Etched

Overview

Etched is a United States AI infrastructure company developing custom silicon and rack-scale systems for AI inference. Its central product, Sohu, is an application-specific accelerator designed around transformer workloads rather than the broader flexibility of GPUs. The company is notable for combining ASIC design with memory, networking, software, rack integration, and manufacturing capabilities, targeting large-scale model serving and emphasizing throughput, latency, power efficiency, and cost.

History

Etched was founded in 2022 by Harvard graduates Gavin Uberti and Chris Zhu, with Rob Wachen and semiconductor executive Mark Ross also associated with the founding team. The company emerged from stealth in 2024 with Sohu, an ASIC designed specifically for transformer-model inference. It later expanded from chip design into rack-scale inference systems integrating compute, memory, networking, software, and manufacturing. The company reports having raised $800 million across four financings and operates a Taiwan factory plus San Jose-area prototyping and production facilities.

Product Lines

Manufacturing

Etched designs its chips and systems in-house but uses external semiconductor fabrication. Sohu is reported to be manufactured by TSMC on its N4P process, with HBM3E memory, while Etched operates a factory in Taiwan and a San Jose-area facility for prototyping, new-product introduction, and surface-mount manufacturing.

Notable Products

• Sohu - Transformer-specialized ASIC for high-throughput AI inference, reported to use TSMC N4P fabrication and 144 GB of HBM3E memory.
• Frontier Inference Clusters - Rack-scale inference infrastructure combining Sohu accelerators, memory, networking, and software for serving frontier AI models.
• Sohu Software Platform - Etched’s software and systems stack for deploying and operating transformer-focused inference workloads across its accelerator clusters.

Reputation

Etched is viewed by semiconductor and AI infrastructure observers as a specialized challenger to GPU-based inference, with attention focused on its transformer-specific architecture and vertically integrated rack systems. Its reported engineering team includes personnel from NVIDIA, Google TPU, Broadcom, SK Hynix, TSMC, and quantitative-trading firms. The main criticism of its approach is the tradeoff between specialization and flexibility: custom transformer silicon may offer efficiency for targeted workloads but can be less adaptable than GPUs as model architectures and customer requirements change. Public information about commercial deployments, pricing, and long-term production performance remains limited.