Cerebras Systems – Customer No. 1

One customer is a tough risk to swallow at $7.4bn.

• Cerebras Systems has filed for an IPO at a mooted valuation of $7.4bn (Bloomberg) but with almost all of its revenues coming from one customer and the tough challenge of taking share from Nvidia in this generation of AI, 28x 2024 revenues and 14x 2025 revenues is a tough pill to swallow.
• Cerebras Systems is a semiconductor company that makes silicon chips and software that allow creators of foundation models and the services built upon them, to train and inference their models in the cloud.
• In that regard it is a direct competitor to Nvidia’s data centre business and of all of the would-be Nvidia competitors, Cerebras has by far the most differentiated offering.
• One of the biggest problems with training massive models is that vast numbers of GPUs are required to do the job within a reasonable time frame and all of these GPUs need to talk to each other.
• This is why high bandwidth memory and inter-chip communication are so important in determining the overall performance of LLM training and inference.
• Cerebras has a novel solution to this problem which is to make chips that are so big that all of the GPUs needed can fit on many fewer chips making the memory and communication problems far simpler to deal with.
• The problem with this is that even with the best manufacturing, there are always flaws on every wafer which with smaller chips are simply discarded or used as lower-performance variants.
• With massive chips, it is impossible to discard the flaws and Cerebras has developed a technology that allows it to deal with any flaws that occur and still run the system optimally.
• This architecture allows Cerebras to offer very high performance which I think makes it ideal for certain niche applications and where the customer is going is fully vertically integrated.
• This means that the customer has its own data centres, builds and trains its own models and then sells or offers the inferencing of those services as a generative AI service.
• This means that the models can be optimised for Cerebras chips without having to worry about 3^rd party developers or a vibrant ecosystem of partners.
• To me, this means governments and military applications.
• This is where this proposition comes unstuck in my opinion and the revenue profile of the company clearly demonstrates this.
• Cerebras has demonstrated excellent growth but almost all of it has come from one customer and one shareholder, G42 which is an AI company based in Abu Dhabi which is owned by the state via its Mubadala Investment Company.
• Revenue concentration is about as high as I have ever seen it with 83% of 2023 and 87% of H1 2024 revenues coming from G42 which creates substantial single-source risk for revenues.
• This dependence is also evident on the balance sheet where during H1 2024, G42 has deposited $300m meaning that there will be no more cash inflows until the company wins new customers, G42 has spent more than another $300m with Cerebras or there is a capital raise such as this one.
• Being dependent on vertically integrated niche applications is not necessarily a problem until one considers the $7.4bn valuation issue.
• In order to grow into that valuation, Cerebras needs to break into the wider market and attract 3^rd party developers and create a vibrant ecosystem rather than relying on vertically integrated customers who will not be large enough in the long run.
• This is going to be difficult because Nvidia’s CUDA platform has a 20-year head start and is the most mature where all AI developers know how to use it and will select it given the option.
• This creates a colossal barrier for Cerebras and I am not convinced that it will be able to overcome this while the control point of the ecosystem remains at the silicon development platform.
• As the AI ecosystem matures, I can see the control point migrating further up through the AI technology stack to the foundation model or services and at this point (maybe 2-4 years from now), Cerebras will be in a much better position to compete with Nvidia directly.
• Until then, Cerebras will be dependent on vertically integrated niche markets which is not a problem other than the company has a high valuation to justify.
• Consequently, while I like the technology and the niche position of Cerebras, the single customer risks and the difficulty of taking share from Nvidia over the next 24 months at least keep me cautious on the 12 to 24 months outlook and valuation.