For years, the pitch sounded like a joke: a startup that would build a chip the size of an entire silicon wafer, taking on Nvidia. On Thursday, Cerebras Systems closed its first day on the Nasdaq up 68%.
Cerebras, which went public in the largest U.S. tech IPO since Uber’s 2019 debut, nearly ran out of money along the way. The company burned roughly $8 million a month in its early years, spending toward about $800 million in cumulative capital before it had a commercially viable product, according to a reconstruction published this week by TechCrunch. The 10-year journey from investor mockery to a Nasdaq listing ended with Cerebras raising $5.55 billion and a valuation near $95 billion.
Cerebras priced its initial public offering on May 13 at $185 a share, 16% above the marketed range, selling 30 million shares. Shares opened at $350 on Thursday and closed at $311.07, valuing the company at about $95 billion. If underwriters exercise their option to buy 4.5 million additional shares, total proceeds could reach $6.38 billion.
Cerebras was founded in 2016 on a contrarian premise: instead of cutting a silicon wafer into hundreds of separate dies and wiring them back together — the way every GPU maker builds chips — the company would leave the wafer intact and build one chip the size of an entire wafer. The idea had been tried and abandoned for decades, dating to attempts in the 1980s, and the semiconductor industry’s engineering consensus said it could not be done profitably: yields would be too low, power would be unmanageable, and no one had ever solved the problem of communicating across a 46,225-square-millimeter slab of silicon.
The early years nearly sank the company. Investors who heard the pitch were often openly skeptical, according to TechCrunch’s account. A startup claiming it could out-engineer Nvidia by building chips bigger than anyone had ever attempted was a hard sell in a market built on scaling down.
The company solved the engineering problems one by one: manufacturing the world’s largest processor, keeping a full-wafer chip cool, delivering power to it, and achieving commercially acceptable yields. Its Wafer-Scale Engine 3, now in production on a 5-nanometer TSMC process, packs 4 trillion transistors and 900,000 compute cores on a single continuous piece of silicon — 58 times larger than Nvidia’s B200 chip, with 2,625 times more memory bandwidth than Nvidia’s B200 package. Cerebras argues the architecture eliminates the off-chip data movement that dominates GPU latency and power consumption, making its systems faster and more power-efficient on AI workloads.
The IPO itself was a long and winding road. Cerebras first filed to go public in September 2024 but withdrew a little over a year later after its prospectus drew heavy scrutiny over reliance on a single customer, UAE-backed G42, which accounted for 85% of revenue in 2024. The company refiled in April and diversified its business in the intervening months. In the refreshed prospectus, G42’s share of last year’s revenue had fallen to 24%, though the Mohamed bin Zayed University of Artificial Intelligence in the UAE accounted for 62% of revenue.
Cerebras also shifted its strategy away from selling hardware systems toward offering a cloud service based on its chips, putting it in direct competition with cloud providers including Google, Microsoft, Oracle, and CoreWeave. The logic: a wafer-scale chip is worth more rented by the hour than sold by the unit, and a cloud business produces recurring revenue that investors can model.
The diversification arrived with two marquee deals. In January, OpenAI signed a cloud deal with Cerebras worth more than $20 billion, running through 2028. In March, Amazon Web Services said it would install Cerebras chips in its data centers so developers can run AI models on them, giving the company a second route to customers. Both Amazon and OpenAI hold warrants to purchase Cerebras stock.
The market backdrop helped. Cerebras is the biggest pure-play AI IPO to hit Wall Street, and the first notable tech offering in months. The VanEck Semiconductor ETF has jumped 58% so far in 2026, with Intel, AMD, and Micron all notching triple-digit gains this year. Nvidia, Cerebras’s most formidable competitor and the world’s most valuable company, paid $20 billion in December for assets from Groq, whose chips more closely resemble Cerebras’s architecture — a sign that even the incumbent takes the wafer-scale threat seriously.
Cerebras’s path from near-bankruptcy to a $95 billion valuation is the rare case where a company bet everything on a technology the industry said was impossible, and won. The $8 million-a-month burn in its early years — against roughly $800 million in cumulative capital before a viable product existed — was the price of that bet, and it paid off at the largest U.S. tech IPO since Uber. The harder test now begins: Cerebras must convert its architectural advantage into recurring revenue against Nvidia, whose ecosystem and market position remain dominant. The IPO proved the skeptics wrong about the chip. The coming quarters will test whether the company can prove them wrong about the business.
