Etched.ai, Inc.
Etched builds transformer-specific ASIC chips, such as Sohu, designed to accelerate AI model inference at lower cost and latency than GPUs.
Etched.ai, Inc. designs application-specific integrated circuits built exclusively for transformer models. Founded in 2022 by Harvard dropouts Gavin Uberti, Chris Zhu and Robert Wachen, the company is a bet on a single technical conviction: that transformers will dominate artificial intelligence, and that chips hard-wired for them will outperform general-purpose alternatives. That premise puts Etched in direct competition with NVIDIA and the wider market for graphics processing units.
The company's flagship product, Sohu, is described as the world's first transformer-specific ASIC. Etched claims the chip delivers an order of magnitude more throughput and lower latency than NVIDIA's B200, and sustains more than 500,000 tokens per second on Llama 70B. The company positions Sohu as an enabler of workloads that are impractical on GPUs, including real-time video generation and extremely deep chain-of-thought reasoning agents, serving enterprise AI workloads and large language model deployments.
Etched raised $120m at a valuation of $5 billion. Its technical scope spans ASIC design, inference acceleration, transformer optimisation, and semiconductor architecture, with memory and interface technologies supplied through partners such as Rambus. The company's stated approach favours specialised single-model silicon over general-purpose computing.