Cerebras Says CS-4 Nears Full Production Run
The November 12 update moves the wafer-scale rack from launch claims to factory output, and the spec sheet still carries the caveats.
The 30x number is a launch claim. The production line is the November news.
Cerebras told investors on November 12 that its CS-4 rack will soon enter full production, three months after the wafer-scale system's August launch.
The machine is three processors, and each one is a whole wafer. Every WSE-3 Turbo carries four trillion transistors, 900,000 cores, and 44 gigabytes of on-wafer memory across 46,225 square millimeters of TSMC 5-nanometer silicon, and the rack totals 750 petaflops with 129.6 petabytes per second of memory bandwidth.
The headline claim needs its footnote attached. Cerebras reports more than 4,400 tokens per second per user on GPT-OSS-120B, up to 30 times a GPU baseline, and its own datasheet says actual throughput varies with model design, context length, precision, and serving setup.
The rack's plumbing is the part built for factories. The Nexus design folds power conversion, liquid cooling, and high-speed I/O into a rear backpack around each wafer, cuts the component count by 50 percent, and moves power conversion to about 0.5 millimeters from the processor against roughly 50 millimeters on a GPU board.
Cerebras also sells the machine as a partner, not only a replacement. A separate engine can read the prompt first and hand the state over, and the CS-4 then generates the answer, with AMD Helios and AWS Trainium named as the complementary platforms in the launch material.
The November update pairs the production line with a capacity claim. Cerebras is expanding manufacturing with Flex and Sanmina and expects to end 2026 at more than ten times the factory capacity it started the year with, and it has already previewed a CS-5 for early summer 2027.
The conversation.
Loading comments…
A little more perspective.
Reading is free. Verified members can comment and reply for $2 a month.