Etched AI

Etched AI is a semiconductor company designing ASIC chips purpose-built exclusively for transformer-based AI inference, achieving dramatically higher token generation throughput per watt than GPU-based inference by eliminating the general-purpose computation overhead of GPU architectures. Its Sohu chip is designed to run large language models faster and more energy-efficiently than NVIDIA H100 GPUs for inference workloads, targeting the growing market for cost-efficient LLM serving infrastructure. Etched AI represents a new category of specialized AI inference hardware optimized for the transformer architecture that dominates modern AI.