Is it bad to rely on one customer for 87% of your revenue? An AI company that’s going public is about to find out

Investors have long been waiting for the proverbial IPO window to reopen, and chipmaker Cerebras Systems may be the answer to their prayers. On Monday, Cerebras, an AI chip maker, filed its S-1 to prepare for an IPO, hoping to capitalize on investor optimism in the sector by joining the public markets. However, Cerebras’s S-1 reveals some glaring issues that I’d like to discuss.

In case you’re unfamiliar with Cerebras (I was), it is the creator of the world’s largest semiconductor. While other chip makers have created smaller and smaller chips over time, Cerebras went big, literally, and its chips are the size of a dinner plate. This larger size provides processing power advantages: While Nvidia’s latest Blackwell chips have 200 billion transistors, Cerebras’s chips have 4 trillion transistors. Additionally, while Nvidia chips are heavily reliant on external memory, Cerebras’s chips host memory directly on the chip, allowing its products to more quickly complete AI inference tasks (inference is the process of using pre-trained AI models to make predictions. The other primary AI function is “training, where a model is taught how to perform a task). Cerebras claims that its chips can complete inference tasks 20 times faster than Nvidia’s GPUs.

That’s great, right? If Cerebras’s chips are 20 times faster than competitors’ products, you would think that big tech companies would be knocking down their door to buy as many as possible. But that hasn’t been happening. While tech giants like Microsoft, Meta, and Alphabet are spending billions on Nvidia’s chips, 87% of Cerebras’s revenue comes from “G42,” an Abu Dhabi-based AI company founded in 2018.

Critics have flagged Nvidia’s revenue concentration as a risk, noting that almost half of Nvidia’s revenue comes from just four customers, but Cerebras’s revenue concentration makes this look like child’s play. And Nvidia’s biggest customers are the largest tech companies in the world, not a six-year-old company based in Abu Dhabi. Cerebras mentioned “G42” an incredible 301 times in its S-1, noting in its risk section:

We currently generate a significant majority of our revenue from one customer, G42, and a significant portion of our revenue from a limited number of customers. A reduction in demand from, or a material adverse development in our relationship with, G42 or any of our other significant customers may harm our business, financial condition, results of operations, and prospects.

The S-1 filing shows that G42 also owns a ~1% stake in Cerebras after investing in the company’s 2021 Series F, and Cerebras has granted its primary customer some… favorable terms to expand its investment. Cerebras listed a separate “Class N” stock offering reserved for G42, allowing the company to purchase $335 million in shares at $14.66 per share, a discount to its Series F price of $27.74 per share in December 2021, and G42 also has the right to purchase Class A shares at a 17.5% discount to their fair market value.

I know everyone wants to invest in the next hot AI company, and plenty of investors will probably be willing to overlook customer concentration risk, especially when that company’s revenue increased from $8.7 million through the first six months of 2023 to $136.4 million in the same period in 2024. That being said, I do think it’s worth proceeding with caution given Cerebras’s dependence on, and weird relationship with, its top customer.