Definition

Sohu is etched’s transformer-specific inference ASIC that hardcodes LLM attention mechanisms into silicon, optimized exclusively for transformer-based model inference (not training).

Key Points

  • 2026-07-23: Parent company etched raises 10.3B to scale inference systems (2026-07-23-etched-300m-series-c-10-3b-valuation)

  • Fabricated on TSMC N4P; first-pass silicon (A0 tapeout) validated

  • Rack-scale systems running DeepSeek, Qwen, Mamba, Llama

  • Company claims 8-chip server replaces 160 H100s on Llama-70B (unverified independently)

  • Architecture concentration risk: obsolete if industry shifts away from transformers

Sources