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

Key points

• **The OpenAI $20B+ deal closed in 24 days.** Sam called mid-summer 2025: 'We've been trying so hard just to keep up with demand. We now see the importance of fast inference.' Trial → testing → 'whoa, we understand now' from OpenAI engineers → term sheet **Thanksgiving eve** → master agreement **Dec 24** = ~4.5 weeks for a $20B+ contract. **Feldman's lesson: 'The art of the possible has been expanded by this push in a way I'd never have expected. The rate at which Elon builds data centres, the speed Cursor grew — these were truncated aspirations. You can do deals like this in 24 days if you work on it 8-10 hours a day.'** Direct cross-reference to [Brad Gerstner on Altimeter calling Cerebras the most important AI-infra IPO since Snowflake](/issues/2026-05-17).
• **15-20x faster than GPUs across the board.** Feldman's headline performance claim — not specific to a use case, but 'big models, small models, US models, Chinese models, trillion-parameter, billion-parameter.' **Architecture decision: 46,000 sq mm chip ('size of a dinner plate'), vs everyone else's 'postage stamp'.** 'You can't get 15-20x better than the GPU with a minor modification to their architecture. If you aspire to radical improvement, your design has to be different.' Cross-references the [Issue 04 Reiner Pope thesis on memory bottleneck](/issues/2026-05-10) — Cerebras's wafer-scale design is the structural answer to memory bandwidth that Pope was identifying.
• **The 7-year wilderness: 2017-mid-2019 couldn't build the chip.** $8M/month burn rate, board meetings every 6 weeks saying 'I can't build it.' Each failure analysis got incrementally better. **Summer 2019: it yielded.** 'We were staring at a computer, which is about as exciting as watching paint dry, and it was working. We couldn't speak for half an hour.' Then **2-3 more years of being ahead of the market** (Gen 1 = ~12 units sold; Gen 2 = ~300 units; now Gen 3 = tens of thousands). **'We had a 2-3 year period where we were blisteringly fast and absolutely nobody cared.'** Direct rhyme with [Shiv Rao's 5-year-wilderness Abridge story from Issue 05](/issues/2026-05-17) — vertical AI companies needed the inflection that came in 2025.
• **G42 as 'the bridge' — the most undervalued strategic decision in the Cerebras story.** Sovereign customer placed **$1B order**. 'Allowed us to transform the company, change our supply chain, deploy in big enough clusters that we could battle-test at scale. One of the challenges in hardware: your QA lab can't be as big as some of the customers you want to deploy to. They worked with us. We trained models for them, did inference for them.' By the time OpenAI/AWS came calling, **they had the capacity, they'd battle-tested, they'd built the bridge over the chasm.** Same structural insight as the [Stargate-UAE thread from Issue 04](/issues/2026-05-10) — sovereign capital is the bridge from supercomputing-pilot scale to hyperscaler scale.
• **Why IPO: 'graduate from corporate adolescence to corporate adulthood' + the pure-play AI bet.** Feldman: *'Going public is exchanging professional investors who specialise in tech for a different class of investors — my dad. In return for that you accept extraordinarily stringent rules.'* But: **'We could offer the public market something unique — the first and only AI pure-play for a period of time. 100% of revenue from this exact market. No gaming, no graphics, no PC.'** Implicit moat: institutional investors who want clean AI-infra exposure now have *one* listed option and Nvidia. **Worth pairing with Eric Ries's [Issue 05 governance frame](/issues/2026-05-17) — Cerebras chose adolescence-to-adulthood, Anthropic chose the LTBT structural-moat path. Both are durable; the bet matters.**
• **Coding tools internal disclosure: $0 → $25-30K/engineer/month in 8 months.** 'It's not useful for everybody — some people have the perfect mindset for it, running 8-10 agents 7x24, governing agents, building QA agents, remedying coding-model weaknesses (verbosity, stripped comments). They've gone from 10x guys to 100x guys.' **'The rest of us are limping along.'** Cross-reference to [Krishna Rao's '90%+ of Anthropic code is written by Claude Code' disclosure from Issue 05](/issues/2026-05-17) — Cerebras is the second public operator confirmation of the 10x→100x discontinuity for the right cognitive type.
• **'I'm a professional David — 5th startup, competing against Goliath.'** 'Every dollar, every million, every billion we sell — if not for our brains, their muscle would have taken it in a heartbeat. You've got to love being a David.' Companion to: 'we would much rather fail in pursuit of the extraordinary than succeed in the ordinary' as the cultural-DNA frame for hiring 800 → several thousand. **Direct echo of Dana White's 'I don't have a Plan B' from [Issue 05](/issues/2026-05-17) — different industries, same founder operating mode.**
• **'Slow inference market = zero.' The killer-feature framing.** 'How big is the market for slow search? Zero. How big is the market for dial-up internet? Zero. That's how big the market for slow inference will be — but we had to wait until AI was smart enough to be useful. That happened in 2025.' **Strategic implication for the entire AI-infra cohort:** inference speed (token/sec, not just total throughput) becomes the dominant performance metric going into 2026-27 as agentic + multi-sensory use cases scale. Implicit threat to inference-cost-optimised competitors who trade latency for $/token.
• **'Netflix-becomes-a-studio' frame for fast-AI 2.0.** 'Speed doesn't make existing business models a little better. Netflix delivered DVDs and thought competition was Blockbuster — when the internet got fast, they became a movie studio. Right now we're replacing things everyone can see (coding, design, SaaS tools). Once we fundamentally reorganise around this, you'll see new business models and fundamental jumps in productivity.' **This is the framing for the next 12-18 months that wasn't in any of the synthesis threads to date.** The agentic-GTM playbook from [Legora in Issue 05](/issues/2026-05-17) is the early version; the structural reorganisation is the next leg.
• **On when to give up: 'It's clearly the right time when you've laid out a set of hypotheses about what it'll take to win and they all come back negative.'** The slippery-slope trap: 'people test one more thing, then one more — the slope is a beast. Having other former CEOs or seasoned entrepreneurs who can remind you of pre-committed exit conditions is critical.' **The frog-in-warm-water analogy + accountability framework is the most-quotable operational discipline on the tape this week.** Companion to Eric Ries's '20% founder retention 3 years post-IPO' — knowing when to keep going is the founder's hardest skill.

Notable quotes

Themes

• Cerebras IPO at $63B as the only pure-play public AI-infra stock
• OpenAI $20B+ deal closed in 24 days — 'art of the possible expanded'
• Wafer-scale architecture as the radical-improvement bet vs GPU-derivative competitors
• 7-year wilderness + G42 sovereign bridge → battle-tested supply chain
• Netflix-becomes-a-studio: fast-AI 2.0 unlocks new business models, not just better existing ones