Guangyu Xinchen has raised a $300 million Series A, according to a Dealroom News post on X, which described it as China’s largest disclosed primary-market round that week. The Shanghai company was founded in 2024 and builds edge AI chips designed to run large models on-device rather than in the cloud.
The company focuses on deploying large models on end devices to address latency, privacy and cost issues associated with cloud-based models. Its founder, Dr. Zhou Qiang, studied microelectronics at Fudan University and previously worked at VeriSilicon, AMD, Enflame Technology and Moore Threads.
Reporting cited by Kechuangban Daily said on September 10 that Guangyu Xinchen had completed a Series A and had raised roughly RMB 2 billion, or about $300 million, cumulatively across multiple rounds over the preceding six months. Available reporting does not specify a single lead investor. Named investors include Longteng Capital, China Life Capital, Dahua Venture Capital, Shandong Development Group, Qianhai Infrastructure Fund, Pudao Xinye, BAIC Investment and Kaitai Capital.
Memory chipmaker GigaDevice Semiconductor was a founding investor, subscribing RMB 1.5 million for a 15% stake on July 26, 2024, the day the company was established, according to corporate registry data cited by Eastmoney.
Guangyu Xinchen calls its architecture EdgeAIon. The company said it combines 3D stacking with compute-storage integration, and that its first chip has taped out and is expected to reach mass production by the end of 2026. It claimed a 10x improvement in computing efficiency with lower power consumption.
Other Chinese chipmakers are pursuing similar 3D-stacked near-memory architectures in response to AI memory-bandwidth bottlenecks amid constrained access to advanced foreign process nodes and HBM, including Tsingway, Suanmiao Technology and Unisplendour.
At the 2026 World Artificial Intelligence Conference, Guangyu Xinchen unveiled its TC1000 series, claiming a single chip can deliver 300 tokens per second on a 3B-parameter model and support on-device operation of models with up to 35B parameters. As of the TC1000’s unveiling, the company had not publicly disclosed wafer volumes, cumulative shipment numbers, paying-customer counts or order values.
