It takes 59 dinner-plate chips to run a single AI model

Cerebras published a technical walkthrough of how it would serve Moonshot AI's Kimi K3 — a roughly 3-trillion-parameter open model — across 59 of its WSE-3T wafer-scale chips, each holding 44 GB of on-chip memory. At 4-bit quantization each of the model's 93 layers needs about 16 GB, so dense layers get a wafer each and linear-attention layers pack two to a wafer: 35 plus 24 wafers, with only activations crossing between stages at roughly two microseconds a hop.

The company advertises 1,000 to 2,000-plus tokens per second for multi-trillion-parameter models — but the post is a deployment recipe, not a measured result: Kimi K3 is explicitly an 'illustrative example,' and Cerebras gives no measured power figure for this layout. The writeup hit Hacker News on October 25, where it drew 272 points and a debate over the gap between the math and a benchmark.