The Story Behind Cerebras' $63 Billion IPO with Founder and CEO Andrew Feldman

2026-05-21 - 30 min - source - Read full transcript

Sarah Guo (host)Elad Gil (host)Andrew Feldman

Key insights

Cerebras was fast for years before anyone cared, because speed only matters once a technology is used every day.

Feldman says the company had already proven wafer-scale chips worked by 2019 and was already blisteringly fast years before demand arrived - from roughly 2023 to early 2025, people 'pointed at AI' without using it daily. Once daily use began in 2025, speed became non-negotiable, comparable to the zero market size for slow search or dial-up internet, and demand exploded overnight.

fast-inference-as-new-category

Feldman argues a radical performance jump requires a genuinely different architecture, not an incremental variant of the incumbent design.

He frames this as a general rule: you cannot get 15-20x better than a GPU with a minor modification to GPU architecture. Cerebras chose wafer-scale (a single 46,000 square millimeter chip) specifically because it could not be a derivative of existing designs, a bet critics called impossible and 'wrong' before it worked.

wafer-scale-contrarian-bet

New computer workloads have historically produced entirely new market winners rather than being captured by incumbents.

Feldman notes that when graphics emerged as a workload, Nvidia (not Intel or AMD) won; when mobile compute emerged, ARM won. He treats AI as the same pattern: a new workload requiring a new, dedicated architecture, which is why Cerebras bet the whole company on being '100% contrarian' from the start.

wafer-scale-contrarian-bet

Cerebras spent roughly two years (2017-2019) unable to build its core product, burning about $8 million a month with no working prototype.

Feldman describes recurring board meetings every six weeks admitting 'I can't build it, it's still not working,' each time doing a failure analysis to improve slightly. The wafer-scale chip finally worked in the summer of 2019, in a makeshift Los Altos office - a moment he says left the team unable to speak for half an hour.

wafer-scale-contrarian-bet

Sovereign and government-lab customers were the bridge that let Cerebras survive the years it was technologically ahead of market demand.

Feldman's sequencing: national labs first (Argonne, Lawrence Livermore, Sandia, LRZ) because that world tolerates immature software and values raw speed; then oil and gas and pharma; then G42, a sovereign investor whose billion-dollar order let Cerebras transform its supply chain and battle-test equipment at a scale its own QA lab could never afford, which in turn made it ready when OpenAI and AWS came calling.

wafer-scale-contrarian-bet

Feldman's rule for knowing when to quit a venture is when every specific hypothesis required to win has come back negative, not a vague gut feeling.

He warns against the 'slippery slope' of always testing 'one more thing,' and says outside CEOs or seasoned entrepreneurs are valuable specifically because they can remind a founder of the threshold they themselves set earlier - the 'frog in warm water' problem of a standard quietly eroding.

founder-persistence-and-psychology

Feldman treats being CEO as fundamentally lonely, and says founders should love the building itself, not the money, because the path is too hard otherwise.

He calls himself 'a professional David,' on his fifth startup, competing deliberately against giants like Nvidia because there are easier ways to make money than competing with a company that strong. He credits the chip-on-the-shoulder response to being told a problem can't be solved as part of what sustains him through periods with no external validation.

founder-persistence-and-psychology

Cerebras used secondary-market sales to give employees and early investors liquidity, decoupling the IPO decision from the traditional four-year option-vesting clock.

Feldman notes the standard Silicon Valley option package timeline was built around roughly four years to liquidity via an IPO; Cerebras opened the secondary market instead so people betting large chunks of their career on the company could find modest liquidity along the way, rather than forcing an earlier public listing.

ipo-and-late-stage-liquidity

Feldman frames Cerebras' public offering as valuable specifically because it is a pure-play AI compute company with no other revenue mixed in.

He argues Cerebras could offer public markets something unique - being the first AI 'pure play' with no gaming, graphics, or PC revenue diluting the story - alongside the credibility and audited-books legitimacy that comes from being public, on top of reducing cost of capital by trading venture investors for public shareholders.

ipo-and-late-stage-liquidity

The $20-billion-plus OpenAI deal went from term sheet to signed master agreement in about four and a half weeks.

Feldman says he first spoke with Sam Altman in the middle of summer 2025 as OpenAI recognized the importance of fast inference; testing showed Cerebras was far faster than competitors, a term sheet was signed the night before Thanksgiving, and the full master agreement was signed December 24 - a pace he says required multiple law firms working seven days a week.

ai-native-development-velocity

Deal and build timelines across the AI industry are compressing well past what Feldman previously assumed was physically possible.

He cites Cognition buying Windsurf 'over a weekend' and Elon Musk's data-center build speed as evidence that things widely assumed to be at 'the speed of light' were not - they could be done much faster if someone works on them every day, eight to ten hours a day, expanding what he calls 'the art of the possible.'

ai-native-development-velocity

Cerebras' own internal AI-generated coding spend jumped roughly 25-30x in eight months, but Feldman says the productivity gain is highly uneven across employees.

He says the company went from under $1,000 per engineer per month on AI coding tokens to roughly $25,000-$30,000 within eight months. A small group with 'the perfect mindset' - running eight to ten agents around the clock, building their own QA agents, and correcting the models' verbosity and comment-stripping tendencies - went from being '10x' engineers to '100x' engineers, while most others, himself included, are still 'limping along' figuring out how to apply it to their own roles.

ai-native-development-velocity

Companies

Cerebras - Feldman's company; makes wafer-scale AI computers, recently went public at roughly a $63 billion market cap, and is the whole subject of the episode.
OpenAI - Signed a deal with Cerebras north of $20 billion at the end of 2025; the term sheet was signed the night before Thanksgiving and the master agreement four and a half weeks later, on December 24.
AWS - Signed an agreement in March to deploy Cerebras hardware in its data centers.
Nvidia - The dominant GPU incumbent Feldman positions Cerebras against; he calls himself 'a professional David' competing against this Goliath.
G42 - The Abu Dhabi sovereign investor and strategic partner that placed a billion-dollar order early on, letting Cerebras transform its supply chain and battle-test equipment at scale before the OpenAI and AWS deals.
Argonne National Laboratory / Lawrence Livermore / Sandia / LRZ - Supercomputing labs that were Cerebras' earliest customers, chosen because that world values raw speed and tolerates immature software.
Cognition - Cited as a fast-moving AI coding company that bought Windsurf 'over a weekend,' illustrating how much faster deals now move in this market; its product Devin is also cited as a magical experience on Cerebras hardware.
Cursor - Named alongside Cognition as an AI coding tool ramping extraordinarily once inference got fast enough to be useful daily.
Lovable - Named as another AI-native company that ramped quickly once inference speed crossed the daily-use threshold.
Anthropic - Cited, with OpenAI and Databricks, as one of the first companies able to raise huge sums privately without going public.
Databricks - Same point as Anthropic - an example of a company able to raise at public-market valuations while staying private.
Netflix - Feldman's central analogy for what fast AI enables: Netflix didn't just deliver DVDs more efficiently than Blockbuster, speed let it become a movie studio - a fundamentally new business.

Techniques and frameworks

Wafer-scale chip architecture - Cerebras' core bet: building a single 46,000 square millimeter chip (the size of a dinner plate) instead of GPU-sized 'postage stamp' chips, on the thesis that a truly new AI workload requires an architecture that isn't a derivative of what already exists.
Ahead-of-market patience via supercomputing beachhead - Feldman's path through a multi-year period of being technologically ahead of demand: win supercomputing labs first (they value speed and tolerate immature software), then oil and gas and pharma, then a sovereign investor (G42) whose billion-dollar order funds the supply-chain and battle-testing needed to serve hyperscale customers.
Failure-hypothesis test for when to quit - Feldman's rule for when to give up on a venture: it is clearly right to quit once you've laid out the specific hypotheses required to win and they all come back negative, as opposed to endlessly testing 'one more thing' - the slippery slope he says outside CEOs and seasoned entrepreneurs are needed to catch.
Late-stage secondary liquidity instead of an early IPO - Cerebras' approach to the option-timeline problem: rather than rushing to IPO on the traditional four-year option-vesting clock, they opened up the secondary market to let employees and early investors find modest liquidity while staying private longer.

Summary

Sarah Guo and Elad Gil host Andrew Feldman, co-founder and CEO of Cerebras, days after the company's IPO valued it at roughly $63 billion. Feldman recounts the arc from being technologically ahead of the market for years - wafer-scale chips (a single 46,000 square millimeter chip, the size of a dinner plate, versus GPU-sized "postage stamps") that critics called impossible - to being overwhelmed with demand once AI models became good enough for daily use in 2025. His central framing is that radical performance gains require architectures that are not derivatives of the incumbent design: Cerebras' bet paid off with inference speeds 15-20x faster than GPUs, and he draws the historical parallel that new computing workloads (graphics, mobile) have always produced new winners (Nvidia, ARM) rather than being captured by incumbents like Intel or AMD.

He details the brutal middle years: roughly 2017 to 2019 spent unable to build the core wafer-scale product, burning about $8 million a month with recurring board meetings admitting failure, followed by the first working chip in a makeshift Los Altos office in summer 2019 - a moment he says left the team speechless for half an hour. Even after that technical breakthrough, Cerebras spent two to three more years ahead of market demand, surviving on a sequenced path through customers who tolerate immature software and value raw speed: national supercomputing labs (Argonne, Lawrence Livermore, Sandia, LRZ), then oil and gas and pharma, then a pivotal billion-dollar order from sovereign investor G42, which let Cerebras transform its supply chain and battle-test equipment at a scale its own QA budget couldn't fund - the bridge that made it ready when OpenAI and AWS arrived.

On founder psychology, Feldman is candid that being CEO is "an extraordinarily lonely thing" and offers a specific decision rule for when to quit: it's right to give up once every hypothesis you laid out for winning has come back negative, not through the "slippery slope" of endlessly testing "one more thing" - a trap he says outside CEOs and seasoned entrepreneurs help catch by reminding founders of thresholds they set for themselves earlier. He describes himself as "a professional David" on his fifth startup, arguing founders should love the building itself because there are much easier ways to make money than competing with Nvidia.

The conversation turns to why and when Cerebras went public: Feldman contrasts it with the small handful of companies (OpenAI, Anthropic, Databricks) now able to raise huge private rounds at public-market valuations, arguing that for everyone else, an IPO still buys legitimacy, audited-books credibility, and lower cost of capital. He notes Cerebras used secondary-market sales to give employees modest liquidity along the way, decoupling the decision from the traditional four-year option-vesting clock, and frames the IPO itself as valuable partly because Cerebras is a rare AI "pure play" with no gaming, graphics, or PC revenue diluting the story. He recounts the OpenAI deal moving from a first conversation with Sam Altman in mid-2025 to a signed term sheet the night before Thanksgiving to a full master agreement by December 24 - a four-and-a-half-week sprint he says required multiple law firms working seven days a week.

Feldman closes on how fast inference reshapes what's buildable, using Netflix as his central analogy: speed didn't just make DVD delivery incrementally better than Blockbuster, it let Netflix become a movie studio - a genuinely new business. He expects the same pattern for AI: today's visible use cases (coding, design, SaaS tools) are just the obvious replacements, but the bigger productivity jump comes once work is reorganized fundamentally around AI speed, the way the cloud and SaaS reorganized work after the PC. On internal AI adoption, he says Cerebras' own AI-coding token spend jumped from under $1,000 to $25,000-$30,000 per engineer per month in eight months, but the gains are concentrated in a small group who've redesigned their workflow around running multiple agents and QA-checking their output - most employees, himself included, are still figuring out how to apply it to their own roles.

Notable Quotes

Note: the source transcript (podscripts ASR) carries no speaker diarization. Guest quotes below are attributed to Andrew Feldman based on first-person content about his own company and decisions; host questions are not separately distinguishable and are not quoted here.

"How big is the market for slow search? It's zero. How big is the market for dial-up internet? It's zero. That's how big the market for slow inference will be." - Andrew Feldman

"You got to love being a David, right? I'm a professional David - this is my fifth startup. I compete against Goliath. That is what I do for a living." - Andrew Feldman

"It is clearly the right time to give up when you've laid out a set of hypotheses about what it's going to take to win and they all come back negative." - Andrew Feldman

"Netflix used to deliver DVDs and envelopes... And when the internet got fast, they became a movie studio. That's what happens with speed." - Andrew Feldman

"Eight months ago we weren't spending $1,000 per engineer on tokens, and we're probably at 25 or 30,000 right now, and it's ripping." - Andrew Feldman