Deepseek intends to install at least 160,000 Huawei Ascend-950DT chips in a new facility in Inner Mongolia. This deployment would create the largest known cluster of Huawei processors. The hardware will handle inference tasks only, while training workloads remain on Nvidia equipment. Bloomberg reports that full delivery is unlikely within a year due to production limits and memory shortages. The project aligns with government efforts to reduce reliance on foreign technology.
The initiative highlights a specific bottleneck in China’s artificial intelligence infrastructure. Domestic memory manufacturers like CXMT are only beginning small batches of HBM3E production. These firms lag significantly behind Samsung, SK Hynix, and Micron, which are already mass-producing HBM4. The memory gap restricts the speed at which domestic chips can scale. Without high-speed memory, the potential of the processor cluster remains limited regardless of quantity.
* CXMT is three to five years behind global leaders in HBM4 production
* Nvidia hardware continues to handle all training workloads for Deepseek
* Delivery of the full order faces delays beyond one year
