Dielet cuts open-model inference down to the measured bottleneck

Why it surfaced

The pitch has a clear engineering constraint: no cutover until outputs match and the clock improves on the same silicon. Its homepage even shows a sample H200 comparison, from 41.2 to 93.8 tokens per second with a claimed 100% output match.

Dielet tests kernels, batching, and decode layouts against recorded traffic, then serves an optimized open-model stack on the GPUs a customer already rents.