China Mobile Cloud Unveils Domestic GPU and Neuromorphic AI System

China Mobile Cloud has launched a new inference system that splits AI model workloads between domestic graphics processors and neuromorphic chips. This approach offers enterprise buyers a way to lower operational costs while maintaining high throughput for large language models. The system targets organizations seeking to reduce dependency on traditional GPU-only clusters for inference tasks.

Enterprise buyers can reduce operational costs by over 40 percent

The platform runs the DeepSeek V4 Flash model using a heterogeneous architecture. Three server racks equipped with Shanghai TianShu ZhiXin GPUs handle the prefill and attention layers of the Transformer model. Beijing LingXi Technology provided the neuromorphic chips that process the feed-forward network layers. This division of labor allows each hardware type to operate in its optimal execution environment.

• Architecture: Domestic GPU + Neuromorphic Chip
• GPU Vendor: Shanghai TianShu ZhiXin (天数智芯)
• Neuromorphic Chip Vendor: Beijing LingXi Technology (北京灵汐科技)
• Model: DeepSeek V4 Flash
• Performance Improvement: Inference output and energy efficiency increased by more than 1x compared to pure GPU clusters

China Mobile Cloud reports that the system doubles both inference output and energy efficiency compared to pure GPU clusters of similar investment. The company states that business operation costs have decreased by more than 40 percent. The project has accumulated 15 authorized invention patents and six software copyrights to protect the hybrid design.

The system is currently in small-batch trial production in China. It was formally unveiled at the 2026 China Computing Power Conference held in Langfang, Hebei province. The specific model number of the neuromorphic chip remains unconfirmed in public reports, though the vendor is identified as LingXi Technology. The solution is available for immediate trial deployment within the Chinese market.