Cerebras Systems' monster debut on Thursday didn't just place it among tech's biggest-ever IPOs — it was a crystal clear signal of unstoppable demand for chips to power AI, as tech giants scramble to find alternatives to the costly, sold-out graphics processing units made by Nvidia.
Cerebras closed its first day trading on Wall Street with a market cap just below $100 billion, putting it near the few companies to close above that mark, such as Facebook-parent Meta and Alibaba. The stock closed 10% lower on Friday, its first full day of trading.
Here's what you need to know about this hot Nvidia competitor.
Cerebras makes a different type of chip than the classic Nvidia GPU, and it's the size of a dinner plate.
"We build the biggest chips in the semiconductor industry," Cerebras CEO and Co-Founder Andrew Feldman told CNBC on Squawk Box Thursday. "Big chips process more information in less time and deliver results more quickly."
Until now, Nvidia has been winning the AI chip race because its GPUs serve as general-purpose workhorses, excelling at the parallel math necessary for training large models. But we've now arrived at the era of agentic AI, where inference is key. While training teaches the AI model to learn from patterns in large amounts of data, inference uses the AI to make decisions based on new information.
Inference can happen on less powerful chips programmed for more specific tasks, such as Cerebras' WSE-3. It falls into a category of chips known as custom ASICs — application-specific integrated circuits. It's an increasingly crowded space, with in-house ASICS now made by the likes of Google, Amazon, Meta and Microsoft.
Cerebras said the WSE-3 is 57 times larger than the largest GPU, and has 50 times the number of transistors.
The most advanced AI chips are made using Taiwan Semiconductor Manufacturing's 2-nanometer process node, currently only possible in Taiwan. Cerebras' chip is also made at TSMC, but on its less advanced 5-nanometer node.
Founded in Silicon Valley in 2016, Cerebras first filed to go public in 2024, but withdrew that submission when it faced scrutiny for its heavy reliance on a single customer, Microsoft-backed AI firm G42 in the United Arab Emirates.
With the firm's successful IPO on Thursday, Feldman and hardware technology chief Sean Lie, two of its co-founders, became billionaires based on their holdings.
For years, Cerebras sought to sell chips to companies, but now largely operates the chips inside its own data centers as a cloud service, pitting it against cloud providers Google, Microsoft, Oracle and CoreWeave.
Cerebras and OpenAI announced a $20 billion cloud deal in January that expires in 2028, while Amazon Web Services announced in March that it's using Cerebras chips in its data centers.
"For our fast inference product, there's so much demand that our biggest challenge is actually trying to supply it. We are adding as much manufacturing and data center capacity as we possibly can, and we're still sold out into 2027," Cerebras CFO Bob Komin told CNBC Thursday.
While hyperscalers make their own in-house ASICs, Cerebras more closely competes with firms that specialize in making them for others. Chief among those is Groq. In its largest purchase to date, Nvidia paid $20 billion for Groq's tech in December, then announced custom Groq Language Processing Units at GTC in March.
SambaNova and D-Matrix are two other notable Cerebras competitors looking to capitalize on unprecedented AI chip demand.
SambaNova counts Hugging Face and Meta among the customers of its SN50 chips, while Intel participated in a $350 million funding round for SambaNova in February. Intel CEO Lip-Bu Tan has served as SambaNova's chairman since 2017.
Cerebras' IPO also paves the way for other custom ASIC start-ups looking to go public, such as Rebellions.
The South Korean chipmaker raised $400 million from the likes of Samsung, at a valuation of $2.34 billion, in March as it prepares for an IPO.
Watch: Breaking down AI chips, from Nvidia GPUs to ASICs by Google and Amazon
Facts Only
* Cerebras Systems conducted an IPO on Thursday.
* Cerebras closed its first day trading with a market cap just below $100 billion.
* Cerebras manufactures chips, such as the WSE-3, which are noted for their large size and high transistor count.
* Nvidia GPUs serve as general-purpose workhorses for training large models.
* The current AI era emphasizes inference, which can be handled by less powerful, specialized chips like the WSE-3.
* Cerebras chips are custom ASICs (application-specific integrated circuits).
* Cerebras' WSE-3 is 57 times larger than the largest GPU and has 50 times the number of transistors.
* Cerebras chips are manufactured on TSMC's 5-nanometer process node.
* Cerebras withdrew an earlier public submission due to reliance on a single customer, G42.
* Cerebras operates its chips largely as a cloud service within its own data centers.
* Cerebras announced a $20 billion cloud deal with OpenAI, expiring in 2028.
* Amazon Web Services announced in March that it is using Cerebras chips in its data centers.
* Nvidia paid $20 billion for Groq's technology in December.
* SambaNova's SN50 chips are customers of Hugging Face and Meta.
* The South Korean chipmaker raised $400 million from Samsung in March as it prepared for an IPO.
Executive Summary
Full Take
The emergence of Cerebras Systems highlights a structural shift in the AI hardware market, moving the focus from generalized compute for training to highly specialized, efficient inference. This shift is not just technological; it represents a re-alignment of power and value away from monolithic general-purpose solutions toward custom application-specific integrated circuits (ASICs). The fact that entities like hyperscalers (AWS) and specialized firms (Groq, SambaNova) are both competing for this custom silicon demonstrates a maturing ecosystem where infrastructure ownership is increasingly tied to proprietary architecture.
The narrative positions Cerebras as a challenger to Nvidia's dominance by arguing that larger, custom chips are superior for the next phase of AI—agentic AI—which relies heavily on fast, localized inference. The pattern here is the leveraging of supply-side limitations: as the demand for AI inference explodes, the need for bespoke hardware creates a niche where players like Cerebras can establish themselves outside the traditional GPU paradigm. However, the central tension lies between centralized power and decentralized innovation. Hyperscalers are building their own in-house ASICs, creating potential redundancy, while specialized firms focus on targeted market segments. The IPO, driven by the deal with OpenAI and the integration into AWS infrastructure, suggests that success in the future will depend less on raw processing power and more on controlling the specialized software and data pipelines surrounding the physical silicon. The critical question is whether this specialization leads to truly decentralized, resilient AI architecture or simply creates a new, segmented oligarchy of chip providers and cloud vendors.
Sentinel — Human
The text exhibits the structure and flow of human-written financial journalism, blending market commentary with technical details effectively.