puts on computing’s Mount Rushmore. From roughly mid-2017 to mid-2019, Cerebras could not build the thing. They were spending about $8 million a month. Every six weeks they told the board it still was not working. Each failure analysis made the next attempt a little better. In the summer of 2019 it yielded. They sat in a makeshift office in downtown Los Altos, in a building not designed for hardware people, staring at a computer as exciting as watching paint dry. It was working. For half an hour nobody could speak. Nobody had been able to do this, and they had.

Solving what Feldman calls the hardest problem in the computer industry produced a shrug. First generation, they might have sold a dozen. Second generation, maybe 300. Third generation, they are looking at tens of thousands. For two or three years they were ahead of the market, blisteringly fast, and absolutely nobody cared. From about 2023 to the beginning of 2025, people pointed at AI without putting it into daily work. Once you use a tool every day, it cannot be slow. How long will you wait for a website to resolve? Guo says she has no attention. Feldman’s punchline is the size of the leftover markets: slow search, zero; dial-up internet, zero; slow inference, the same. They had to wait until the models were smart enough to be useful. That, he says, is 2025, and why Cognition, Cursor, Lovable, OpenAI, and a swarm of other companies, including ones the hosts had invested in, exploded toward them.

New computer architectures, he argues, have a worn path. You start in supercomputing, because those customers love speed and do not mind immature software. Cerebras ran that table: national labs, Lawrence Livermore, Sandia, Europe’s EPCC and LRZ. Then oil and gas, then pharma, industries already accustomed to burning extraordinary amounts of compute. Historically a chasm sits after that, because none of those buyers supply mainstream volume. The company that bridged it was the Abu Dhabi group G42, which became a strategic partner and placed a billion-dollar order. That order let Cerebras change its supply chain and deploy clusters big enough to battle-test at scale. Hardware has a structural bind: your QA lab cannot be as large as the customers you want. You cannot park a hundred million dollars of your own gear in-house. With G42 they trained models, ran inference, and had what Feldman calls extraordinary partners in CEO Peng and chairman Sheikh Tahnoun. When OpenAI and AWS arrived, Cerebras had capacity and a bridge across the chasm.

Elad Gil underlines the path dependence. Going from tens or hundreds of millions in orders to $20 billion of backlog is not a jump. It is years of work in the middle. Software companies can scale almost overnight. Hardware cannot. If you want to double, you call the contract manufacturer, they find power, rent a building, add lines, make test fixtures, and every step takes real time. Cerebras is trying to increase manufacturing ten times this year, which Feldman says is about as fast as anyone in the history of hardware. Software had its own decade. When they started, co-founder Gary told Feldman a compiler would take about ten years. Feldman called that big-company talk and said five. It took about ten. Compilers are extraordinarily difficult software. They have a good stack now. The slowness was not a vibe. It was the actual clock of making physical things and the software that talks to them.

Inside Cerebras, the same revolution is eating engineering time. Eight months before the taping, they were not spending a thousand dollars per engineer on tokens. They are now at $25,000 or $30,000, and it is ripping. Feldman is honest that it is not useful for everybody. A small group has the mindset for it: they run eight or ten agents around the clock, have shifted from writing code to governing agents, put a QA agent on the output, and treat model weaknesses as a puzzle, verbosity, comments getting cut. Those people have gone from 10x engineers to 100x. Everyone else, the CEO included, is limping along, trying to make the tools work for being a CFO, an accountant, or someone in marketing, and showing the rest of the company the best practices of the few who have cracked it.

The company is about 800 to 850 people, which the hosts note is a lot of market cap per head. With a backlog north of $20 billion, delivery is the daily job, but Feldman is more worried about culture. Companies that grow to one, two, or three thousand people often stop taking the risks they used to take. A fearless engineering culture turns into “what can we get into the next rev.” He wants to hire people who do fearless work and to guard a rule he states without hedging: they would much rather fail in pursuit of the extraordinary than succeed in the ordinary. The other death is hiring. Openings multiply, and it becomes easy to put a reasonably good body in a seat. He spends a meaningful part of every day talking to candidates so that does not happen.

Being CEO, he says, is extraordinarily lonely, especially if you pick problems everyone else says cannot be solved. The chip on the shoulder is fuel: they say it cannot be done, and in your head you say you cannot solve it. You also have to love the building. Doing this for the money is a horrible reason. There are easier ways to get rich than creating something extraordinary and competing with someone as strong as Nvidia. “I’m a professional David. This is my fifth startup. I compete against Goliath.” Every dollar they sell, he tells himself, their brains took from muscle that would have grabbed it in a heartbeat. If you do not love that, it is a very long road.

Gil asks the founder question that sits under all of this: when do you give up? One camp says keep going no matter what. The other says you should keep reassessing, and that quitting can be the smartest move. Feldman wants hypotheses written down in advance about what it will take to win. When they all come back negative, it is time. Gil’s worry is the sequential version: one more test, it fails, one more test. The slippery slope, Feldman says, is a beast in ethics and in company-building. Former CEOs or seasoned entrepreneurs on your side can pull you back by reminding you of the sentence you uttered a year ago. The frog in warm water is the same story: you said you would get out at a certain heat, and the water kept rising. Put a time frame on the things that have to change. Lots of efforts, he thinks, ought to be truncated so people can redeploy. Time is the expensive input. Life’s opportunity cost is high.

Why list now? Feldman’s version is unusually plain. Going public is exchanging professional technology investors for a different class of investors and, in the process, shaving a little off the cost of capital. You go from people like the hosts to his dad, and in return you agree to be governed by extraordinarily stringent rules. The question is complicated, he says, by a historical novelty: four or five companies can now raise huge amounts of money without going public. That was not a thing before OpenAI, Anthropic, and maybe Databricks. Silicon Valley’s four-year option vest, the hosts note, comes from an era when four years was how long it took to get public and to a valuation in the hundreds of millions. Cerebras took ten.

A decade changes the social contract with employees. Cerebras opened the secondary market so people could sell along the way. If you are going to bet a big chunk of a career with the company, Feldman thought modest liquidity along the path was reasonable. For a very small number of firms, those three in particular, public-market money at public-market valuations is available in private. For everyone else, if you want the high valuations and the legitimacy, large American companies still like doing business with other public companies. Audited books, and the ability to see who you are, are different from being private. Cerebras also thought it could offer the public market something unique: for a period, the first and only AI pure play, a company whose revenue is 100 percent this market, with no gaming, no graphics, no PC. More than anything, he frames the listing as graduating from corporate adolescence to corporate adulthood.

The OpenAI deal, in Feldman’s telling, started with a call to Sam Altman in the middle of summer 2025. Altman said for the first time that they had been trying so hard just to keep up with demand, and that they now saw the importance of fast inference. Trials followed. Cerebras was so much faster than the competition that the other side understood immediately. Feldman has a consumer rule he will not break: if his mother buys it or uses it, he does not want to make it or sell it. He wants super-smart customers doing interesting things with the machines. The night before Thanksgiving they signed a term sheet. On December 24 they signed a master agreement. A $20 billion-plus deal in four and a half weeks, several law firms on the other side, Cerebras working seven days a week. Gil calls the whole market’s hunger to keep up with demand something he has not personally seen before.

Similar compressions are everywhere. Cognition bought Windsurf over a weekend. Elon Musk puts up data centers at a pace everyone says is impossible, unless you are him. You cannot buy a $300 million company in three days, except you can. You cannot do a deal like this in 24 days, except if you work eight or ten hours every day, you can. What people thought was the speed of light was not. The art of the possible has been expanded by this push in a way Feldman says he never would have expected. Ambition for speed is an advantage only if you believe the faster thing is possible. Cursor and Cognition show growth rates that were not supposed to exist. Truncated aspirations, he thinks, were the real constraint.

Open source, in his view, fed the market. When closed-source models were too expensive, the open-source community kept the flame going and pushed the closed labs. Techniques from Chinese makers were a “we cannot rest on bigger clusters and more data” alarm. That made a vibrant ecosystem, and it is fun, as an infrastructure person, to watch other people’s ideas take flight on your hardware. If you do not love that, the job is wrong. Guo mentions Devin on Cerebras as a magical high-speed coding experience. The punchline is not a slightly better version of today’s products. Netflix used to mail DVDs in envelopes and thought Blockbuster was the competition. When the internet got fast, they became a movie studio. Speed did not make DVD logistics more efficient. It opened an entirely different business. The PC first replaced typewriters and ledgers; the productivity jump came later, when work reorganized around the cloud and SaaS put tools in reach that a small company could never have afforded. AI is in the replacement phase now, coding, design, some SaaS. Feldman is waiting for the reorganization, the new business models, and the real jump in productivity.