Part IThe Story
Watching Paint Dry on San Antonio Road
The computer did nothing interesting. That was the point.
In July or August of 2019 (Andrew Feldman, telling the story years later, has never quite pinned down which month), the founders of Cerebras Systems stood in a tiny lab in Los Altos and watched a machine run. Nothing about it was cinematic. It was, in Feldman's words, "just a big metal box with some lights flashing," an experience "about as exciting as watching paint dry." The lights blinked. Five men who had worked together for the better part of a decade looked at each other, and, as Feldman puts it, "we were stunned."
What had stunned them was the thing inside the box: a single piece of silicon roughly the size of a dinner plate. It carried 1.2 trillion transistors and 400,000 processing cores, and it was the first commercially workable wafer-scale processor in the history of computing. For seventy years the industry had treated the idea as a graveyard. Gene Amdahl, one of the architects of modern computing, had tried it and failed. The conventional chip was the size of a postage stamp, or a thumbnail, and it stayed that size because of a manufacturing boundary called the reticle limit, along with the brutal arithmetic of defects. Print a chip big enough and the odds that a flaw lands somewhere on it approach certainty. So you cut the wafer into many small chips, throw away the bad ones, and wire the survivors together. Everyone did this. It was physics plus accounting, and nobody argued with it.
Cerebras argued with it, at a cost of roughly $8 million a month over a stretch that Feldman has variously described as fifteen months, eighteen months, and two years. The length of the desert depends on who is asking and how long ago it was. In that time they spent about $200 million on a single problem, packaging: how to attach an enormous slab of silicon to a board, feed it power, cool it, and get data in and out without the whole assembly tearing itself apart as it heated and expanded. Every six or eight weeks Feldman walked into a board meeting with the same report. Nope, still can't make it.
The company was then quartered on San Antonio Road. Visitors in early 2020 found no receptionist, just an iPad that asked them to sign a nondisclosure agreement, and across from it a skeleton dressed in Cerebras merchandise to keep them company. A few blocks down the same road, where it crosses into Mountain View, is 391 San Antonio Road. In 1956 that address housed Shockley Semiconductor Laboratory, the first silicon device company in what would become Silicon Valley. William Shockley's lab is where the Valley started, and it is where the eight engineers who left him to found Fairchild Semiconductor started too.
Feldman grew up nearby. In the summer of 2026 he sat for a podcast whose final segment carried the plain title "Andrew's childhood next to William Shockley." By then he was 56 and newly a billionaire, and the company that had spent two years unable to turn its chip on had just completed the largest American technology IPO since 2019. It is tempting to call this a homecoming. That framing assumes he ever left, and the most interesting thing about Andrew Feldman is how far he went in trying to.
By the Numbers
The Wafer and the Man
4 trillionTransistors on a single WSE-3 wafer
900,000AI-optimized cores on one chip
$8MMonthly burn during the years Cerebras couldn't build one
$5.55BRaised in the May 14, 2026 Nasdaq IPO
86%Share of 2025 revenue from two Abu Dhabi customers (MBZUAI and G42)
$510M2025 revenue, with $87M in net income
8Colleagues from his first startup still with him 25+ years later
Plumbing
Start with the tennis court, since Feldman does.
Throughout his childhood his father played doubles on Saturday and Sunday mornings in a rotation of six friends, and the boy watched. Both of his parents were professors at Stanford, and the family lived on campus. By the time Feldman finished high school, three of those six tennis partners had won Nobel Prizes. "Imagine a Nobel laureate," he wrote decades later, "making time for an sixth grader with a biology question." He took two lessons from the court and still uses both. The first is that once you have seen how the best minds work a problem, "how the complex can be explained simply," your standard moves for good. The second is "that truly exceptional need not be arrogant." He also learned, he notes, "how not to hit a backhand."
The headline he gave the post was "I learned more about business from watching tennis than from my Stanford MBA." For a man who holds two Stanford degrees and grew up inside the university's fence, the sentence carries some quiet defiance. Stanford supplied the environment. The education happened at the edges of it, in the gap between sets.
He earned a B.A. in economics and political science from Stanford in 1991. He did not want to build things. He wanted to become a professor of development economics, and to that end he went to work in Pakistan's Northwest Frontier Province, which he later described as "about as far from Silicon Valley as you can get." Then he came home, enrolled in a PhD program, finished his coursework and qualifying exams, "Then I got bored." He decided to add an MBA, figuring he would end up teaching at a business school. Stanford again: the MBA came in 1997.
His housemate Jim was joining a tiny startup. The startup asked Feldman to join too. He declined. They asked a second time and he declined again. "I had no interest in data networking," he recalled. "I thought it was plumbing." He eventually gave in far enough to write their business plan, and then he joined. A former colleague who left a comment on Feldman's LinkedIn post identifies the company as YAGO Systems and the year as 1998, and the story matches what Bloomberg has reported: Feldman was the first non-engineering employee of a hardware networking startup that sold for $280 million while he was still in his twenties. Feldman says the sale came "less than a year later." Bloomberg says eighteen months. In either version it was the moment the economist discovered he loved the plumbing.
"We built chips and systems for switching and routing," he wrote. "We played a small role in changing the world." Then comes the sentence that matters most in this profile, because it explains everything after it: "8 of those guys are still with me, more than 25 years later, at Cerebras today."
He calls the tidy, retroactive career narrative "the myth of LinkedIn," the idea "that we all had our careers mapped from the time we left high school." His own path was "a series of accidents where I fell into things I enjoyed." That sounds humble, and it is. It is also slightly untrue in the way any good founder's story is untrue. The accidents kept landing him in the same place, among the same people, working on the same layer of the stack: the unglamorous infrastructure that other people's ideas run on.
Sometimes Feldman frames it as a matter of identity. "As a chip and system guy, that's all we know," he told the podcaster Nataraj Sindam in 2026. "We weren't going to build a web app; that's not who we are." It is a strange declaration from a man who once planned to spend his life studying poverty in South Asia. Identity here was not given at birth. It was adopted, in a shared house in Palo Alto, after he had said no twice.
The Man Who Was Not a Technical Visionary
In 2011, while running his fourth company, Feldman gave an interview to Electronic Design that contains the most precise self-portrait he has ever offered.
— Andrew Feldman, Electronic Design, June 2011I'm not a technical visionary, I'm a person who sees market opportunities and identifies technical talent.
Set that line beside his later claim to be "a chip and system guy," with the wafer-scale engine sitting between the two statements like a hinge, and you have the man's central puzzle. He is an economist who learned to talk to engineers and a marketer who built the largest chip in history. His job title has always been the business side, and his reputation now rests on an engineering feat. He resolves the contradiction by refusing to claim the feat. "There is much on which I am not an expert," he wrote in 2026. "And I never claim to be. When I don't know, I say I don't know."
The career that led there was marketing and corporate development at the switching and routing layer. From March 2000 to August 2003 he was vice president of corporate marketing and corporate development at Riverstone Networks, there "from inception through IPO" in 2001, near the very top of the networking bubble. From August 2003 to December 2006 he was vice president of marketing and product management at Force10 Networks, which Dell later bought for a reported $800 million.
Riverstone also left a mark that does not appear in the conference-bio boilerplate. As Bloomberg reported at the time of the IPO, a Cerebras filing disclosed that Feldman pleaded guilty in 2007 to a federal felony charge related to circumventing accounting controls during his time at Riverstone. He received probation and a monetary fine. The speaker introductions that circulate at IEEE and university events, which repeat the same paragraph about SeaMicro and Force10, leave it out. The S-1 could not. The resulting profile is a man whose professional writing now dwells, at length and with evident sincerity, on transparency and on never surprising the people who trust you with their money, while carrying a felony plea from the moment he learned what happens when the numbers are made to say something they shouldn't. Whether the later convictions grew out of the earlier conviction is something only Feldman could answer, and the record does not show him answering it. The juxtaposition sits there regardless.
Next came a period of waiting. In 2007 he was an entrepreneur-in-residence at Crosslink Capital and U.S. Venture Partners, and from that seat he watched what he estimates were hundreds of pitches, ten a week for six or eight months. The lesson he took from them was counterintuitive:
"What surprised me was that the more senior and the more credible the CEO, the more likely they were to answer with 'I don't know' or 'we aren't sure' or 'we need to test that hypothesis.'"
He noticed "the calming effect on the room" when a CEO admitted ignorance and then described how to resolve it. The amateurs bluffed and the veterans didn't. "Now I am a repeat CEO," he wrote. "I say 'I don't know', fairly often." He states the alternative bluntly: "So is bullsh*tting your team."
He also chose a mentor during these years. Mark Leslie, the founding CEO of Veritas and later a lecturer at the Stanford Graduate School of Business, was "foundational in my thinking about leadership, transparency, and communication." Twice a year Feldman helped Leslie teach a GSB class, and the students always asked the same question: how do you learn to lead? He gave the same answer every time: "Leading is a muscle like any other muscle. You learn by repetition."
Five Hundred and Twelve Small Things
SeaMicro began as a thread, and Feldman pulled it.
Around 2007 he noticed that large data-center operators had become anxious about power: how much their servers drew, how much heat they produced, how much it cost to cool rooms full of machines doing what was mostly simple work. "I saw the growing importance of power in the data center," he told Electronic Design, "and I began to explore the cause of the power issue." The research process he describes sounds less like engineering than like fieldwork, which is what a failed development economist would know how to do. "I started having breakfasts and lunches with different people in my network." Through those meals he was introduced to Gary Lauterbach, a server architect with ideas about building a very different, much lower-power machine. "After talking to 30 or 40 customers we came to believe that we were onto something very interesting."
The product they built in 2007 inverted the server industry's logic. The industry's flagship processor was Intel's Xeon. SeaMicro skipped it and used the Atom, a small, low-powered chip best known for netbooks, and packed 512 of them into a box 17.5 inches tall. The company claimed that box did the work of 40 conventional Xeon servers with one-quarter of the power and one-quarter of the space. In effect it was a swarm, intelligence through multiplicity.
Feldman raised money both before and after the 2008 financial crisis, which is a harder trick than it sounds. In 2009 he pitched Steve Vassallo of Foundation Capital. Vassallo passed on SeaMicro. "But I knew I wanted to invest in Andrew," he wrote later, "whatever he'd do next."
In February 2012, AMD bought SeaMicro. Feldman has put the price at $334 million. Various conference bios say $355 million or $357 million, a twenty-three-million-dollar drift in the retelling that nobody seems inclined to settle. He joined AMD as corporate vice president and general manager of its server business and stayed until June 2014. The AMD executive who had wanted his company was Lisa Su. Feldman would later rank her, with Jensen Huang and Hock Tan of Broadcom, among "the three great CEOs over the last 10 or 15 years." He also tends to introduce her, almost casually, as the person "who bought my last company."
The AMD period produced a document worth reading closely in light of what came next. On May 15, 2013, Feldman spoke at J.P. Morgan's technology conference and made the case for ARM processors in servers. The reporter covering the talk summarized his thesis as a bet "on the server industry's history of picking smaller, more mass-producible options over larger and more complex offerings." The future of the data center, in Feldman's 2013 telling, was "lots of little power efficient cores to solve simple problems." He brushed off the tribal loyalties of the instruction-set wars with one line: "This isn't religion, this is instruction sets."
Two years later he would co-found a company dedicated to the largest and most complex single piece of silicon ever made.
This looks like a reversal, and it is not one, and the reason it is not one is the key to understanding him. SeaMicro and Cerebras answered the same question, which is what a particular workload actually needs, and that question has no permanent answer. In 2007 the workload was search queries and web serving: many small, independent, cheap tasks. Tiny chips suited it. In 2015 the workload was neural networks, which needed to move enormous quantities of data between memory and compute over and over, so the cost of communication between chips became the binding constraint. A large chip suited that. His loyalty was to fit. In his own phrase, it isn't religion.
Two Lines on a Whiteboard
When Feldman left AMD in mid-2014, Vassallo did what he had promised himself he would do. "When SeaMicro got acquired by AMD, we set a proverbial egg timer, knowing he'd pop out at some point." The two of them then "spent the next two years riffing on a half dozen ideas until he converged on building systems for AI workloads." Feldman remembers the venue: "We were doing some of that thinking together in your office, actually."
He described the way a computer architect frames an opportunity as a series of nested questions: "We see the rise of a new workload and ask: can we build a better machine that's really good at that work? And if it's sufficiently better at that work, should we make the processor happen? Is the market going to be big enough?" Then came the question that decided everything: why should anyone expect a graphics processing unit, "a machine that had been built to push pixels to a monitor," to be good at AI? Vassallo had a line for the answer that Feldman would have liked to write himself: "just because a piano can float doesn't make it a good life raft."
In 2015, Feldman got the SeaMicro band back together. Gary Lauterbach, the CTO who had come to him over breakfasts in 2007, returned. So did Sean Lie, Jean-Philippe Fricker, and Michael James. Feldman calls them "some of the leading computer architects and engineers in the industry." At the first meeting, he recalls, "we wrote two things on a whiteboard – that we wanted to work together, and that we wanted to build something that would transform the industry and be in the Computer History Museum."
Look at what the whiteboard leaves out. It mentions no market size, no product, no revenue target, and not even AI. The first line is about the people and the second is about permanence. A museum is a place for things that are finished, things whose meaning has been settled. Five men starting a company wrote down, as one of their two founding objectives, that they wanted their work to be history.
The company was formally founded in 2015 (some accounts say 2016), and Feldman has served as CEO since April 2016. In May 2016 Cerebras raised a $27 million Series A led by Benchmark, Foundation Capital, and Eclipse Ventures. The Benchmark partner was Eric Vishria, who would later say that he invested despite having no chip experience and that he almost didn't take the meeting. Eclipse brought Feldman closer to Pierre Lamond, a semiconductor pioneer whose career ran back through Fairchild Semiconductor and Gordon Moore, and who once tried to hire Morris Chang straight out of college, years before Chang founded TSMC. Feldman credits Lamond with teaching him "to set the bar far higher than I could have imagined and challenged me to think about what a great team could really accomplish."
In early 2016 there was also the meeting with Sam Altman and Ilya Sutskever. Feldman describes the moment as two sketches meeting: "OpenAI was an idea and Cerebras was a PowerPoint." AI then meant classifying cats and dogs, "finding a cat in a picture and making sure it was not a chair." The OpenAI founders were talking about artificial general intelligence, superintelligence, safety. "You're looking at them saying this is crazy talk," Feldman recalled. Elsewhere he gave his more private reaction: "I remember leaving that meeting thinking… holy sh*t." Altman, Sutskever, and many of their colleagues invested in Cerebras's early rounds. "What Ilya said was true," Feldman said a decade later. "It happened."
He is honest about the limits of his own foresight in a way most founders aren't. "We clearly didn't think it would be this big," he said, "or we would have raised at a higher valuation."
By 2017 Cerebras had about thirty employees, and it needed TSMC, the world's dominant contract chipmaker, to alter its manufacturing process for them. Normal chips are printed as a grid of separate dies with the lines between them left blank for cutting. Cerebras needed the dies connected across those lines so that the entire wafer could act as one processor. A question in a 2026 interview with Feldman puts it in a single clause: "Why did TSMC agree to modify their manufacturing process for a 30-person startup in 2017?" The question explains why the whiteboard's first line came first. The answer, insofar as there is one, involves who the thirty people were.
Still Can't Make It
Feldman explains the core idea of the wafer-scale engine with a near-insolent simplicity that he clearly enjoys. "Build really big chips," he said at a company event. "Except nobody's ever done it before."
The physics run like this. A chip computes quickly, but it has to fetch its data from memory, and the trip costs time and power. GPUs use high-capacity, relatively slow memory sitting off the processor and connected by a narrow channel. Feldman's analogy, which he has repeated so often it has worn smooth, is this: to generate a single word from a 70-billion-parameter language model, "which is not very big," you have to move "about 100 full-length movies' worth of data." If that data has to squeeze through "a thin little pipe to the GPU," you wait. The alternative is SRAM, memory that is extremely fast but historically small. On a postage-stamp chip you cannot fit much of it. On a dinner plate you can "stuff it to the gills." Put the SRAM next to the compute cores on the same silicon, and, in Feldman's arithmetic, you move "more than 2,600 times more data more quickly." He calls memory bandwidth "a known Achilles' heel of GPU architectures."
So the logic was clean, and the manufacturing nearly killed them.
Cerebras had to design cooling methods to keep a massive chip from burning itself up under power. It had to write software that routed around the microscopic defects every wafer contains, which meant building in redundancy so the chip could tolerate flaws instead of dying from them. It had to invent a machine that could drive forty screws into the assembly simultaneously without cracking the silicon. The hardest problem was the one Feldman later named in technical terms: "the coefficient of thermal expansion mismatch between the chips and the interposer." In plain language, the silicon and the board beneath it grow by different amounts when they heat up, and a structure that large will shear itself apart.
— Andrew Feldman, on the Startup Project podcast, 2026It wasn't right the first time or the second time or the eighth time. In fact, we had about a 15-month period where we were spending eight million a month, and we couldn't make one. You're going to a board meeting every six or eight weeks saying, "Nope, still can't make it."
In another telling he flattens it further: "You have board meetings every six weeks. All you've got to say is, still can't make it. Right? That's the board meeting. I mean, what else are we going to talk about?"
The money still came in. A Series B led by Coatue closed in December 2016, and a Series C led by VY Capital in January 2017. In November 2018, partway through the desert, Cerebras raised $88 million in a Series D that made it a unicorn while it could not yet manufacture its product. Altimeter joined. Benchmark, Foundation, Eclipse, Coatue, and VY all put in more. It is one of the odder facts of the story: the company's valuation crossed a billion dollars during the period its CEO was reporting failure at every meeting.
Feldman later developed a theory that explains why this was possible, and it reads like the sort of thing a person works out on long drives home from bad board meetings. Startups reduce two kinds of risk, engineering risk and market risk. The work of reducing engineering risk is continuous, "a smooth line through hard work day by day." Investors only pay for it in steps, "when you hit a milestone. Daily work is hard to understand. Milestones are easy to evaluate." The worst place to be is "just below the riser in the step function," where you have done nearly all the work and can prove none of it. Cerebras sat beneath that riser for well over a year. What carried it was investors who, according to one account of Feldman's reflections, understood that the payoff for solving this particular problem was "a monopoly on speed," along with failure analysis rigorous enough that each attempt failed slightly better than the last.
The other thing that carried it was the group. In a 2026 post on company culture, Feldman described the discipline of "fearless engineering" partly as humility: "Having the humility to back out of the 'engineering cul-de-sac' and try again." A cul-de-sac is a nice image for fifteen months of attempts. You drive in confidently and find no exit, and then you have to reverse all the way back out, in front of everybody.
He has never claimed the outcome was inevitable. At Stanford in 2026, a student asked how he stayed committed to one company for more than a decade. His answer, as an audience member recorded it, was this: "I'm not necessarily right. If this company had failed, it would have been one of the most highly funded failures in history. We were fortunate that it worked."
Then came July or August of 2019 and the box with the flashing lights. In all, Feldman estimates, the first working system took "about three years and, when all was said and done, about half a billion dollars."
Nobody Cared
On August 19, 2019, the New York Times ran Cade Metz's story under the headline "To Power A.I., Start-Up Creates a Giant Computer Chip." Cerebras announced the WSE-1 and the CS-1, a 19-inch rack-mounted appliance built around the chip. The first government customer was Argonne National Laboratory, and the first commercial customer was GlaxoSmithKline. AstraZeneca followed, along with Lawrence Livermore, the National Energy Technology Laboratory, the Pittsburgh Supercomputing Center, and the Edinburgh Parallel Computing Centre. In November 2019 the company raised $270 million at a $2.4 billion valuation.
Feldman's verdict on that period is harsh and he repeats it with relish. "In 2019, Cerebras solved the hardest problem in the computer industry," he wrote in 2026, "and nobody cared." On another podcast he put it more vividly: "Everybody said we were crazy, it would never work." When it did work, "nobody bought it, and nobody cared." Or: "for years, no one gave a sh*t."
That is true in the sense he means and not quite true in the literal sense, and the gap between the two is revealing. People did buy the machine, but they were the wrong people for the scale he had in mind. The customers were national laboratories and pharmaceutical researchers working on drug discovery, computational fluid dynamics, genomics, and, starting in 2020, COVID-19. In October 2022, Sandia National Laboratories began using the CS-2 for nuclear stockpile stewardship, which means running simulations to determine whether America's nuclear weapons will work as intended. In November 2022, Argonne researchers who had used Cerebras hardware alongside Nvidia and HPE systems won the Gordon Bell Special Prize for work modeling SARS-CoV-2 variants with large language models. The National Energy Technology Laboratory found the CS-2 nearly 500 times faster than its Joule supercomputer at field equation modeling. A 2024 simulation with Sandia, Livermore, and Los Alamos tracked 800,000 atoms in one-femtosecond increments, each step computed in microseconds, far faster than on Frontier, then the flagship of American supercomputing.
It was serious, important, glamour-free science. It did not create a market. In 2016 AI meant cats in pictures, and in 2020 it meant protein folding and weapons physics and genomics. These were real problems, but they belonged to the few institutions with the budgets to explore them, and the CS-2 cost up to several million dollars a node. The engineering was a decade ahead and the demand was not.
Cerebras kept building through those years. The CS-2 came out in April 2021 with 850,000 cores and 2.6 trillion transistors on TSMC's 7-nanometer process. A $250 million Series F in November 2021, led by Alpha Wave Ventures and the Abu Dhabi Growth Fund, valued the company above $4 billion and brought total financing to $720 million. The Abu Dhabi name in that round would matter later. In November 2022 the company unveiled Andromeda, sixteen wafers clustered into 13.5 million cores and one exaflop of AI compute, drawing 500 kilowatts. In March 2024 it launched the WSE-3, built on a 5-nanometer process, with 4 trillion transistors and 900,000 cores, twice the performance of its predecessor, and Time named the CS-3 one of the best inventions of the year.
One event in those years had nothing to do with revenue and probably meant more to the founders than any of it. In August 2022, the Computer History Museum in Mountain View opened a display featuring the WSE-2 under the title "The Biggest Chip In the World." The company's literature says the museum honored it as an "epochal" achievement. The second line on the whiteboard was done seven years after it was written, while the business still had not found its market.
"We were honored," Feldman said, with the understatement of a man ticking off a box he had drawn himself.
Too Busy to Hate
The market, when it came, came first from the desert in a geographical sense.
In July 2023, G42, the Abu Dhabi technology group that Feldman describes as "a sovereign institution and a national champion for the UAE," agreed to pay around $100 million for the first of potentially nine Cerebras supercomputers. Condor Galaxy 1 delivered 4 exaflops across 54 million cores. Condor Galaxy 2 was announced that November. Ground was broken on Condor Galaxy 3, rated at 8 exaflops, in March 2024. The first of these machines was built in the United States, not the Gulf. Feldman said Cerebras "recently received licenses to deliver equipment in the UAE." In August 2023, Cerebras, G42's subsidiary Inception, and the Mohamed bin Zayed University of Artificial Intelligence released Jais, a large language model in Arabic.
The dependence that resulted was extreme. According to Cerebras's own registration statement, MBZUAI accounted for 62 percent of the company's 2025 revenue and G42 for another 24 percent. Eighty-six cents of every dollar came from two related institutions in one emirate. G42 also held 3.5 million shares. Customer concentration of that kind is usually the line in an S-1 that makes investors wince, and it was. It is also the line that kept the company alive while American hyperscalers were still buying GPUs by the warehouse.
Feldman talks about the relationship in terms that are partly geopolitical and partly something that sounds like personal conviction. Asked in 2026 for his most contrarian belief, he did not mention chips:
— Andrew Feldman, on the Startup Project podcast, 2026I think we'll have peace in the Middle East sooner than people think. The returns to being moderate have been demonstrated by the UAE. In 2005, the GDP of Dubai was the same as Gaza.
"In those regions," he went on, "people are too busy to hate right now; they're busy working and building cool things." He then gave the clearest example he could think of, from his own inbox: "I got an email from a Hindu manager reminding me, a Jewish guy, to wish a Muslim team member Happy Eid." That happens in Silicon Valley, he said, "because we're all working together to build stuff. We don't care. The only question is, can they build something cool?"
You can read that skeptically. A CEO whose company depends on Abu Dhabi for most of its revenue has every commercial reason to speak warmly about the Gulf. You can also hear in it the young man who chose development economics, who went to the Northwest Frontier Province to study how poor places become less poor, and who still reasons in terms of "returns to being moderate," a phrase only an economist would use about peace. The two readings don't cancel each other. The Pakistan detour is not a quirk in his biography. It is the lens he uses to understand his biggest customer.
He keeps a sense of humor about all this. In early 2025 he gave a keynote at Web Summit in Doha and found he was the warm-up act. "My talk was followed by Will Smith," he wrote, "marking the first time I've opened for a celebrity. I offered to open for him in Vegas as well but he politely declined." He added a hashtag: #aibutnotinvegas.
The Speed of a Word
The turn came from a direction Feldman admits he saw late.
From the founding through 2023, Cerebras was a training company. Training is how AI models are made. Inference is how they get used, the generation of an answer one token at a time each time someone types a question. "Until ChatGPT, nobody was doing large-scale inference in production, so all of us were doing training," Feldman explained. Then usage exploded, along three dimensions he likes to enumerate: more users, using it more often, for more complicated tasks. "Each of those increases the compute needed, which is a product of three rapidly growing dimensions."
This was precisely the workload the wafer was built for, because inference is where the hundred movies have to cross the thin pipe for every single word. In August 2024, Cerebras launched an inference service that it said was ten to twenty times faster than systems built on Nvidia's H100. Feldman is unusually candid about the timing: "I was probably too slow in recognizing that and adding resources to build out our inference program. I wish I had done that six or eight months earlier." He also draws the methodological lesson: "There is no substitute for working with customers and building things for learning and for product strategy. It's very hard to do it in a conference room."
For a while, being fast was a novelty. Between 2023 and 2025, by one account of his reflections, Cerebras was "blisteringly fast," and the market didn't much care because AI was not yet doing daily work. "Until about early 2024, AI was mostly a novelty," he said. "It was cool, but it wasn't doing real work." Then it began doing real work, and from then on speed was a requirement. His analogy is Netflix. "Netflix used to deliver DVDs and envelopes, and when the Internet got fast, they became a movie studio." Speed doesn't just do the old thing faster. It creates businesses that could not exist before.
The customers arrived. In January 2025, Cerebras ran DeepSeek's R1 70B reasoning model at 1,600 tokens per second. In February, Mistral moved its Le Chat service onto Cerebras and claimed a speed record, and Perplexity's Sonar model began running there at 1,200 tokens per second. In March, Cerebras announced six new data centers in Dallas, Minneapolis, Oklahoma City, Montreal, New York, and France. In April, Meta chose Cerebras to power its Llama API. In May, independent benchmarking had Cerebras running Meta's 400-billion-parameter Llama 4 Maverick at more than 2,500 tokens per second per user against roughly 1,000 for Nvidia's Blackwell.
Feldman made one more decision in this period that tells you a lot about him. At the end of 2024, he scrapped multi-year contracts. "My investors thought I had lost my mind," he wrote. Customers had been telling him, "We can't even see six months out," and Cerebras was asking them for multi-year commitments. "That just didn't seem right." The company moved to short contracts and pay-as-you-go pricing, rebuilt its sales motion from scratch, and accepted that it would forgo money from customers who would have overcommitted. When customers such as AlphaSense and Cognition "scaled like crazy, we could scale with them."
Every startup knows the fundraising strategist's temptation, booking the long contract now so it looks good on a slide later. A man who once pleaded guilty to circumventing accounting controls declined that temptation and described the decision publicly as "one of the hardest calls." The record makes no further connection, and none is needed.
The software moat that was supposed to protect Nvidia turned out to matter less than everyone assumed. "When you do inference, you don't use CUDA," Feldman said. "The rise of inference weakens CUDA." Switching from a GPU cloud to Cerebras took, in his telling, "ten keystrokes." In another interview it took eight. You change one word in your API call and pick a model. "That's it."
In January 2026 came the deal that ended the "nobody cared" era for good. OpenAI agreed to buy 750 megawatts of Cerebras computing power through 2028. CNBC put the value at more than $10 billion. Feldman has since called it "$20B+." It was negotiated, he said, in four and a half weeks. OpenAI received warrants to purchase up to 33.4 million Cerebras shares, tied to its compute purchases. As part of the agreement, Cerebras was temporarily barred from selling to Anthropic. The idea that had been a PowerPoint in early 2016, sitting across a table from another idea, was now that idea's supplier.
Two months later Amazon Web Services agreed to buy CS-3 systems. Then AMD announced a "historic partnership," with AMD's Helios hardware handling the prefill stage of inference and the Cerebras wafer handling decode. "Proud to partner with Lisa Su and the AMD team again," Feldman wrote. An industry analyst left a comment for anyone with a short memory: Lisa had once bought Andrew's previous company, and Andrew had once worked for Lisa.
Ties
The IPO itself took a detour. Cerebras first filed its S-1 on September 30, 2024. A year later, in early October 2025, a week after closing a $1.1 billion round at an $8.1 billion valuation, the company withdrew the filing without saying why. Rumors filled the gap, most of them about G42 and national-security review. Feldman responded on LinkedIn: "We didn't explain why – that was a mistake." The data was stale, he said, and "the numbers were too small." It was "a procedural step, not a strategic shift." In January 2026 the company raised another $1.0 billion at $89.01 a share. Benchmark, whose partner had almost skipped the first meeting, put in $225 million of it.
On May 14, 2026, Cerebras listed on Nasdaq under the ticker CBRS. It sold 30 million shares at $185, raised $5.55 billion, and opened at $350. Shares rose about 68 percent on the first day. The company's valuation that day depends on which source you trust and which minute of trading you choose. Wikipedia says $95 billion and Bloomberg roughly $67 billion. Podcasters in the following weeks said $63 billion, $56 billion, and $60 billion, sometimes in the same episode. Feldman's 10.1 million shares were worth an estimated $3.2 billion by Bloomberg's count. Sean Lie, who had followed him from SeaMicro, held a stake worth around $1.6 billion. Feldman spoke of minting a thousand millionaires.
The share prices along the way trace the step function he had written about: $14.66 in 2024, $36.23 in September 2025, $89.01 in January 2026, $185 at the offering, $350 at the open. In a 2026 post he gave the three jobs of a CEO as interesting work, amazing colleagues, and the third one, "Do everything in your power to make them (and their families) rich." For a decade that sentence had been aspirational. Now it could be audited.
Feldman's favorite memory of the day has nothing to do with the stock. When the company brought employees who had been there nine years or more to the listing, along with their families, he found out that engineers do in fact own ties. He also noticed the weight of it for the children of immigrants, whose parents had waited a decade for this.
Public life arrived fast, as it does. In June 2026, after the first earnings report, Feldman was on CNBC defending a gross-margin outlook. After a later report the stock fell, and he said the margin forecast had been "misunderstood." In September, regulatory filings recorded him selling 37,441 shares. Time put him on its list of the 100 most influential people in AI. He began talking about a $25 billion backlog. He also began writing, often, about the malaise that hits companies as they grow from 1,000 to 3,000 people. "Meetings replace bold ideas. Asses get covered, rather than cool stuff built." He insisted, the way people insist on things they are afraid of, that "we refuse to let it happen here."
In the spring of 2026 he gave two talks at Stanford, the campus where he had been raised and from which he had tried to leave for Pakistan, for a doctorate, for a professorship. One was to the IEEE club in the Paul Allen Building. The other was to graduate students in computational and mathematical engineering. He noted where the second talk took place: "In an irony lost on nobody, my talk was in the Jensen Huang building." He did not dwell on it. "In the spaces created by the success of others, I shared my experiences."
Meanwhile, a few miles north on Shoreline Boulevard in Mountain View, the second line on the whiteboard sits behind glass. It is a wafer about eight and a half inches on a side with 850,000 cores and 2.6 trillion transistors. It has no power running to it and no lights blinking. It is a few minutes' drive down San Antonio Road from the address where Shockley's lab opened in 1956, and its placard reads: The Biggest Chip In the World.
Part IIThe Playbook
The principles below are drawn from three decades of Feldman's decisions across five companies, from the YAGO-era networking startup through SeaMicro to Cerebras. They are not general management advice. They come from a specific operating situation: building capital-intensive infrastructure against incumbents with nearly unlimited resources, over time horizons long enough to break most founders, using a team whose continuity is itself the competitive advantage. Some of them contradict the standard software playbook on purpose.
The principles
- 1.Fall in love with the plumbing.
- 2.Bet on the workload, not against the incumbent.
- 3.Against Goliath, ten percent is not a strategy.
- 4.Keep the band together for twenty-five years.
- 5.Sell the risk you've retired, not the work you've done.
- 6.Bring the bad news before the board meeting.
- 7.Say "I don't know" out loud.
- 8.Sell the system, not the part.
- 9.Price to the customer's horizon, not yours.
- 10.Let usage tell you what you already built.
- 11.Recruit with impossibility, retain with ownership.
- 12.Treat the crisis like Tuesday before lunch.
Principle 1
Fall in love with the plumbing
Feldman turned down his first networking job twice because he thought the work was "plumbing." Thirty years later his whole career is plumbing: switches, routers, microservers, and now the silicon underneath every fast token. The lesson isn't that infrastructure is secretly glamorous. Infrastructure compounds in ways applications don't, and the people suited to it are often not the ones who picked it.
What matters is the self-knowledge that came after the accident. "We weren't going to build a web app; that's not who we are," he said of the 2015 founding. "Entrepreneurship, like many things in life, pays dividends if you stay true to who you are." He thinks the same logic favors young founders in consumer software, where "they and their friends look like their customers," and works against them in chips, where "there are a lot of returns to experience." He applies the rule to himself. He would never have invented Facebook. "I've been married 30 years; that's not a question I ask."
Tactic: Before choosing a market, write down which layer of the stack you would still find interesting after ten years of failure, and build there, even if the layer looks like plumbing from the outside.
Principle 2
Bet on the workload, not against the incumbent
Feldman's theory of market entry is historical, not heroic. Dominant designs entrench themselves. In x86, Intel and AMD traded share for decades with "no meaningful entrants." When a new workload appears, the incumbents' advantages stop applying. His favorite example is mobile: "When the cell phone workload emerged, who was better positioned to take advantage of that than AMD and Intel? Both failed completely, zero share." ARM, Apple, and Samsung won instead.
This explains what looks like the great reversal of his career, from championing many small chips at SeaMicro and AMD to building the largest single chip in history at Cerebras. The architecture changed because the workload changed. Web serving rewarded swarms of cheap cores. Neural networks rewarded keeping enormous quantities of data on-chip. Feldman's loyalty was never to a design. It was to the question of fit. "This isn't religion," he said in 2013, "this is instruction sets."
⇄
Same Founder, Opposite Architectures
SeaMicro and Cerebras look like contradictions. They are the same question asked of different workloads.
| Dimension | SeaMicro (2007–2012) | Cerebras (2015–) |
|---|---|---|
| Workload | Internet tasks: search, web serving | Neural network training, then inference |
| Binding constraint | Power and space in the data center | Memory bandwidth and data movement |
| Architecture | 512 tiny Atom chips in one box | One wafer-sized chip, SRAM next to compute |
| Claimed advantage | Work of 40 Xeon servers at 1/4 power and space | 15–30x faster than GPUs on inference |
| Outcome | Sold to AMD for ~$334M | Public at a valuation in the tens of billions |
Tactic: Identify the workload whose demands are growing fastest and ask which incumbent assumption it breaks. Build for that break, not for a slightly better version of what the incumbent already sells.
Principle 3
Against Goliath, ten percent is not a strategy
Feldman calls himself a "professional David." His rule for the fight is arithmetic. Being 10 percent better, or even twice as good, "is not an available strategy" against a company with Nvidia's resources, as he told Jack Altman and Eric Vishria. You have to aim for 100x. The 2015 thesis was "not a little bit better or a little bit cheaper, but orders of magnitude faster." In his view, building a chip that resembles the competitor's gives you roughly zero chance of beating it.
He pairs this with real respect for the giants. He calls Jensen Huang, Lisa Su, and Hock Tan "dazzling." But "their size also creates opportunity for us. They can't move as quickly as we can. They can't take the type of engineering risks we take." The asymmetry is in risk tolerance, not resources. A large company cannot spend fifteen months reporting "still can't make it." A startup with the right investors can.
The defensibility follows from the difficulty. Asked why Nvidia hasn't built a wafer-scale chip, Feldman says it would take "five or seven years and two or three billion dollars, and we would still be 10 years ahead." He adds that Cerebras solved in 2018 the thermal-expansion problem he says delayed Nvidia's B200 by eighteen months.
Tactic: If your advantage over the incumbent can be stated as a percentage under 100, either find a different architecture or a different fight.
Principle 4
Keep the band together for twenty-five years
The first line on the Cerebras whiteboard was "we wanted to work together." It came before any product, any market, or any mention of AI. Eight engineers from Feldman's first startup in 1998 still work with him. All four of his Cerebras co-founders came from SeaMicro. Gary Lauterbach, whom he met over the breakfasts and lunches that led to SeaMicro, has been his CTO across two companies.
Feldman treats this continuity as a financing asset, not just a cultural one. "There are very few great hardware teams—maybe eight or 10 in the world—and we have one of them. They've been with me for 20-plus years, which made it easier for us to raise money for a big idea." Investors who could not evaluate wafer-scale packaging could evaluate a team that had already shipped together. That is very likely part of why TSMC modified its process for a thirty-person startup.
The team is also how the company got through the desert. People who have already survived one company together can survive fifteen months of failed attempts together.
Tactic: Treat your best collaborators as a portfolio you carry across companies. Stay in touch between ventures, and when you start the next one, recruit the team before you finalize the idea.
Principle 5
Sell the risk you've retired, not the work you've done
Feldman says he has raised more than $4 billion across 17 financing rounds over 25 years, and he reduces all of it to one graph. Startups reduce engineering risk and market risk. The work is continuous and the rewards come in steps. "Daily work is hard to understand. Milestones are easy to evaluate. A working demo. A first customer. A signed contract." The danger zone is "just below the riser," where the work is nearly done and nothing can be demonstrated.
Cerebras's share price history traces that graph exactly.
↗
The Step Function, Priced
Each riser corresponds to a demonstrable reduction in market risk.
2016
$27M Series A led by Benchmark, Foundation Capital, and Eclipse. The company is a PowerPoint.
2018
$88M Series D makes Cerebras a unicorn, while it still cannot manufacture a working chip.
2019
First working system in July or August. $270M Series E at $2.4B follows in November.
2024
Series F-1 shares sell at $14.66. Inference service launches in August.
2025
Series G at $36.23 per share, an $8.1B valuation, after Meta, Mistral, and Perplexity sign on.
2026
Series H at $89.01 after the OpenAI deal. IPO prices at $185 and opens at $350.
His practical advice follows from the graph: raise enough to "comfortably get through the next major demonstrable milestone," frame every use of funds as risk reduction, and "when you complete a fundraise, I recommend immediately writing the slide deck you want to use for the next raise."
Tactic: Write your next fundraising deck the week you close the current round, then work backward to the milestones that would make each slide undeniable.
Principle 6
Bring the bad news before the board meeting
Feldman's first rule for boards is absolute: "Never surprise your board with bad news. Ever. If something breaks, do not wait for the board meeting. Call each board member individually." You can surprise them with good news. "Trust grows when there are no surprises. It evaporates when there are."
That rule is the reason the Cerebras desert didn't end the company. A CEO who walks into meeting after meeting with "still can't make it" survives only if the board has never been ambushed, because the report then reads as candor instead of collapse. His other two rules support the first. Never be defensive, and you don't have to take the board's advice, but you do have to show that you thought it through. "It is unusual for a board member to know your business better than you do."
The S-1 withdrawal in 2025 was the exception that proved the rule. Cerebras pulled the filing without explanation, and the vacuum filled with speculation. Feldman's correction was public and unhedged: "We didn't explain why – that was a mistake."
Tactic: When something material breaks, phone each director individually within 48 hours with the problem and your plan, before the next scheduled meeting gives anyone the chance to discover it themselves.
Principle 7
Say 'I don't know' out loud
As an entrepreneur-in-residence in 2007, Feldman watched what he estimates were hundreds of pitches and noticed that "the more senior and the more credible the CEO, the more likely they were to answer with 'I don't know.'" The admission, followed by a plan to find out, calmed the room. Bluffing put people on edge.
He has made this into an operating principle and extended it to failure. "When I mess up, I say I messed up. And I tell the team why." His definition of transparency is narrower and more demanding than most: "Sharing what it is easy to share is not transparency. Everyone can do that. True transparency is sharing the sausage making." The goal is to remove "the illusion that good decisions spring like Zeus's children fully formed from the CEO's head."
For a non-technical CEO of a deep-tech company with more than 130 patents across lithography, packaging, and machine learning, this is not a matter of style. It is survival. "Bullsh*tting engineers is the surest way to lose their respect." He applies the same honesty to his own record: he says he was slow to the inference opportunity, that he priced the first rounds too low, and that the company could have failed.
Tactic: In your next all-hands, name one decision you got wrong this quarter, explain the reasoning that led you there, and describe what you changed.
Principle 8
Sell the system, not the part
Cerebras does not sell chips. It sells a computer, "the chip in a system," a complete rack delivered and installed, or it sells cycles on that system through its own cloud and through marketplaces at Amazon, Microsoft, Hugging Face, and OpenRouter. Feldman's reasoning is historical and blunt: "It's very hard to get paid for software when you're selling chips," and selling a chip on a PCI card leaves you "dependent on Dell or Supermicro for your I/O and power."
The system also protected the invention. A wafer-scale chip is useless without the cooling, packaging, power delivery, and defect-routing software around it, which were the very problems that consumed $200 million. Selling the whole assembly means the customer never has to understand any of it. The cloud layer takes this a step further by removing the hardware decision altogether. The customer changes a word in an API call. "It's literally 10 keystrokes."
Tactic: If your core innovation only works because of the surrounding engineering, sell the surrounding engineering as part of the product, and make adoption a configuration change instead of a procurement decision.
Principle 9
Price to the customer's horizon, not yours
At the end of 2024, Feldman dropped multi-year contracts, the structure "so beloved in the SaaS world," in favor of short terms and pay-as-you-go pricing. His investors "thought I had lost my mind." The reasoning came straight from customers: "We can't even see six months out." Asking them to sign multi-year commitments "just didn't seem right."
The trade was explicit. Cerebras gave up guaranteed revenue from customers who would have overbought and kept the ability to grow alongside those who exploded, like AlphaSense and Cognition. In a market where usage was compounding on three axes at once, locking in today's volume meant capping tomorrow's. As Feldman put it, "AI is not SaaS."
Tactic: When your customers' planning horizon is shorter than your contract length, shorten the contract. Bet on usage growth over booked revenue, and make it easy for them to say yes now.
Principle 10
Let usage tell you what you already built
Cerebras built a training machine and discovered it had also built the best inference machine on the market. The architecture's main advantage, memory bandwidth, matters most in exactly the phase of AI that didn't exist at scale until ChatGPT. Feldman admits the recognition was late: "I wish I had done that six or eight months earlier."
His conclusion is that strategy comes from contact, not contemplation. "There is no substitute for working with customers and building things for learning and for product strategy. It's very hard to do it in a conference room." The same habit started SeaMicro: thirty or forty customer conversations before anyone quit a day job. It also drove the AMD partnership, which split inference into prefill and decode and gave each half to the hardware best suited to it.
Tactic: Every quarter, list the three ways customers use your product that you didn't design for, and staff the fastest-growing one before your roadmap tells you to.
Principle 11
Recruit with impossibility, retain with ownership
Feldman's account of the CEO's job has three parts: give people interesting work, surround them with great colleagues, and "do everything in your power to make them (and their families) rich." He rejects the military and sports models of leadership. "In Silicon Valley, nobody has to salute. And everyone can leave tomorrow."
The first part, interesting work, is where the impossible problem pays for itself. "Nobody became an engineer to build ordinary things." A company whose founding goal is a museum display attracts engineers who want that too, along with mentors like Pierre Lamond and Mark Leslie. "Big, interesting projects have a gravitational pull." On compensation he is unapologetic: "No company ever went bankrupt paying extraordinary people too much. You go if you wanna go bankrupt, pay mediocre people too much."
The danger he watches for most closely after the IPO is settling. "With hundreds of openings, there is tremendous pressure to put a body in a seat and move on. One settle becomes two."
Tactic: Hire against the hardest problem you have, pay at the top of the market for the people who can solve it, and never fill a seat just because it is empty.
Principle 12
Treat the crisis like Tuesday before lunch
Feldman's wife worked in an emergency room for 25 years, and he credits her with two lessons that shape how he runs a company. The first is perspective: "In business, our crises aren't that serious." Panic burns energy and solves nothing. "You can get kicked in the gut 8 times before lunchtime and have it still be a good day."
The second is about asymmetry. A patient's ER visit is a day they remember forever. For the doctor, "It was Tuesday before lunch." Inside a company, the same gap separates a CEO's twentieth decision of the day from the employee who spent a week preparing the pitch. The right response is not a quick no but an explanation of why not now, and of what would have to change to make it yes.
Both lessons treat calm as an advantage, and both treat the other person's experience as information. A company that spent fifteen months unable to build its product needed a leader who could treat each failed attempt as ordinary and still take seriously the engineer who had staked everything on it.
Tactic: When you turn down an idea, spend two extra minutes explaining the reasoning and the conditions under which you would revisit it.
Free playbook
Get The Business Model Playbook
58 business models, one visual page each: how the money flows, the metrics that matter, and who runs it. Free when you join the Faster Than Normal email.
Free. No spam. Unsubscribe anytime.
Part IIIQuotes and Maxims
In their words
— Andrew Feldman, LinkedIn, 2026We would rather fail in pursuit of the extraordinary than succeed in the ordinary.
— Andrew Feldman, on whether AI compute will be commoditizedAs you know less about something, the details and complexity go away. From the moon, the Earth is just a blue and white orb. Get up close, and there are religious, political, and technology battles.
— Andrew Feldman, LinkedInWhen you visit the emergency room it's a day you will remember your entire life… For my wife, who looked after you, it was like every other day in the ER. It was utterly unmemorable. It was Tuesday before lunch.
— Andrew Feldman, LinkedInThe reticle limit was the line. Cerebras engineers are the only ones to have stepped across it.
— Andrew Feldman, LinkedInI call it the myth of LinkedIn - this idea that we all had our careers mapped from the time we left high school. Mine wasn't a straight path. It was a series of accidents where I fell into things I enjoyed.
Maxims
- The workload picks the architecture. Loyalty to a design is religion. Loyalty to the workload is engineering.
- Ten percent is a rounding error to Goliath. If you can't be orders of magnitude better, find a different fight.
- Hire the band before you write the song. A team that has already shipped together is the one asset an investor can't diligence away.
- Milestones get paid, and work doesn't until it reaches one. Raise enough to reach the next thing you can demonstrate, not the next thing you can describe.
- Bad news belongs on the phone. Never let a board meeting be where your directors first hear it.
- "I don't know" calms a room. Bluffing loses engineers faster than anything else.
- Sell the box, not the chip. The surrounding engineering is the product, and the easiest switch is a few keystrokes.
- Contract to the customer's horizon. When buyers can't see six months ahead, a three-year deal is a fiction.
- Strategy comes from customers, not conference rooms. What you built may be better at something you didn't design it for.
- Write the museum on the whiteboard. Aim for work that will outlast the company, and let the market catch up when it does.
