Cerebras founder Andrew Feldman says his AI chips are 15 to 20 times faster than GPUs at inference, a speed edge he ties to new AI business models.
Feldman says Cerebras spent about $8m a month from mid-2017 to mid-2019 unable to build a working wafer-scale chip before it finally succeeded.
Feldman says going public gives Cerebras access to a broader investor base and lower cost of capital, while accepting strict governance rules, after a roughly ten-year path.
Cerebras's giant chip bet pays off as AI speed matters
Cerebras bet on a radically different wafer-scale chip years before the market wanted it. It now sells inference it says is 15 to 20 times faster than GPUs. Demand arrived when AI became useful for daily work in 2025 and speed became essential. The result is a backlog above $20bn, an OpenAI deal and a public listing near $60bn.
Takeaways
Cerebras claims inference 15 to 20 times faster. Feldman says this holds against GPUs across big and small models, US and Chinese alike. Speed became essential because models turned useful around 2025 and people began using them daily in their work.
Fast AI should create new business models. Feldman compares it to faster internet enabling Netflix. He argues speed will do more than replace existing tools such as coding and SaaS. Demand ramped at Cursor and Cognition, whose growth he calls unprecedented.
New workloads let architects beat incumbents. Feldman notes Intel and AMD gained no share when graphics and mobile arrived. Radical improvement needs a different design, so Cerebras built a dinner-plate wafer chip where rivals use postage-stamp chips. Gene Amdahl failed at the same goal.
The OpenAI deal closed at unusual speed. Feldman spoke with Sam in mid-summer 2025, trials showed Cerebras far faster, a term sheet came at Thanksgiving and a master agreement on December 24. A deal above $20bn took about four and a half weeks.
Going public cut capital cost and added credibility. An IPO swaps professional venture investors for a broader group, lowering the cost of capital, in return for strict governance. Cerebras pitched itself as the only AI pure play, and the path took ten years.
AI coding tools are transforming internal engineering. Engineering token spend rose from about $1,000 eight months earlier to roughly $25,000 to $30,000. With about 850 employees, Feldman worries that growth to thousands will erode the fearless engineering culture he calls essential.
Founders should decide in advance when to stop. Feldman calls the CEO job extraordinarily lonely and says he thrives as David against NVIDIA's Goliath. Outside former CEOs can hold founders accountable against the slippery slope of endless testing.
From failed chip to $20bn backlog
From mid 2017 to mid 2019 Cerebras could not build a working wafer-scale chip while spending about $8m a month.
In summer 2019 the chip yielded and began to work. Feldman says the company then spent two or three years ahead of a market that did not care.
The first generation sold about a dozen systems, the second about 300 and the third tens of thousands.
Tap a name to go there · tap a line to see why they are linked
On this episode 3
Show 1
Supercomputing customers such as Argonne and Sandia came first, followed by oil and gas users.
G42 placed a billion dollar order that let Cerebras change its supply chain and test at scale.
The backlog now exceeds $20bn, and manufacturing is set to grow tenfold this year, which Feldman calls as fast as anyone in hardware history.