Cerebras Systems

What Cerebras builds

Cerebras Systems is an American company that designs AI processors and the computers built around them. Its central idea is wafer-scale computing: a conventional chipmaker builds hundreds of circuits on a silicon wafer and then cuts the wafer into individual chips, whereas Cerebras keeps the wafer intact and treats it as one very large processor ^[1]. The motivation is that moving data between separate chips is far slower than moving it inside one chip, so keeping more compute and memory on the same piece of silicon can pay off. The difficulty is that a single defect anywhere on a wafer can spoil the whole device, and when the company was founded it was not clear that the approach was technically possible ^[1].

Origins

Cerebras was founded in 2015 by Andrew Feldman, who is chief executive, together with Gary Lauterbach, Michael James, Sean Lie and Jean-Philippe Fricker. The stated goal was to bring wafer-scale computing to market ^[1].

The first Wafer Scale Engine

In August 2019 the company unveiled the Wafer Scale Engine (WSE), a chip with more than 1.2 trillion transistors and 400,000 cores, and launched it in a system called the CS-1 ^[2]. Transistors are the basic switches on a chip and cores are the units that carry out arithmetic, so both numbers describe how much circuitry sits on the one wafer.

WSE-3 and CS-3

In March 2024 Cerebras announced the third-generation WSE-3 and the CS-3 system built on it. The company says the chip doubled the performance of WSE-2 at the same power and price, with 4 trillion transistors, 900,000 AI cores and 44 GB of on-chip memory on a 5 nm process ^[3]. In the same announcement Cerebras says the WSE-3 delivers 125 petaflops of peak AI performance, that its 44 GB of on-chip memory is SRAM made by TSMC on a 5 nm process, and that CS-3 machines can be grouped into clusters of up to 2,048 systems ^[3]. On-chip memory means storage built on the same slab of silicon as the arithmetic units, which is why the company treats the size of that memory as a headline number: data that stays on chip does not have to travel out to external memory and back.

Service, listing and CS-4

Cerebras says that in 2025 its inference service was powering Meta's Llama API, Perplexity and Mistral. It says it began trading on Nasdaq under the ticker CBRS in May 2026, and that in August 2026 it launched the CS-4, a system powered by three WSE-3 Turbo processors ^[1]. Inference here means running a trained model to answer requests, as opposed to training one.

Where it sits

Across the generations the design argument has stayed the same: the larger the single piece of silicon, the more cores and memory sit inside one device, and the fewer separate chips a system has to connect ^[1]. The approach trades manufacturing difficulty for integration density, and each generation is the company's own test of whether that trade holds up.