Etched is building specialized AI inference hardware around a single thesis: transformer architectures are important enough to justify a purpose-built chip. Rather than serving robotics directly, the company sits in the embodied-AI stack as a compute enabler for low-latency multimodal inference.
In October 2024, Etched and Decart released Oasis, an interactive world model designed to be highly optimized for Sohu. The demo became Etched’s clearest public example of co-designing models and products around its hardware roadmap.
Sohu And The End Of General‑Purpose AI Chips
A look at Etched’s transformer‑only ASIC and what it means for data centers, humanoid robots, and the next phase of AI infrastructure.
Jessica Alvarez
Who Owns The Brains And Minds Of Humanoid Robotics
How next generation chips and foundation models will quietly decide which humanoids win.
Jessica AlvarezEtched is a San Jose semiconductor startup building Sohu, a transformer-specific ASIC for AI inference. Founded in 2022 by Harvard dropouts and systems engineers, the company is designing rack-scale hardware optimized for dense models, sparse MoEs, diffusion workloads, and real-time generative applications.
Etched was founded in 2022 around the view that transformers would dominate AI workloads and that the architecture was stable enough to hardwire into a dedicated inference chip. The company began building Sohu as a rack-scale system optimized around that assumption.
In June 2024, Etched described Sohu as the world’s first specialized ASIC for transformers and argued that custom silicon would be required to deliver a major leap in tokens per dollar for large-scale production inference.
In October 2024, Etched and Decart released Oasis, an interactive world model designed to be highly optimized for Sohu. The demo became Etched’s clearest public example of co-designing models and products around its hardware roadmap.
