Be 100x Better or Don't Bother

Andrew Feldman is the co-founder and CEO of Cerebras, the company he started in 2016 with four longtime colleagues to build a chip the size of a dinner plate for AI. Cerebras is his fifth startup: he has sold three companies, including one to AMD, taken one public, and has now taken Cerebras public. He grew up on the Stanford campus as the child of faculty, with William Shockley next door, and left a PhD for Stanford's business school before getting pulled into startups. Cerebras now serves OpenAI and a hyperscaler alongside government and sovereign cloud customers. These signals are from Andrew Feldman's interview with Jack Altman on Uncapped with Jack Altman.

1. The 100x Rule. Feldman says a startup attacking a Goliath with high margins can't win by being a little better or a little cheaper, because the incumbent can cut prices, bundle, or give parts away. Benchmark's Eric, a Cerebras board member since the start, gave the math. If the incumbent doubles performance every year and a startup needs 5 years to reach scale, the bar is already 2 to the fifth power, or 32x, and a real edge on top of that puts the target near 100x. "You can't incremental your way to vastly better," Feldman says, so the only route is radical innovation.
2. Own Every Layer. To get that advantage under its own control, Cerebras built the chip, board, system, software, and everything up to the API. That made the company slower, costlier, and gave it more ways to fail. Radical designs have no vendors waiting: no catalog sells a heat sink for a chip the size of a dinner plate when everything before it was the size of a postage stamp. Feldman says the payoff is expertise that compounds: Cerebras didn't start out as a packaging leader, and now calls itself the best in the world at it, earned "failure after failure year after year."
3. Only New Mistakes. For about 18 months Cerebras couldn't make wafer-scale work while spending roughly $8 million a month. Board meetings every 6 weeks came down to "still can't make it." The team ran a full failure analysis after each attempt under the mantra "only new mistakes," moving from wafers that shattered in seconds to minutes to an hour. In July 2019, in a small office converted into a lab with a hole drilled in the wall for exhaust, the temperature held flat. Feldman calls it one of the great minutes of his life: solving a problem nobody in 75 years of computing had solved.
4. Hop-to-Hop Strategy. Feldman contrasts two ways to plan: an old telephone circuit that sets a fixed line to the destination, and an internet router that decides at each hop. Cerebras knew there was "a pot of gold" out there and treated the path as unknowable, so each milestone earned a new view and a new decision. That is how a company founded as a training system ended up focused on inference once it saw that smarter AI would mean everyone using it. "Earn the next viewpoint," he says.
5. Bet on the Math. Hardware companies can't change course overnight, so a few big early decisions carry years of consequence. In its first architecture Cerebras chose not to hard-wire support for convolutional networks, which dominated in 2016, and instead accelerated the linear algebra underneath all of AI. When transformers arrived, Cerebras was the fastest at them too, despite never having seen them. "In chips we measure three times and cut once," Feldman says.
6. Speed Creates Markets. Feldman argues speed has historically created markets: nobody pays for slow search or dial-up. When the internet was slow Netflix mailed DVDs, and when it got fast Netflix became a movie studio. His suggested punishment for a teenager is a week at dial-up speed, and he asks why people accept slow AI. Fast frontier intelligence, he says, is already leading customers to dream up entirely new applications.
7. Disaggregated Inference. Cerebras is also chasing throughput by splitting inference work between its chips and other vendors' hardware. With AMD, Feldman says the split delivers 5x more throughput at the same speed, and AWS is showing similar numbers. Its architecture lets it do this across the GPU field: of the 4 major chipmakers (Nvidia, AMD, Google's TPU, and AWS), Cerebras already works with two and would like to partner with Nvidia.
8. Nvidia's Decade of Grit. Feldman thinks most people misread what makes Nvidia great. He doesn't credit CUDA or the chip architecture. From roughly 2004 to 2014 the stock traded badly and the company fought for every sale while nobody listened, and that intensity got into its DNA before it became the most valuable company in the world. He points to Jensen Huang still fighting at events and says no earlier tech leader at that scale fought that way.
9. Speed of Real Estate. "AI is moving at the speed of software and data centers are moving at the speed of real estate," Feldman says, and that gap explains the shortage. A leading-edge fab costs $40 to $50 billion, takes about 5 years to build, and only TSMC has the skills, so it can't build 12 at once; ASML is the sole maker of the lithography machines. A 50-megawatt data center block takes about 18 months from raw land with a good builder, and generators and electrical switchgear are the long-lead items. In 5 years, he predicts, the industry will still be chasing data centers, chips, and HBM.
10. Only Sam Got It Right. Feldman says everyone underestimated the AI buildout, including Cerebras, TSMC, Nvidia, and the memory makers. The exception was Sam Altman, who "saw an exponential and wasn't afraid" while others kept predicting a slowdown. People laughed at the early Stargate numbers, and Feldman now calls the initial Stargate "sadly small." Eric cited a chart putting AI capex at about 3.5% of GDP a year, against 2% for railroads and 1% for highways and telecom.
11. Pay Your Own Way. Feldman says the industry hurt itself early by pushing costs onto local communities and trying to take advantage of municipalities. A data center should pay its own way and needs little water with closed-loop systems: California almond growers use 4 to 7 times more water than every US data center combined. The industry failed to explain the thousands of construction jobs and the ongoing tax base, and is now paying for it. He also wants a 20-year waiver of local ordinances so TSMC, Samsung, and GlobalFoundries can build US fabs, after three decades of bad policy pushed fabs, tool vendors, and packaging firms such as Amkor and ASE offshore.
12. Experience Where It Counts. Cerebras staffs product and go-to-market with young people promoted from within and no tenure rules, because no one has deep experience in what AI customers want. Chip, system, and mechanical work is different: years of building chips best predict great chips, and few people get to tape one out even in a PhD. Feldman says five founders is too many unless they've worked together before, and co-founder Sean, hired at 25 or 26 as an individual contributor at his last company, became an AMD corporate fellow and is now CTO of a public company. Extraordinary people show it fast, he says: within 6 to 8 weeks, and often in the first email, where every list is in descending order of importance.

Watch the full video at https://www.youtube.com/watch?v=HFOhG1jH2tk.