Etched.ai designs transformer-specific ASIC chips, including Sohu, built to increase AI inference throughput and reduce latency versus general-purpose GPUs.
Etched.ai, Inc. builds specialised AI hardware: transformer-specific ASIC chips designed to improve inference performance and cost relative to general-purpose GPUs. Its product, Sohu, is described as the world's first transformer-specific ASIC. The company reports over 500,000 tokens per second on Llama 70B, and an order of magnitude more throughput and lower latency than NVIDIA's B200. The chips are intended to support workloads such as real-time video generation models and extremely deep chain-of-thought reasoning agents.
The company's technical work spans AI accelerator hardware, ASIC chip design, transformer model inference, inference optimisation, and memory and interface technology. Etched.ai takes a partnership-driven approach to components, working with Rambus on memory and interface technologies. Its stated position is a bet on single-model ASICs rather than general-purpose computing, founded on the conviction that transformers will dominate the AI landscape.
Founded in 2022 by Harvard dropouts Gavin Uberti, Chris Zhu and Robert Wachen, Etched.ai raised $120m in a recent round and is valued at $5 billion. It operates in the semiconductors and AI infrastructure sectors, competing with established players such as NVIDIA in data-centre inference hardware.
