Cerebras CS-4

Wafer-scale inference system on Cerebras' new Nexus rack architecture

N/A Datacenter
Visit Product Page →
44GB SRAM
43200 GB/s bandwidth
23000W TDP
900000 cores
5nm process

Cerebras unveiled the CS-4 on August 18, 2026, built around its new Nexus rack architecture that houses three WSE-3T wafers per system instead of the single wafer used in the CS-3. Each wafer carries 900,000 AI cores and 44GB of on-wafer SRAM, and Cerebras claims up to 30x more tokens-per-second-per-user than GPU-based inference. The modular backpack design roughly doubles power and cooling headroom over the CS-3, and the company has already outlined CS-5 and CS-6 as the next steps on the same platform.