Etched, founded in 2022 by three Harvard dropouts, builds specialized AI chips and rack-scale systems designed exclusively for running transformer-based AI models (the architecture behind GPT, Claude, and nearly every major LLM). Its flagship chip, Sohu, is an ASIC fabricated on TSMC's 4nm process that hardwires the transformer attention mechanism directly into silicon rather than running it as software on a programmable GPU. This extreme specialization lets Etched claim roughly 500,000 tokens per second on Llama 70B versus ~25,000 from a comparable H100 server. Etched sells complete "frontier inference clusters" — full rack systems including custom cooling and interconnects — not just chips. Buyers are hyperscalers, quant trading firms, and enterprises running inference at scale who need lower cost-per-token than Nvidia GPUs can deliver. The company co-designs chips, racks, software, and manufacturing under one roof, including an 80,000 sq ft Milpitas production facility. First racks shipped to Jane Street in mid-2026. Etched's bet is that transformer dominance is durable enough to justify permanently sacrificing chip flexibility for peak inference efficiency.