TSMC actually bet on a new startup with "0 products, 0 revenue"! Andrew Feldman, co-founder and CEO of Cerebras, an American AI chip company, personally revealed a past cooperation nine years ago on the social platform X. In 2017, Cerebras had only 30 people and flew to Taipei to seek cooperation with a giant chip idea that had repeatedly failed in the past. He bluntly said that TSMC had "1,000 reasons to say no" at the time, but unexpectedly, it directly approved the plan at the meeting. This decision also allowed him to see the two keys to TSMC's ability to "consistently win."
Federman said Silicon Valley often talks about "long-term partners," but there are few better examples than TSMC.
Back in 2017, Federman and his team flew to Taipei, hoping that TSMC would be willing to "bet on us." However, at that time, Cerebras had only 30 employees, the product had not yet been released, and there was no revenue. The entire company was even smaller than the service team in a TSMC restaurant.
What’s even more surprising is that what Cerebras wants TSMC to manufacture is not an ordinary chip. Federman described TSMC's wafer fabs as "the pyramids of our generation," with investments of up to US$40 billion to US$50 billion each. The entire manufacturing process was originally built around postage stamp-sized chips, but the idea proposed by Cerebras is a giant chip that is 55 times larger than any previous chip.
This idea is not a path that has been successfully followed. Federman pointed out that in the past 70 years of computer development, there have been attempts, but every time they failed.
A start-up company with only 30 people, 0 products, and 0 revenue requires the world's top chip manufacturer to accompany it to challenge a technology that has repeatedly failed before. Federman bluntly said: "TSMC has 1,000 reasons to say no."
At that time, TSMC was already an industry giant and was the best in the world in manufacturing completely different types of chips. Federman even asked: "Why take this risk?"
But TSMC’s final decision was exactly the opposite. Federman revealed that TSMC approved the plan at that meeting and has since become an important partner of Cerebras.
Looking back nine years later, Federman believes that this decision just explained why TSMC can continue to win. He pointed out two key points. The first is to "remain open to new ideas"; the second is "to be willing to cooperate regardless of the size of the company."
Today, the 30-person startup that had zero products and zero revenue has continued to invest in the AI chip market. According to recent public information, Cerebras launched a new generation of AI server hardware CS-4 this year, which is equipped with wafer-level giant chips and continues to use TSMC’s 5nm process. It is expected to be launched in the third quarter of this year.
Reuters pointed out that Cerebras produces AI hardware and chips that compete with Huida, focusing on "inference" computing in the field of AI, which is the process of making chat robots such as Anthropic's Claude generate responses.
