<div id="readability-page-1">AI chip startup Cerebras has completed one of the biggest IPOs in semiconductor history, issuing 30 million shares to raise around $5.5 billion on the NASDAQ (comparable to Arm’s 2023 offering, which raised $4.9 billion).In the weeks leading up to the IPO, the expected price range for the shares grew from $115-$125 to $185 at launch on Thursday (May 14). By the end of Thursday, they were trading at an enormous $311. This price puts Cerebras’ market capitalization at around $66 billion, a figure unthinkable when Cerebras first announced its intent to IPO in 2024.Amongst a sea of data center chip startups, Cerebras and competitor Groq had both pivoted from chip supplier to cloud service provider, building and deploying their own data centers to serve developers directly via API. The idea was to avoid having to rely on lukewarm hardware sales in a market still dominated by Nvidia, but it also showed off what Nvidia’s competitors could do—both companies had identified Nvidia’s Achilles heel as single-user token speeds, an area both Groq and Cerebras excelled in. Demand grew for fast tokens as applications such as coding and agentic AI took off, validating the market for speed.Nvidia’s $20-billion deal for Groq set the bar for the valuation of an AI chip startup and immediately validated the market for non-GPU hardware. With a similar vertically integrated chip offering and cloud service to Groq, perhaps the $20 billion figure provided a baseline for Cerebras’ valuation, though it seems clear now that Nvidia’s interest was in Groq’s technology, not its cloud deployments.Groq’s deal and Cerebras’ IPO may serve to boost valuations of companies such as D-Matrix, Tenstorrent, SambaNova, Rebellions, and others, though it seems unlikely there could be many more companies shopping for an acquisition at $66 billion or above. Rather, the market has shown it is now ready to support large semiconductor IPOs again.Cerebras had filed an S-1 in October 2024, which it later withdrew. That paperwork revealed that a single customer, Abu Dhabi-based G42, was responsible for 87% of Cerebras’ revenue in the first half of 2024. This concentration of customers was widely criticized at the time.Since then, Cerebras has broadened its customer base from one… to three significant customers. The second, MBZUAI, is an Abu Dhabi-based University and research institute. The third is none other than OpenAI.Cerebras will supply 750 MW of compute capacity to OpenAI for a fee of around $20 billion, plus a significant advance to fund the required buildout. OpenAI also got a significant stake in Cerebras—as much as 10%. This may seem like a lot of equity to give away, but it’s likely Cerebras’ public association with OpenAI’s golden brand was worth a lot more.(These circular deals, in which the same company is both customer and investor, are becoming more commonplace in the industry. OpenAI did a similar deal with AMD recently, and Cerebras’ existing deal with G42 has similar elements.)Cerebras’ crown jewel is its wafer-scale engine (WSE), currently in its third generation. This dinner-plate-sized chip is a huge feat of engineering in its own right—let alone the custom server built to house it. The company can generate fast tokens since its architecture is SRAM-based; it is not reliant on memory bandwidth between compute and DRAM like GPUs and other architectures.However, there are limits to what the wafer can do. The WSE3 has 44 GB of SRAM, considerably less memory than a state-of-the-art GPU’s HBM, but with much higher bandwidth. 44 GB isn’t enough to hold the whole of a model on a single piece of silicon anymore. Instead, multiple wafers need to be connected together, against the original thesis for the architecture. The wafer has relatively limited I/O, which Cerebras currently gets around by passing only activations between wafers (which are smaller than weights or data). However, it’s unclear whether this can continue to scale for even bigger models. And since the wafer was originally designed for training, it doesn’t natively support lower precision number formats used for inference (current support is down to FP16). (SemiAnalysis has a detailed technical report here).A fourth-generation WSE could fix some of these problems, but SRAM density scaling is leveling off for nodes beyond the WSE3’s, likely a key reason the company hasn’t announced a fourth generation yet. Could a WSE4 use 3D stacking on top of the wafer to add more memory? And will Cerebras go to optical wafer-to-wafer connectivity in some form? Both would be significant engineering challenges, but that’s something Cerebras is no stranger to.With such a high share price, expectations for Cerebras are extremely high, meaning the company will need to show very high growth in the coming quarters. It is also facing much, much stronger competition from the new Nvidia-Groq combination in its key segment, fast token generation, due to reach general availability in the third quarter. As the first of its cohort to IPO, Cerebras has proven Wall Street’s appetite for funding Nvidia alternatives, but many other companies’ futures now depend on its success or failure.Fallout From Nvidia-Groq Deal Validates AI Chip Startup LandscapeFallout from Nvidia’s deal with Groq, in which the GPU giant reportedly paid $20 billion for a non-exclusive license to Groq’s technology and hired most of its technical team, has had two big ramifications across the AI chip industry.RELATED TOPICS: AI, DATA CENTER, IPO, SEMICONDUCTORS, SRAM, STARTUPSCOMPANIES: CEREBRAS, CEREBRAS SYSTEMSSally Ward-Foxton covers AI for EETimes.com and EETimes Europe magazine. Sally has spent the last 18 years writing about the electronics industry from London. She has written for Electronic Design, ECN, Electronic Specifier: Design, Components in Electronics, and many more news publications. She holds a Masters' degree in Electrical and Electronic Engineering from the University of Cambridge. Follow Sally on LinkedInYou must Register or Login to post a comment.This site uses Akismet to reduce spam. Learn how your comment data is processed.</div>

CBRS

NVDU

SOXL

Cerebras has completed a significant IPO, raising approximately $5.5 billion by issuing 30 million shares, with shares trading at $311, giving it a market cap of around $66 billion. The company has shifted from being a chip supplier to a cloud service provider, competing with Nvidia. Cerebras has expanded its customer base, now including OpenAI, which will receive 750 MW of compute capacity for $20 billion. However, the company faces challenges with its wafer-scale engine's memory limitations and increasing competition.

Cerebras IPO Revives AI Chip Startup Fever