Andrew Feldman on Building Cerebras and the Future of Chips

⬅️ Back to Videos

🎥 Andrew Feldman on Building Cerebras and the Future of Chips

Jack Altman with Andrew Feldman (Cerebras) and Eric Vishria (Benchmark). Duration: 1 hr 3 min

Timestamps

  • 0:00 Intro
  • 4:31 Attacking Goliath takes 10 to 500x
  • 7:45 Vishria’s 100x math
  • 11:43 Only new mistakes
  • 12:20 The July 2019 breakthrough
  • 16:05 From sand to chatbot answer
  • 21:55 Three decades of bad fab policy
  • 23:00 Being a good hardware board member
  • 29:15 Hop to hop, not telephone circuits
  • 42:30 Everyone got the exponential wrong except Sam
  • 48:15 Data centers move at the speed of real estate
  • 53:00 No market for slow search
  • 55:45 What actually makes Nvidia great

Feldman started Cerebras in 2016, before transformers existed, to build a chip the size of a dinner plate. A decade later it is a public company selling speed. This conversation is the clearest account I have seen of what that decade cost.

  1. Attacking an incumbent with a slightly better or cheaper product is not a strategy. Feldman says Nvidia had the margins to undercut, bundle, or give product away against any modest challenger. You need something 10, 100, or 500 times faster, a proposition that wins even if the other side charges zero.
  2. Vishria puts numbers on it. The incumbent improves about 2x a year, and reaching scale takes a new company five years even if everything works first try. That is 32x before you start, plus a 3x edge on top, so roughly 100x. Most pitches he hears promise 2x against today, which he calls aiming at the wrong target.
  3. Radical bets force you to own the whole stack. Nobody sells a heat sink for a dinner plate chip, so Cerebras built chip, board, system, software, and API itself. Slower and pricier, Feldman admits, but it made them world class at packaging through years of failure. You cannot inch your way to vastly better.
  4. For 18 months they could not make wafer scale work while burning about 8 million dollars a month. Board meetings every six weeks had one agenda item, in Feldman’s words: still can’t make it. The mantra was only new mistakes, full failure analysis every time. Wafers went from shattering in seconds to surviving minutes to an hour.
  5. The breakthrough came in July 2019. A run held with flat temperature while the team stared at a server in a converted office with a hole drilled in the wall for air. Feldman calls it one of the great minutes of his life, solving a problem nobody had cracked in 75 years of computing.
  6. His sand to chatbot walkthrough is worth the episode alone. A cutting edge fab costs 40 to 50 billion dollars and lives four or five years. ASML holds a true technology monopoly on lithography machines. Out the other end of all that genius comes a 22 dollar chip, Feldman notes, cheaper than a Mission District burrito. And the chip is only a fourth of the problem.
  7. Good hardware investors stay out of the engineering. Vishria is blunt about his small circle of competence and says his job was letting engineers engineer and keeping them financed. Feldman adds that boards should ask long game questions about hiring and specialization, since the first chip is rarely good. Even Google needed four tries at the TPU.
  8. Cerebras navigated hop to hop, like internet routers, not along a fixed telephone circuit. Each summit revealed the next move, which is how a company founded for training ended up in inference. One early call made that possible. They accelerated the algebra under all AI instead of hard coding convolutional networks, so transformers ran fastest on their chip before transformers had been invented.
  9. Nearly everyone underestimated the exponential. Feldman says he got it wrong, TSMC got it wrong, Nvidia got it wrong, and only Sam Altman read the curve without flinching. The once laughable 7 trillion Stargate figure now looks small, and AI capex runs near 3.5 percent of GDP, past railroads at their peak.
  10. The binding constraint moved to concrete and copper. AI advances at the speed of software while data centers move at the speed of real estate, about 18 months for a 50 megawatt block. Feldman calls the US grid pathetic, says AI may be what brings back nuclear, and dismisses the water panic. American data centers use several times less water than California almond growers.
  11. Speed creates markets, slowness kills them. There is no market for dial up or slow search, Feldman argues, and fast inference births products nobody planned, the way broadband turned Netflix from envelopes into a studio. Cerebras pairs that with disaggregation work alongside AMD and AWS for 5x throughput. And his read on Nvidia stung me a little. Not CUDA, not architecture, but a decade of public market failure from 2003 to 2013 that burned grit into the company. He wakes up, he says, as a professional David.

Related TMFNK Content

Crepi il lupo! 🐺