Is this company an “Nvidia killer”? What you need to know about Cerebras’ IPO

Conventional wisdom holds that Nvidia (NVDA 1.69%) Since the introduction of ChatGPT, it will continue to dominate the artificial intelligence (AI) chip market. But it’s facing stiff competition not only from competitors and cloud giants that make their own accelerators, but also from AI chip startups.

One such startup, Cerebras, has just filed its prospectus ahead of its impending initial public offering (IPO). After reading this, I think Cerebras is a name that all Nvidia investors should keep a close eye on. But is it really a threat to the graphics processing unit (GPU) giant?

Image source: Getty Images.

What is Cerebras?

Cerebras was founded in 2016 by current CEO Andrew Feldman and a group of engineers who founded or worked at a company called SeaMicro more than a decade ago. SeaMicro developed efficient high-bandwidth microservers and was subsequently acquired by Advanced Micro Devices in 2012.

Cerebras sold its first AI chip in 2019 and has seen a significant acceleration in demand recently, leading to this IPO filing.

giant brain chip

Cerebras’ big differentiator is that its AI chip, called a wafer-scale engine (WSE), is huge. And by huge, I mean a chip that occupies the entire semiconductor wafer. Foundries typically produce many chips per wafer, some of which are defective and are scrapped. But Cerebras is aiming for one giant chip per wafer.

The result is a massive processor that is 57 times larger than Nvidia GPUs, with 52 times more compute cores, 880 times more on-chip memory, and 7,000 times more memory bandwidth. A single Cerebras WSE contains an astonishing 4 trillion transistors. That’s 50 times the 80 billion transistor count of Nvidia’s H200. Like Nvidia, Cerebras’ chips are manufactured by Taiwan Semiconductor Manufacturing.

The theory behind making the chips so large is that by performing more processing on the chip, WSE eliminates the need for Infiniband or Ethernet-based network connections to stitch together hundreds or thousands of GPUs. It’s a thing. Cerebras says this architecture allows WSE to achieve training and inference more than 10 times faster than an 8-GPU Nvidia system.

Feldman said in a recent interview that recent tests showed Cerebras’ chips were 20 times faster at inference than Nvidia’s chips. Is it impressive? When Feldman was asked at a summer conference how much market share Cerebras intended to take from Nvidia, he said, “Everything.”

Finances greatly accelerated

Not only does Cerebras talk about big games, but as you can see, it’s also showing strong revenue acceleration and improved profitability this year.

Data source: Cerebras S-1. H1 = first half of the corresponding year.

As you can see, between H1 2023 and H1 2024, Cerebras’ revenue surged by 1,474%. Technically, gross margin decreased from 50.5% to 41.1%, primarily because nearly all of last year’s revenue came from high-margin services. In fact, Cerebras’ hardware gross margin increased over that period. Even better, operating losses narrowed by $40 million, a great indicator that the company could benefit from scaling.

This exponential expansion should continue next year. Cerebras’ largest customer, Abu Dhabi’s G42, has agreed to buy $1.43 billion in equipment by the end of 2025, according to filings. This is a six-fold increase compared to the current occupancy rate in 2024.

Risks to the Cerebras story

However, there are some risks to the Cerebras story. One is that manufacturing one giant chip can introduce a large number of defects. Nvidia and other chipmakers can scrap all the bad chips on their wafers, but Cerebras has to take it all on and expose the WSE to imperfections.

Cerebras says it assumes that many chips have defects, so it has created “redundant” cores and interconnects on the chips to avoid this. “Deficiencies are designed to be recognized, blocked and avoided,” the application states.

However, building in redundancy also means that Cerebras cannot exploit the full potential of the chip’s surface area. Clearly, management believes that the “big chip” architecture more than compensates for this inefficiency.

The second risk, and perhaps the biggest one, is Cerebras’ customer concentration. Currently, UAE AI company G42 accounts for 87% of Cerebras’ revenue in the first half of 2024. G42 and its affiliates are also behind next year’s $1.43 billion order intake, meaning the concentration will further increase.

In the early stages of a company’s growth, some degree of concentration is expected. However, if something goes wrong with the relationship or with the G42 itself, Cerebras’ plans could be seriously derailed. G42’s close ties to foreign governments (the UAE’s national security adviser is the company’s founder and largest shareholder) certainly pose a risk in the event of geopolitical escalation.

Celebras is worth paying attention to

If it goes public, Cerebras will become a hot new AI player and will likely sell for a high price. Therefore, investors need to be cautious about how much they pay for stocks in the market.

Nevertheless, the company has an architecture that differentiates it from other companies. So it’s worth keeping an eye on whenever it goes public, especially if you’re a large Nvidia or AMD shareholder.