Cerebras Systems is an AI chip company that builds its processor from an entire silicon wafer instead of cutting the wafer into many smaller chips. It sells that processor in rack systems, the CS-3 and the CS-4 announced in August 2026, and rents its speed through the Cerebras inference cloud, which in the second quarter of 2026 brought in more revenue than hardware. Cerebras has been listed on the Nasdaq Global Select Market under the ticker CBRS since May 14, 2026.

On this page

Why a wafer-scale chip

A conventional AI accelerator is one die cut from a wafer, with its working memory (HBM) in separate stacks beside it, and large models are spread across many such chips. Cerebras keeps the wafer whole and puts memory on it as SRAM, so more of a model’s data sits next to the compute cores and less has to travel between chips.

The WSE-3, announced March 13, 2024, has 4 trillion transistors on TSMC’s 5nm process, 900,000 “AI optimized compute cores,” 44 GB of on-chip SRAM and 125 petaflops of peak AI performance, according to Cerebras. Its chip page gives the wafer’s area as 46,225 mm².

WSE-3T and the CS-4

On August 18, 2026, Cerebras announced the CS-4 and its processor, the WSE-3T (WSE-3 Turbo). The WSE-3T keeps 4 trillion transistors, 900,000 cores and 44 GB of SRAM, and Cerebras says memory bandwidth doubles to 43.2 petabytes per second. Cerebras’s chip page lists the WSE-3T at 250 petaflops of AI compute.

The CS-4 is a rack-scale system built from three WSE-3T wafers. Cerebras says it “delivers 750 PFLOPs of AI compute, 7.2 terabits per second of I/O, and 129.6 petabytes per second of memory bandwidth,” and that “first CS-4 shipments begin this quarter.” Its headline claim is “more than 4,400 tokens second per user” on GPT-OSS-120B, which Cerebras calls “up to 30 times faster than GPU solutions.” That comparison is Cerebras’s own, on that model.

WSE-3 (CS-3) WSE-3T (CS-4)
Announced March 13, 2024 August 18, 2026
Transistors per wafer 4 trillion 4 trillion
AI cores per wafer 900,000 900,000
On-wafer SRAM 44 GB 44 GB
Peak AI compute per wafer 125 PFLOPs 250 PFLOPs
Wafers per system 1 3

For the CS-3, Cerebras offered external memory options of 1.5 TB, 12 TB or 1.2 PB and clusters of up to 2,048 systems. Other accelerators are compared in our AI chip catalog.

The inference cloud

Cerebras launched Cerebras Inference on August 27, 2024, as an API compatible with OpenAI’s format, served from Cerebras Cloud, with private-cloud and on-premises deployment for enterprise customers. At launch it reported 1,800 tokens per second for Llama 3.1 8B and 450 for Llama 3.1 70B.

The cloud is now the larger part of the business. In its second-quarter 2026 results, released August 12, 2026, Cerebras reported GAAP revenue of $180.1 million, of which $126.0 million was cloud and services and $54.1 million hardware. It said GAAP cloud revenue grew 281% year over year. In the first quarter, reported June 23, 2026, the split was the other way: $110.6 million hardware and $82.8 million cloud and services out of $193.4 million.

Quarter (2026) GAAP revenue Hardware Cloud and services
Q1 $193.4 million $110.6 million $82.8 million
Q2 $180.1 million $54.1 million $126.0 million

Customers and contracts

On January 14, 2026, OpenAI announced a partnership for 750 MW of “ultra low-latency AI compute” from Cerebras, coming online “in multiple tranches through 2028.” Cerebras describes it as a multi-year deal valued at over $20 billion. Cerebras says it was OpenAI’s launch partner for GPT-5.6 Sol, serving it at 750 tokens per second.

In its second-quarter release, Cerebras said it had over 600 MW of data-center capacity under contract for delivery by the end of 2027, and named new agreements with Cognition and Lovable. It lists Block, Figma, AlphaSense and GSK as users in agent workflows, security work with CrowdStrike, a multi-year partnership with AWS to distribute its inference, a planned Amazon Bedrock integration, and joint disaggregated-inference work with AMD. Cerebras also says it is expanding manufacturing capacity more than tenfold in 2026 with Flex, Sanmina and Rocket EMS. Its data-center build-out is tracked with other projects in our AI data center tracker.

Public company

Cerebras shares began trading on the Nasdaq Global Select Market on May 14, 2026 under the symbol CBRS, and the company announced the closing of its initial public offering on May 15. Cerebras says it raised $6.4 billion in the IPO. Our AI companies directory records the listing alongside other AI chipmakers and labs.

Sources