Cerebras launches the CS-4 inference system

Cerebras announced the CS-4, a server rack built around three of its wafer-scale WSE-3 Turbo processors, which the company calls the largest AI semiconductors ever built at 4 trillion transistors each. Cerebras says each chip delivers 250 petaFLOPS of compute with 44GB of on-chip SRAM, and claims the system reaches 30 times the per-user inference speed of rival Nvidia GPUs. The new design uses 50% fewer components than its predecessor to simplify setup. The machine is being sampled by a small group of customers now and will become more widely available later in the third quarter, with another chip-and-system generation planned for 2027.

What it means for you

Cerebras’s whole argument is that keeping an entire model’s weights in on-chip SRAM, instead of shuttling them across a GPU’s memory bus, is what makes token generation fast, and the CS-4 pushes that further. The number that matters for builders is per-user latency: if you run interactive workloads where a user waits on each token, a genuine 30x speedup changes what feels responsive. Two caveats keep this in perspective. The claim is Cerebras’s own and needs independent benchmarks, and limited sampling now means most teams cannot buy one until later this quarter. The launch fits a wider scramble for inference capacity and the money behind it, from Groq’s neocloud pivot to Nvidia’s $105B financing for OpenAI’s Ohio data center. Access, for now, likely means Cerebras’s hosted inference cloud rather than your own rack.