Amid continued tightening of overseas export controls on high-end AI chips,
Chinese large-model developer DeepSeek is accelerating a move to domestic
compute. Industry sources say DeepSeek’s next-generation model, DeepSeek V5, is
planned to use Huawei’s Ascend 950 as its core training hardware, with
deployment targeted next year. DeepSeek and Ascend core R&D teams are now
co-located in Hangzhou and completed extensive low-level operator adaptation and
custom development during V4; Ascend was used for inference only in the prior
generation, while V5 would be the first large model trained natively on Ascend
950.
