Etched

Etched designs specialised hardware for running frontier AI models. The company describes its category as frontier inference clusters: chips, racks, software and manufacturing methods co-designed so that large models can run with high throughput, low latency, good cost and strong power efficiency, for both prefill and decode workloads. Its inference silicon, publicly known as Sohu, is an ASIC built around the transformer architecture. Etched is headquartered in San Jose, California, and does not make game or video models.

The product is a rack-scale inference system rather than a standalone card. Two design techniques anchor it: Low Voltage Inference, which runs the chip's math blocks at much lower voltage to raise sustained compute density without thermal throttling, and Cluster Scale Memory, a shared low-latency memory pool across the scale-up domain that uses a proprietary interconnect and a hybrid of HBM and SRAM. Targeted workloads include very large mixture-of-experts models, long context and agentic use. First racks have shipped to a quantitative trading firm that is its first named customer, and the company runs a Taiwan factory plus a data center, test house and prototyping lab at its San Jose site to speed production.

Etched sells inference systems to AI labs, cloud providers, hyperscalers and other organisations that run large models, and is vertically integrated across chip design, packaging, thermal and system engineering to scale production toward gigawatt-class capacity.

keywords: AI inference chips, transformer ASIC, Sohu, inference clusters, AI hardware, rack-scale systems, low voltage inference, cluster scale memory, HBM, mixture of experts, AI accelerators, semiconductors, data center, TSMC, San Jose