What You Need to Know

Cerebras Systems’ momentous debut on Thursday not only positioned it among the largest technology IPOs in history, but also served as a resounding testament to the unyielding demand for chips powering artificial intelligence. This surge comes as major tech players intensify their search for alternatives to Nvidia’s coveted, and often scarce, graphics processing units (GPUs).

The company concluded its inaugural trading day on Wall Street with a market capitalization hovering just shy of $100 billion, placing it in elite company alongside giants like Meta and Alibaba. While the stock experienced a 10% dip on Friday, its first full trading session, the initial market reception underscored a significant investor appetite for AI infrastructure solutions.

Here’s a closer look at Cerebras, a formidable contender in the burgeoning AI chip landscape.

Cerebras engineers a distinct class of silicon, characterized by its imposing size – roughly the dimensions of a dinner plate – setting it apart from traditional Nvidia GPUs.

“We construct the largest chips in the semiconductor industry,” stated Andrew Feldman, CEO and Co-Founder of Cerebras, in an interview. “These expansive chips are engineered to process vast amounts of information with unparalleled speed, delivering results more rapidly.”

For a considerable period, Nvidia has dominated the AI chip arena, largely due to the versatility of its GPUs. These processors excel at the parallel computations essential for training complex AI models. However, the industry is now transitioning into an era of agentic AI, where inference—the process of an AI making decisions based on new data—takes center stage. While training imbues AI models with the ability to discern patterns, inference leverages this learned knowledge to execute tasks and generate outputs.

Inference, a more specialized task, can be efficiently handled by less powerful, custom-designed chips. Cerebras’ WSE-3 falls into this category of application-specific integrated circuits (ASICs), a domain experiencing rapid expansion. Leading tech conglomerates such as Google, Amazon, Meta, and Microsoft are all developing their own in-house ASICs to cater to their specific AI workloads.

Cerebras asserts that its WSE-3 chip is an astonishing 57 times larger than the most advanced GPUs and boasts 50 times the transistor count. This architectural difference is key to its performance claims in specialized AI tasks.

The pinnacle of AI chip manufacturing currently relies on Taiwan Semiconductor Manufacturing Company’s (TSMC) cutting-edge 2-nanometer process node, a technology exclusively available in Taiwan. Cerebras also leverages TSMC’s manufacturing prowess, though its chips are produced using the company’s 5-nanometer node.

Established in Silicon Valley in 2016, Cerebras initially filed for its IPO in 2024. However, the company withdrew its filing amid scrutiny over its significant reliance on a single major client, G42, an AI firm backed by Microsoft in the United Arab Emirates. This strategic shift underscores the company’s efforts to diversify its customer base and de-risk its business model.

Following its successful IPO, Feldman and Sean Lie, Chief Hardware Technology Officer and co-founder, have emerged as billionaires, reflecting their substantial equity stakes in the company.

Shifting its strategy from direct chip sales to enterprises, Cerebras now primarily operates its chips within its own data centers, offering them as a cloud service. This pivot places it in direct competition with cloud behemoths like Google, Microsoft, Oracle, and specialized cloud providers like CoreWeave.

The company’s strategic positioning has already garnered significant attention. In January, Cerebras secured a substantial cloud deal with OpenAI, valued at over $20 billion and set to run through 2028. Furthermore, Amazon Web Services announced in March its intention to integrate Cerebras chips within its data center infrastructure.

“The demand for our fast inference product is so immense that our primary challenge is simply meeting that demand,” disclosed Bob Komin, Cerebras CFO. “We are aggressively expanding our manufacturing and data center capacity, yet we remain completely sold out through 2027.”

While hyperscalers are increasingly developing their own proprietary ASICs, Cerebras’s competitive landscape is more directly shaped by companies specializing in custom chip design and manufacturing for third parties. A prominent player in this space is Groq, a startup that Nvidia acquired for an estimated $20 billion in December. Nvidia subsequently unveiled custom Groq Language Processing Units at its GTC conference in March.

SambaNova and D-Matrix are other significant competitors vying for market share amidst the unprecedented demand for AI chips. SambaNova has attracted prominent customers such as Hugging Face and Meta for its SN50 chips. In a significant display of industry confidence, Intel participated in a $350 million funding round for SambaNova in February, with Intel CEO Lip-Bu Tan serving as SambaNova’s chairman since 2017.

Cerebras’s successful IPO not only validates its business model but also signals a promising future for other custom ASIC startups aspiring to go public, including Rebellions.

The South Korean chipmaker Rebellions, preparing for its own IPO, raised $400 million from investors including Samsung in March, achieving a valuation of $2.34 billion. This move highlights the growing global interest and investment in specialized AI hardware solutions.