Industry AI Minute Newsroom 2026-08-19

Same chip, same factory, roughly double the clock — and a rack that draws as much power as forty homes

Cerebras announced the CS-4 on 18 August, the first machine in what it calls the Nexus rack-scale platform. A rack holds up to three of its dinner-plate-sized wafers and adds up to 750 petaflops of sparse FP16 compute, 129.6 petabytes per second of memory bandwidth, 7.2 terabits per second of input and output, and 132 gigabytes of on-wafer memory. The company says it is up to twice as fast as the CS-3 and delivers up to 30 times more tokens per second per user than GPU systems on frontier models. The silicon is not new: it is the same TSMC 5-nanometre design, and The Register calculates that the doubling comes mainly from running it at roughly twice the clock speed, around 2.8 gigahertz, which pushes each wafer to an estimated 33 kilowatts and a full rack to between 120 and 140. First systems come online later this quarter. Cerebras is also pairing its racks with AMD Helios systems and AWS Trainium chips, which handle the prompt-reading half of the work while Cerebras does the answer-writing half.

Why it mattersRead the headline number carefully. The 750 petaflops figure depends on sparsity, which as a rule does not help language model inference much, and the peak memory bandwidth may be theoretical rather than achievable — The Register flags both. What is not in doubt is the direction: the fastest way to serve a model to one impatient user is no longer more chips but hotter ones, and the electricity bill is the honest measure. A rack drawing 140 kilowatts is not a machine you slot into an existing data hall. It is a reason to build a different data hall.

✓ Verified · 3 sources

Same chip, same factory, roughly double the clock — and a rack that draws as much power as forty homes

Related stories