At the 2026 China Computing Power Conference, held in Langfang, Hebei, from September 11 to 13, a coalition of Chinese tech organizations unveiled the country's first heterogeneous hybrid inference system for large language models that pairs domestic GPUs with brain-inspired chips. The system, developed by China Mobile Cloud (移动云), the China Electronics Technology Nanyang Research Institute, Lynxi Technology, Shanghai Tianshu Zhixin (Dmetas), Tsinghua University, and Peking University, aims to break the performance and cost barriers of traditional single-GPU inference.
How It Works: Divide and Conquer
The core principle, described as "divide and conquer, synergy for efficiency," involves splitting a large model into PD (prefill-decode) and AF (attention-FFN) components, then assigning each to the chip best suited for the task. Attention computation is handled by domestic GPUs, leveraging their general-purpose computing strengths, while the FFN (mixture-of-experts) latency-sensitive modules are offloaded to brain-inspired chips, which excel in compute-in-memory and feature large on-chip SRAM. A custom model compiler, high-speed interconnect protocol, and unified inference engine coordinate task decomposition, scheduling, and result aggregation across the two chip types.
According to IThome, tests with DeepSeek V4 showed the system delivers more than double the cost-performance ratio of comparable domestic GPU-only clusters, and reduces business operating costs by over 40%. The system is designed for high-realtime scenarios such as token factories, AI code generation, multi-agent collaboration, and intelligent manufacturing, and offers a fully domestic inference foundation for security-sensitive industries like finance, security, and telecommunications.
Implications for the Industry
This development is significant for China's AI infrastructure push, as it reduces reliance on high-end GPUs from abroad and optimizes the use of domestic silicon. For readers, it signals a shift toward more cost-effective and efficient AI inference, which could lower the barrier for deploying large models in various applications. The collaboration between major state-backed entities and universities underscores the strategic importance of self-reliant AI computing in China.
The announcement was made alongside other tech news, including China Mobile's AI trusted computing initiative and NASA-IBM's lunar research model, but this heterogeneous inference system stands out as a domestic breakthrough in AI hardware integration.
