A startup, founded in 2025 by a former GPU architect from Nvidia, is developing a processor from scratch intended for physical artificial intelligence—that is, for robots and machines capable of perceiving, reasoning, and acting in the real world. The company's first product, named Y10, aims to reduce the entire perception, output, and action cycle to less than one millisecond, as reported by Xu in an interview with the Chinese startup publication Cyzone on October 10.
Xu asserts that most modern robots use central processing units (CPUs), graphics processing units (GPUs), and autonomous driving chips. He emphasized that GPUs are designed for high throughput when working with large language models, not for the continuous cycle where a robot senses its environment, performs an action, and then corrects it at a high frequency based on feedback. For physical AI, the overall time of the entire cycle, as well as energy and cost expenditure, is more important than just peak TOPS performance.
According to Xu, the chip architecture determines about 70% of the parallel processor's utility, while software accounts for the remaining 30%. Instead of using a specialized ASIC, Y10 is a general-purpose processor managed by a customizable and highly programmable set of instructions, allowing it to adapt as robot models change. According to early company materials, the design is based on an architecture they have named SomaArch. Xu added that members of the core team had experience working on several generations of Nvidia GPUs and participated in the transition of large chips, including 6-nanometer level components, from design to mass production, and the instruction set is designed to simplify migration for engineers.
The company reports that it is conducting testing with leading Chinese robotics manufacturers. Based on simulations, which are proprietary company data and not silicon measurements, its solutions demonstrate five times better overall performance while reducing costs by 80% compared to current x86 plus Orin configurations. Xu clarified that the team's simulations usually align with final results within 3%. Chip development is halfway complete, and a version of the computing platform based on FPGA is planned for release by the end of this year so that customers can test the actual hardware.
The initial target is industrial robotic manipulators, not humanoids. Quoting a partner, Xu noted that existing manipulators perform only about 2% of manual operations on real production lines. Furthermore, the company is considering creating small world model servers based on its chips. In Xu's opinion, the current computational power of robots is roughly equivalent to the level of a 386 processor from 1985.