DeepSeek to Train Its Next AI Model on 160,000 Huawei Chips
Chinese AI company DeepSeek has confirmed plans to train its next-generation V5 model on a fleet of 160,000 Huawei Ascend 950DT chips. The chips will be deployed at a new gigawatt-scale data center under construction in Ulanqab, Inner Mongolia, according to a report published on November 4, 2026.
The move marks a major strategic shift for DeepSeek. The company plans to begin pre-training workloads on the domestically manufactured chips as early as late 2026. The Ulanqab facility was originally intended for inference tasks, but will first be used to assess the hardware before expanding to full pre-training operations.
Each Huawei Ascend 950DT chip features:
- 144 GB of high-bandwidth memory
- Approximately 4 TB/s of memory bandwidth
The specifications are designed to support large-scale model training. The chip order is valued at an estimated $2.56 billion and is widely viewed as a landmark endorsement of Huawei’s domestic AI chip ambitions. It also comes amid U.S. semiconductor export restrictions that have limited Chinese labs’ access to Nvidia’s latest hardware.
Analysts at Bernstein forecast that Nvidia’s share of China’s AI chip market could fall from roughly 40% to just 8% by the end of 2026, with Huawei poised to capture a near-majority share.
Source: Dataconomy — DeepSeek Plans To Train V5 Using 160,000 Huawei Chips