Custom ASIC per model is going to be insane. 20k tokens per second achieved on llama 8b already. Imagine codex 5.3 but 100x faster.