Etched.ai designs transformer-specific ASIC chips optimised for AI inference workloads. The company's product, Sohu, is engineered to deliver substantially higher throughput and lower latency than general-purpose accelerators for running large language models and other transformer-based applications. The chip reportedly achieves over 500,000 tokens per second on Llama 70B, with performance characteristics claimed to exceed leading GPU alternatives by an order of magnitude.
The company was founded in 2022 by Gavin Uberti, Chris Zhu, and Robert Wachen, and has raised $120 million in funding at a reported valuation of $5 billion. Its technical approach prioritises specialisation in transformer inference rather than general-purpose computing. The firm has partnered with Rambus to integrate memory and interface technologies into its designs.
Etched.ai targets use cases where transformer performance and cost efficiency are critical, including real-time video generation and complex reasoning tasks. The company positions itself as a direct alternative to incumbent chip manufacturers for AI inference applications, focusing on the inference phase of model deployment where latency and throughput directly affect product viability and operating costs.
