DeepSeek reportedly plans to train its next model on Huawei Ascend chips

What was reported

As Informatization Observer reported, word emerged in early November that DeepSeek's next large model, V5, is planned to be pre-trained on Huawei's Ascend chips. Pre-training is the first, most expensive stage of building a model, where it learns from huge amounts of data.

Until now Ascend's role in DeepSeek's work was limited to inference, meaning running a finished model to answer users. V5 would be the first DeepSeek model trained natively on Ascend, aimed at the latest Ascend chip and a rollout next year. This comes from reports, not an official announcement.

The groundwork already laid

The article says the two sides did a lot of low-level adaptation during the V4 generation. That includes custom operators, the small math routines a model runs thousands of times, tuned for DeepSeek's designs.

In late September, DeepSeek also open-sourced software for Ascend that mirrors what it already offered for Nvidia: a compiler tool called TileLang, a computing library and a library for chips to talk to each other. DeepSeek said each TileLang operator used in its training has a high-performance Ascend counterpart. Without that matching work, the V5 plan would not be workable.

A wider shift in Chinese chips

The piece argues the question has moved from whether domestic chips can be used to whether they are worth using. On the serving side, firms such as MetaX and Iluvatar CoreX have launched GPUs with 144GB of memory, above the 141GB on Nvidia's H200. Memory matters because it caps how large a model and how much context a chip can hold.

Citing IDC, it says Nvidia's share of China's data-center AI chip market fell in the first half of 2026 while Huawei moved close behind. On the training side, it points to Zhipu AI's image model, trained fully on Ascend, and Moore Threads' cluster runs as earlier signs.

What could still go wrong

The author does not claim a clean win. Chinese suppliers still lag in high-bandwidth memory, advanced chip-design software and lithography machines, which print circuits onto chips. Nvidia's CUDA software has over a decade of developer habits behind it.

To compensate for weaker single chips, Chinese makers are linking many chips into supernodes, rack-scale systems sold as one unit. The article calls whether V5 trains successfully on schedule the first real exam for domestic training chips.