Etched designs Sohu, a transformer-specific ASIC that the company says delivers over 500,000 tokens per second on Llama 70B, outperforming general-purpose GPUs for AI inference.
Etched.ai, Inc. was founded in 2022 by Harvard dropouts Gavin Uberti, Chris Zhu and Robert Wachen on a single conviction: that transformer architectures will dominate artificial intelligence. The company designs transformer-specific ASIC chips, betting that specialised silicon, rather than general-purpose GPUs, will define the economics of AI inference. That bet puts it in direct competition with entrenched incumbents such as NVIDIA.
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. According to Etched, such performance makes previously impractical applications viable - real-time video generation and reasoning agents with extremely deep chains of thought among them. Technical work spans ASIC design, inference acceleration, high-performance computing and memory and interface technologies, the latter developed in partnership with Rambus.
Etched raised $120m at a valuation of $5 billion. For engineers, the appeal is the difficulty of the problem: custom silicon for a single dominant architecture demands excellence across the full stack, from chip design to systems and software, and rewards those willing to work at a rapid pace against well-resourced incumbents.
