Cerebras Maps 3T Models Across 59 Wafers

The important story is the deployment math, not just the speed claim: Cerebras shows a credible way to keep enormous model weights resident while streaming activations through a wafer-scale pipeline, but the headline throughput still needs independent benchmarking.

• –Dense layers get dedicated wafers for KV-cache capacity, while two linear-attention layers fit per wafer.
• –Only activations cross wafers over roughly 2-microsecond RDMA hops; model weights generally stay in place.
• –Pipeline concurrency keeps wafers busy across multiple requests, with micro-batch size trading latency for aggregate throughput.
• –The post does not tie the 1,000–2,000+ tokens-per-second figure to a specific Kimi K3 run or provide matched GPU and tokens-per-watt comparisons. [Independent analysis](https://www.aiinsiders.net/article/how-cerebras-fits-a-3-trillion-parameter-ai-model-across-59)
• –The approach is compelling for latency-sensitive agents and real-time applications, but prefill performance, hardware cost, availability, and failure-domain risks remain open questions. [Developer discussion](https://news.ycombinator.com/item?id=50227356)