No PriorsThe Story Behind Cerebras’ $63 Billion IPO with Founder and CEO Andrew Feldman
CHAPTERS
- 0:00 – 0:37
Cold open: Why speed changes entire business models
Andrew frames fast AI as a qualitative shift—not just making existing tasks cheaper, but enabling entirely new products and organizational workflows. He analogizes to Netflix evolving from DVD delivery to becoming a studio once bandwidth made streaming viable.
- •Speed enables new categories, not incremental improvements
- •Fast AI will move beyond replacing visible tasks (coding/design)
- •Expect reorganization of work and new productivity leaps
- •Netflix streaming as an example of speed-driven business reinvention
- 0:37 – 2:22
Cerebras today: AI computers built for ultra-fast inference
Elad introduces Cerebras and its public-market scale, then Andrew explains what the company builds and why inference speed suddenly became decisive. He describes demand exploding as models became useful in everyday workflows.
- •Cerebras builds purpose-built AI computers
- •Claims 15–20x faster inference than GPUs across model sizes/types
- •Shift in 2025: models became useful → speed became mandatory
- •Massive demand surge as inference moved into daily work
- 2:22 – 2:55
From “nobody cared” to whirlwind growth: OpenAI deal and AWS deployment
Andrew outlines the major commercial catalysts: a landmark OpenAI agreement and subsequent AWS partnership. He characterizes the period as chasing supply to meet demand.
- •OpenAI deal described as north of $20B
- •AWS agreement to deploy in AWS data centers
- •Operational reality: demand outstripped supply
- •Rapid market pull once inference speed mattered
- 2:55 – 5:11
Contrarian architecture: The wafer-scale bet and why it had to be different
The hosts probe why Cerebras looked “weird” and why radical gains require radical design changes. Andrew explains wafer-scale computing: a dinner-plate-sized chip versus postage-stamp chips, and how this enabled step-function performance.
- •Radical performance requires a non-derivative architecture
- •Wafer-scale chip (~46,000 mm²) vs conventional small dies
- •Skepticism was intense: many said it was impossible
- •2019 marked proof and delivery of wafer-scale viability
- 5:11 – 7:09
Early AI era context: From cats vs. chairs to foundation models
Elad and Andrew reflect on how early ML workloads differed from today’s foundation-model world. This chapter situates Cerebras’ early conviction in a period when “AI” still felt fringe and use cases were limited.
- •2016-era ML: CNNs, RNNs, early GANs
- •Use cases were narrow (basic image classification)
- •Magnitude of progress highlights why timing mattered
- •Cerebras built ahead of mainstream demand
- 7:09 – 8:32
Two years of near-failure: solving wafer-scale manufacturing and yield
Andrew recounts the hardest technical stretch: years of failed attempts, massive burn, and repeated boardroom uncertainty. The breakthrough came in mid-2019 when they achieved yield and saw the system work end-to-end.
- •2017–2019: couldn’t build it despite ~$8M/month spend
- •Wafer-scale had a long history of failed attempts (incl. Gene Amdahl)
- •Iterative failure analysis gradually improved outcomes
- •Emotional moment when the first working system finally ran
- 8:32 – 9:21
Being ahead of the market: selling a handful, then hundreds, then tens of thousands
After technical success, Cerebras faced market indifference—speed didn’t matter when AI wasn’t yet embedded in daily work. Andrew explains why the “market for slow inference” will go to zero once usage becomes habitual.
- •Post-breakthrough: initial sales were tiny; market didn’t value speed
- •AI was still a novelty through ~2023–early 2025
- •Once used daily, latency becomes unacceptable (search/web analogy)
- •2025 inflection: widespread adoption drives demand explosion
- 9:21 – 11:09
Crossing the chasm with early adopters: supercomputing, industry, and G42 as the bridge
Andrew details the classic hardware go-to-market path: supercomputing and specialized industries first, then a difficult leap to mainstream volume. A billion-dollar order and partnership with G42 provided scale, real-world testing, and supply-chain transformation.
- •Initial beachhead: national labs and supercomputing sites
- •Next: oil & gas and pharma—compute-heavy early adopters
- •The “chasm” problem: early markets don’t provide enough volume
- •G42 partnership and ~$1B order enabled scale QA and readiness
- 11:09 – 12:34
Scaling constraints in hardware (and software): manufacturing 10x and the 10-year compiler lesson
The conversation turns to operational scaling: hardware growth requires supply-chain and factory realities that can’t be spun up overnight. Andrew also emphasizes that a mature compiler/toolchain takes years, correcting early optimism.
- •Hardware scaling requires physical expansion: lines, buildings, power, fixtures
- •Goal: 10x manufacturing increase—rare in hardware history
- •Software stack maturity is also a scaling bottleneck
- •Compiler development took ~10 years, not 5
- 12:34 – 14:09
AI-generated coding inside Cerebras: from token spend to ‘agent governors’
Sarah asks how AI coding tools affect Cerebras internally. Andrew describes rapid adoption and a split: some engineers become dramatically more productive by orchestrating multiple agents, while others are still learning best practices.
- •Token spend ramped quickly (from near-zero to tens of thousands)
- •Some engineers run many agents continuously and focus on QA/governance
- •Coding models have quirks (verbosity, missing comments) that teams adapt to
- •Productivity gains vary widely by mindset and role
- 14:09 – 15:37
Leadership at scale: protecting a ‘fearless engineering’ culture and hiring bar
Andrew discusses the risks as organizations grow from hundreds to thousands: risk-aversion, settling in hiring, and losing the appetite for bold bets. He argues culture must prioritize extraordinary outcomes—even at the cost of failure.
- •Growth can cause a shift from bold bets to incrementalism
- •“Fearless engineering culture” as a core competitive advantage
- •Hiring discipline matters; ‘butt in a seat’ is described as fatal
- •CEO time spent heavily on recruiting to protect the bar
- 15:37 – 17:48
Founder psychology: loneliness, loving the journey, and competing as ‘professional David’
In advice to founders, Andrew highlights the emotional reality of leadership and the need to enjoy building, not just outcomes. He frames Cerebras’ story as repeatedly taking on giants and winning through ingenuity rather than scale.
- •CEO/leader role is inherently lonely and difficult
- •Motivation must come from the craft and mission, not money
- •Competing against Nvidia requires embracing the ‘David vs. Goliath’ dynamic
- •Winning is framed as brains overcoming incumbents’ muscle
- 17:48 – 20:03
Quit vs. persist: hypothesis-driven decision making and avoiding the slippery slope
Elad asks how to decide when to abandon an effort. Andrew argues for pre-committed hypotheses and checkpoints, plus outside accountability to prevent rationalizing continued failure through incremental “one more test” thinking.
- •Quit when core hypotheses repeatedly test negative
- •Beware sequential rationalizations (“just one more test”)
- •Use experienced advisors/peers to hold you accountable to prior criteria
- •Define what must change and a timeframe to reassess
- 20:03 – 23:28
Why Cerebras went public: cost of capital, credibility, liquidity, and ‘corporate adulthood’
Andrew explains the trade-offs of going public versus staying private in an era where a few AI companies can raise massive private rounds. He emphasizes reduced cost of capital, legitimacy with large customers, employee liquidity, and becoming an AI ‘pure play’ for public investors.
- •IPO swaps specialized VCs for broader public investors; lowers capital costs
- •Public-company governance adds stringent requirements
- •Secondary liquidity matters when timelines stretch to a decade
- •Cerebras positioned as a rare AI pure-play; IPO as ‘adulthood’ transition
- 23:28 – 26:25
The OpenAI deal: how fast inference became strategic and the market learned to move at warp speed
Andrew recounts first conversations with Sam Altman, trials that showed a large speed gap, and an exceptionally rapid contracting process. The hosts generalize this to a broader market trait: ambitious operators compress timelines once thought impossible.
- •Mid-2025: OpenAI recognizes fast inference as crucial
- •Trials/testing validated Cerebras’ speed advantage
- •Term sheet to master agreement executed in weeks for a $20B+ deal
- •Broader lesson: industry timelines are shrinking (M&A, data centers, deals)
- 26:25 – 30:33
Open source, post-training workloads, and what fast inference will unlock next
Andrew credits open source with keeping the ecosystem vibrant and pushing frontier players, including through innovations from global competitors. He ends by returning to the speed thesis: fast AI will enable new business models and reorganize work, not merely automate existing tasks.
- •Open source sustained adoption when closed models were too expensive
- •Ecosystem competition accelerates innovation and prevents complacency
- •Infrastructure builders should enjoy others’ ideas flourishing on their systems
- •Fast inference will unlock new products and workflow reorganization (Netflix analogy)
