Frontier Inference Clusters
Etched offers frontier inference clusters optimized for running large-scale AI models, claiming significant performance and efficiency advantages over traditional GPU-based infrastructure for inference workloads.
Background
Etched is a hardware startup building a chip (the "Sohu" ASIC) purpose-built for running transformer neural networks — the architecture behind models like GPT, Claude, and Gemini. Unlike GPUs or general-purpose AI accelerators, Etched's chip is hardwired specifically for transformers, meaning it can't run other kinds of models but claims huge efficiency gains (speed, cost, power) for the dominant AI paradigm. The "Frontier Inference Clusters" page describes their plan to offer complete server racks of Sohu chips, aimed at companies that need to serve large transformer models in production. The central pitch: if transformers remain the dominant AI architecture, specialized hardware beats flexible hardware; if a new architecture emerges, Etched's rigidity is a liability.