The chip is fast. The rack is the trick.

CS-4 is Cerebras's fourth-generation system: three WSE-3T wafer-scale processors plus a Nexus rack that treats compute, power, cooling, and I/O as one design. The headline is more than a faster chip: Cerebras reports more than 1,000 tokens per second on models above 10 trillion parameters, enabled in part by wafer-to-wafer latency as low as 2 microseconds. 12

That architecture attacks several waits at once. Power conversion moves close to the processor; modular compute backpacks separate compute from facility infrastructure; and disaggregated inference lets a GPU or ASIC handle prefill before CS-4 handles low-latency decoding. First CS-4 shipments begin this quarter. 1

One caution matters: up to 30× faster is Cerebras's comparison across the models shown, based on third-party benchmarking and internal testing. Results vary with workload, configuration, date, and model. The durable mechanism is the co-design, not the single multiplier. 12