DeepSeek Reportedly Plans to Train V5 on Huawei Ascend 950DT Chips

Chinese AI startup DeepSeek is reportedly preparing to train its next-generation model, DeepSeek V5, entirely on Huawei chips. The plan would be a major test of China's push to build a domestic AI stack without relying on Nvidia GPUs.

The reports first surfaced through Tencent Tech and STAR Market Daily, citing industry sources, and were picked up widely by international tech media on November 3, 2026.

What the Reports Say About DeepSeek V5

According to Tencent Tech, as cited by PANews, DeepSeek plans to begin large-scale pre-training on the Ascend 950DT as early as the end of this year. A complete pre-training run may require more than 10,000 chips running continuously for several weeks, depending on testing progress.

The same report says Huawei engineers have been stationed long-term at DeepSeek's offices to fold large-model training needs into next-generation chip design. Huawei is also said to plan to allocate most of its supply of the Ascend 960DT, due in the first quarter of 2027, to DeepSeek.

Wccftech, citing earlier reporting, says DeepSeek has ordered about 160,000 Ascend 950DT chips for a 1GW data center in Inner Mongolia, with deliveries ramping up in Q4 2026. That figure has not been independently verified.

Ascend 950DT and Atlas 950 SuperPoD Specs

Specifications as reported by Wccftech.

DeepSeek's Track Record on Ascend

The DeepSeek-Huawei relationship is not new. Around Lunar New Year 2025, Huawei reportedly deployed more than 3,000 Ascend chips to run DeepSeek V3 and R1 on Huawei Cloud. After DeepSeek V4 launched in April 2026, Huawei said its Ascend SuperPoD systems could fully support the V4 model family.

On September 30, 2026, DeepSeek open-sourced infrastructure components adapted for Ascend, including the TileLang compiler toolkit, computing libraries and distributed communication libraries, widely seen as groundwork for training on Huawei hardware.

Caveats: Still Only Reports

The plan has not been confirmed by DeepSeek or Huawei. Geopolitechs notes its viability will depend on training stability, cluster performance, software maturity and the resulting model's capabilities. Social media claims about V5's size, ranging from 2 trillion to around 8 trillion parameters, are also unverified.

If it works, DeepSeek V5 would show that a frontier-class model can be trained entirely on Chinese chips, something long doubted because of stability problems with earlier Ascend generations.

Conclusion

Training DeepSeek V5 on Huawei's Ascend 950DT, if the reports are accurate, would be a milestone for China's AI self-sufficiency. Until there is an official announcement, the news should be read as industry-source reporting.