This company is relatively obscure to the general public, yet it holds extraordinary technological achievements in the hardware and AI space — and despite a few asterisks, it may soon be heading toward a massive IPO.
-
The company's name is Cerebras.
It is preparing to go public and raise up to $4.8 billion, at a valuation of more than $26 billion.
But what makes it so special that you simply can't find it at any of its giant competitors?
-
Chips are etched by laser onto silicon wafers 300 millimeters in diameter, then sliced into small pieces — each one a separate processor.
There are two reasons for this.
The first is that the etching process is so complex that some transistors always come out defective; dividing the wafer into many processors means only a small fraction need to be discarded while the rest are usable.
The second reason is that etching is performed using reticles, the largest of which cover 850 square millimeters. This means the entire wafer cannot be exposed in a single shot, and the gaps between successive exposures become the boundaries along which the chips are cut.
Cerebras designed its entire architecture from the ground up, and manufactures (via TSMC) monstrous processors featuring 900,000 cores and 44 GB of RAM distributed directly across the die and connected directly to the processing cores.
This architecture comes with significant engineering challenges: routing around defective cores, complex cooling and power-delivery systems, and stitching together the seams between reticle exposures to unify the die into a single giant processor.
For these reasons, the company sells complete server cabinets rather than the processor alone.
The practical upshot is that instead of routing data between separate processors or between a processor and external RAM, data simply flows within the chip itself at a staggering 21 petabytes per second.
These chips host language models from companies such as OpenAI and are capable of generating tokens at remarkable speed. That speed delivers advantages across several domains: voice applications that sound more natural, real-time systems such as fraud detection, and commercially, because the processor sells tokens at an exceptionally high throughput.
-
Several concerns cast a shadow over the IPO plans.
The company's current revenue stands at $510 million, but 86% of that comes from just 2 customers in the Gulf states region.
Cerebras also holds an order backlog of $25 billion, though those orders stretch many years into the future — and $20 billion of them are from OpenAI, which also holds stock options in the company (alongside Amazon). The problem is that OpenAI itself is deep in trouble, and barring a miracle it may be forced to dissolve within two years.
Greg Brockman and Sam Altman, two of OpenAI's co-founders, are among the company's earliest investors, giving them a significant stake in that $20 billion deal.
The dependence on a handful of customers and investors places the company in a position of considerable risk relative to established players like Nvidia.
That said, even if Cerebras struggles to grow as a standalone commercial entity, its technological capabilities are impressive and highly valuable — and it will most likely simply need to choose the best offer from the acquisition bids that are sure to come its way.
--
👋 Hi, I'm Shlomo Strauss — follow me for more content on science and technology.