Chinese AI research lab Deepseek plans to deploy a massive cluster featuring at least 160,000 Huawei Ascend-950DT processors at a data center in Inner Mongolia. The proposed installation would represent the largest known cluster of Huawei AI hardware to date. According to report details, the Huawei processors will be used strictly for inference workloads, while Deepseek continues to rely on Nvidia hardware to execute heavier model training.
The deployment faces supply chain constraints, as Huawei is expected to require over a year to fulfill the processor order due to production bottlenecks and memory shortages. While domestic memory supplier CXMT has begun producing small volumes of HBM3E high-speed memory, its technology remains three to five years behind global leaders Samsung, SK Hynix, and Micron, which are producing HBM4.
The initiative reflects broader efforts by the Chinese government to advance domestic hardware independence amid export controls. However, the operational split between Huawei inference and Nvidia training highlights China’s ongoing challenges in replacing Western chips for frontier model development.
Why it matters
China’s leading AI labs are shifting inference infrastructure to domestic chips while remaining tied to Nvidia for frontier training.
Memory bottlenecks remain the primary operational constraint delaying large-scale deployment of Huawei AI hardware.
Enterprise inference costs in China could drop as domestic chip clusters achieve massive regional scale.
Source: the-decoder.com
