DeepSeek plans to utilize Huawei Ascend 950DT chips to train its upcoming V5 model. The company intends to deploy at least 160,000 of these accelerators at a new data center located in Ulanqab, Inner Mongolia.
The lab aims to begin pre-training workloads on the new hardware by late 2026. This timeline aligns closely with Huawei's scheduled commercial release of the Ascend 950DT, which is expected in the fourth quarter of 2026.
To address previous stability and performance issues encountered with earlier Ascend generations, Huawei engineers are working directly with the DeepSeek team. This collaboration aims to optimize chip design for specific model requirements.
The Ascend 950DT features 144 GB of HBM-style memory and provides approximately 4 TB/s of bandwidth. DeepSeek has also developed open-source software libraries that reportedly achieve over 95% of the chip's peak performance.
DeepSeek founder Liang Wenfeng noted that four Ascend 950 chips are required to match the functionality of a single Nvidia GB300 GPU. This shift toward domestic hardware is driven by US export restrictions on advanced AI chips.
