Cerebras Systems Inc., a maker of supersized artificial intelligence chips, today disclosed the financial terms of its upcoming public offering.

The company hopes to raise as much as $3.5 billion by selling 28 million shares for $115 to $125 apiece. Cerebras’ underwriters, the banks entrusted with coordinating its listing, have the option to buy an additional 4 million shares. The value of the offering could increase even further if Cerebras boosts its price target, which fast-growing tech firms often do when there’s strong investor demand.

The chipmaker’s sales jumped 76% in 2025, to $290.3 million. Moreover, it turned a $87.9 million profit after losing $485 million a year earlier.

Cerebras’ improving financial performance is likely one of the factors beyond its ambitious valuation target. The public offering terms the company is seeking would give it a market capitalization of $26.6 billion at the top end, a $3.6 billion increase from February. That month, Cerebras raised $1 billion in funding at a $23 billion valuation.

The company sells a wafer-size AI chip called the WSE-3 that is several times the size of Nvidia Corp.’s Blackwell B200 graphics processing unit. One of its main selling points is a 44-gigabyte pool of SRAM. SRAM is a memory variety that features significantly more transistors per square millimeter than standard server DRAM. As a result, it’s considerably faster and several orders of magnitude more expensive.

Processors include a quartz crystal that moves millions of times per second in response to an electric current. Each movement corresponds to a clock cycle, the basic unit of time in chips. Cerebras says WSE-3’s 900,000 cores can access the onboard SRAM pool with latency of one clock cycle, which is significantly faster than what a standard graphics card offers. The result is faster performance on inference, the process of providing responses to queries.

Cerebras ships the WSE-3 as part of a 1.8-ton appliance called the CS-3. Customers can link together multiple CS-3 systems into a cluster with the help of auxiliary devices that are likewise provided by the startup. 

There’s a specialized storage system, MemoryX, that is optimized to hold activations. Those are intermediate data points that an AI model generates while processing a prompt and discards immediately after the calculation is complete. A switch called SwarmX manages the task of moving data between the MemoryX and CS-3 systems in a cluster. Additionally, the switch compresses some of the data to optimize storage hardware utilization.

Cerebras filed to go public last month after inking a broad chip supply deal with OpenAI Group PBC. Under the agreement, it will provide the ChatGPT developer with as much as 2 gigawatts of computing capacity through 2030. Cerebras expects the contract, which is reportedly worth over $20 billion, to account for a substantial portion of its revenue in the coming years.

The public cloud market could emerge as another important revenue source. A few weeks after announcing the OpenAI deal, Cerebras teamed up with Amazon Web Services Inc. to make the WSE-3 chip available in the cloud giant’s platform. AWS’ adoption of the chip may prompt other infrastructure-as-a-service providers to follow suit, which could translate into significant sales opportunities for Cerebras.

Photo: Cerebras