Cerebras Systems Inc., the developer of the wafer-size WSE-3 artificial intelligence chip, today filed to go public.

The move comes about 18 months after the company’s first attempt to list its shares. It filed for an initial public offering in September 2024, but withdrew the paperwork late last year. Cerebras explained at the time that its original IPO filing “no longer reflected the current state of our business.”

In 2024, the chipmaker lost $485 million on sales of $290.3 million. Last year, it swung to a $87.9 million profit. Cerebras’ revenue rose by 76% in the same time frame, to $510 million.

The company disclosed in today’s IPO filing that it has secured a $125 million revolving credit facility from Morgan Stanley. The funds will be used to finance deals with data center developers and operators. Cerebras is expanding its data center capacity to support the growth of its Training Cloud and Inference Cloud services, which provide access to hosted AI infrastructure.

The company’s services are powered by its flagship WS-3 chip. It’s an AI accelerator 58 times the size of the B200, a high-end Nvidia Corp. graphics card that debuted in 2024 and remains highly popular. The WSE-3 contains 4 trillion transistors organized into 900,000 cores.

According to Cerebras, Morgan Stanley will upscale its revolving credit line to as much as $850 million following the IPO. The company also disclosed in the filing that it has received a separate $1 billion loan from OpenAI Group PBC.

Last December, the ChatGPT developer agreed purchase 750 megawatts worth of inference infrastructure from Cerebras. The chipmaker revealed today that the deal is worth more than $20 billion. Additionally, it gives OpenAI the option to add an additional 1.25 gigawatts of capacity through 2030.

Cerebras has issued OpenAI warrants to purchase up to 33.4 million shares. Those warrants will vest if the AI model developer goes through with its plan to purchase 2 gigawatts of computing capacity by 2030. Cerebras stated in its IPO filing that the contract “represents a substantial portion of our projected revenues over the next several years.” 

Last month, Cerebras inked a high-profile chip deal with another high-profile customer. Amazon Web Services Inc. agreed to deploy the WSE-3 in its data centers as part of a new “disaggregated architecture.”

The workflow through large language models process prompts comprises two steps known as the prefill and decode stages. AWS’ disaggregated architecture will use its internally developed AWS Trainium chips to perform prefill calculations. The WSE-3, in turn, will be responsible for the decode phase.

Decode calculations are similar to those used by the prefill workflow, but they require more memory bandwidth. That’s a measure of how fast data can move between a chip’s logic and memory circuits. The WSE-3 provides 27 petabytes per second of memory bandwidth, more than 200 times the amount offered by Nvidia’s NVLink interconnect.

Cerebras’ product roadmap “includes the development of a disaggregated inference-serving solution,” the company stated in its IPO filing. “Disaggregated inference would allow Cerebras to operate alongside other architectures, serving as the high-performance engine for decode while other systems handle prefill.”

The filing hints that Cerebras’ disaggregated inference platform will work with other third-party chips besides Trainium. It’s possible the planned offering will be built on a version of the CS-3, the company’s data center appliance. It combines a single WSE-3 with cooling equipment, power management components and other supporting hardware. A built-in software tool called the Cerebras Cluster Manager makes it possible to link thousands of CS-3 appliances into a single cluster.

Cerebras plans to list its shares on the Nasdaq under the ticker symbol “CBRS.”

Photo: Cerebras