Cerebras, which produces AI chips in a life-size “wafer-scale” format, was an instant favorite among Nasdaq investors. The long-awaited IPO led to a 68 percent increase in the value of the brand-new Cerebras stock. On paper, the company is thus worth $95 billion at the time of writing. The question now is whether the company can live up to the sky-high expectations as a publicly traded company.
Part of the explosive growth can be explained by the fact that no one wants to miss out on the latest Nvidia. However, while the GPU maker benefited immediately from the AI boom as an established player, this has been a more difficult path for Cerebras to navigate. First of all, it had to produce its first hardware with “only” tens of millions of dollars—a pittance when it comes to funding the development of advanced processors.
Wafer-scale AI
A new AI company needs to stand out. Many players fail to do so, whether they focus on software or hardware. In the early years following ChatGPT, many software companies did little more than repackage LLMs, and many “AI chip companies” had only an edge computing chip available that could be useful in very limited scenarios. For AI training and AI inference at scale, GPUs from Nvidia or AMD, TPUs from Google, or Trainium chips from AWS were and are required. Cerebras’ unique processors may also be included in that list, though with $510 million in annual revenue projected for 2025, it does not yet hold a significant market share.
Where Cerebras has carved out a unique position is in designing an AI chip that utilizes an entire chip wafer. These are the round silicon wafers that, after processing, contain dozens or hundreds of unique chips. At least, that is normally the case. After the wafer is processed (often in an ASML machine, among other methods), it is split up. The best-performing chips are often high-end processors, while the less-performing sections of the wafer are often still usable for a cheaper product. That paradigm doesn’t fit Cerebras, where the chip wafer isn’t split up at all.
This means that Cerebras uses a completely different design. In other words: the company assumes that there is some margin of error in the manufacturing process, but as a whole, a wafer must be able to function as a single chip. The advantage is that communication between the cores and the on-chip memory on that wafer is incredibly fast, especially compared to memory that is connected to the silicon after the fact or accessed via DRAM or SSD storage.
AI chips revolve around both hardware and software
Cerebras may be over ten years old, but it is, in fact, a relative newcomer to the chip market. Nvidia, Google, AWS, and AMD are the players with varying levels of adoption, software adaptation, and large-scale users. Even though Google and AWS owe the use of their chips in part to large-scale investments in a company like Anthropic, it does mean that their chips power companies’ day-to-day production AI. The biggest disadvantage for AMD is that, unlike Nvidia, it does not dominate a software ecosystem, which is why the latter is still considered the standard in the field of AI training and inference. Eventually, the sky-high costs will make the wide range of alternatives—particularly in inference—an increasingly diverse playing field. That is where the opportunities lie for a challenger like Cerebras.
For now, it is unclear whether Cerebras truly has what it takes to grow significantly. That is to say: it will most likely continue to report revenue increases in the quarterly earnings reports it is now required to file as a publicly traded company, but the question remains whether those increases will lead to an ever-growing market share. If not, the numbers could still turn out well for Cerebras—and for a long time. But without a software ecosystem, it isn’t “sticky.” The same problem seemed to apply to Groq, another alternative AI chipmaker, but that company was effectively acquired by Nvidia itself.
As for market capitalization: that’s always speculative, but we’ve seen spectacular first days on the stock market regularly followed by disappointing declines. Dotcom companies were notorious for this (see: TheGlobe.com or eToys.com), but Bumble, Rivian, Peloton, and Figma can also attest to it.