Image: Huawei
DeepSeek plans to pretrain its next big AI model, DeepSeek V5, on Huawei's Ascend chips, according to a report from the Chinese financial outlet Star Market Daily. Reporter Huang Xinyi wrote on November 3 that several people in the chip supply chain shared the plan. DeepSeek and Huawei had not commented publicly at the time of this writing.
Pretraining is the first and most expensive stage of building a model. It is when the model learns from huge amounts of text, and it usually needs thousands of powerful chips running for weeks. Until now, Star Market Daily says, DeepSeek has used Ascend chips mainly for inference, the work of answering user questions with a model that is already trained.
The story also ran on the WeChat account of Cailian Press, the news group behind Star Market Daily, and was picked up by Phoenix News. It spread quickly on X after users shared screenshots of the Chinese text.
"DeepSeek plans to pretrain its next-generation large AI model, DeepSeek V5, using Huawei Ascend chips."
— Jukan (@jukan05) November 3, 2026
👀
The report traces the history between the two companies. Citing the book Ascend Rising by Fang Xingdong of Zhejiang University, it says Huawei and DeepSeek set up more than 3,000 Ascend chips over the 2025 Lunar New Year holiday to run DeepSeek V3 and R1 on Huawei Cloud.
After DeepSeek V4 came out in April 2026, Huawei said its Ascend SuperPoD systems could fully support the V4 models, the report says. It adds that DeepSeek's own API page said V4 Pro capacity was limited by a shortage of high end chips and that prices should drop once Ascend 950 SuperPoD systems ship in volume.
Image: Huawei
On September 30, DeepSeek open sourced a set of software tools built for Ascend hardware. They match tools the company had earlier released for Nvidia chips. Star Market Daily says this shows the two companies working together deep in the software that both training and inference rely on.
The outlet adds that Chinese chips are still used mostly for inference. It points to a recent project in Shenzhen that used Ascend 910C chips to fine tune the full 1.6 trillion parameter DeepSeek V4 Pro, but notes that work did not include pretraining. If V5 is pretrained on Ascend, the report says, it would be the first time a top Chinese model has shown that this can be done on Chinese chips.
Other outlets added details that Star Market Daily did not include. TrendForce, citing a Chinese outlet it calls Caiwen, reported that V5 would be trained mainly on Ascend 950 chips and could arrive in 2027. It said the model's size, features and launch date are still undecided, and that Huawei holds no stake in DeepSeek.
TrendForce also said, citing Star Market Daily, that Huawei now expects its newer Ascend 960DT chip to be ready in the first quarter of 2027, three quarters earlier than first planned. Zekun, an AI researcher on X, flagged the same chip timeline and said the V5 news was not much of a surprise for people following DeepSeek.
Some on X read the report as part of China's push to rely less on American chips.
This represents the next significant step that China - and by extension, Huawei - is attempting to take: achieving as much independence as possible from the U.S. and from NVIDIA chips.
— Chubby (@kimmonismus) November 3, 2026
That is why, for DeepSeek v5, they have opted to conduct the training almost entirely on Huawei’s own chips.
The rationale behind this is that by breaking away from U.S. chips to the greatest extent possible and keeping development within the domestic market, China retains its leverage - specifically regarding rare earth elements - over the West, without having to rely on Western goodwill for the supply of chips.
That view is commentary, not part of the original report. Names linked to V5 were spotted in DeepSeek code in October, but the company has not announced the model.
Comments