tl;dr: Cerebras runs trillion-parameter models like the 3T-parameter Kimi K3 by pipelining layers across multiple WSE-3T wafers (59 in this example), with each 44GB wafer holding one dense layer plus KV cache, or two linear-attention layers. Weights stay resident on-chip and only activations pass between wafers via RDMA over Ethernet (~2µs per hop). Concurrency is maximized by keeping the pipeline full with micro-batched requests at each stage, with KV cache capacity on dense-layer wafers setting the ceiling on simultaneous requests.