DeepSeek to Train V5 on Huawei Chips Instead of NVIDIA

According to Chinese outlets Caiwen News and Star Market Daily (as reported by TrendForce), DeepSeek plans to train its next model, V5, primarily on Huawei Ascend 950 accelerators, with an expected release in 2027.

Compiled by Sergey KostenchukPublished 2026-11-03Updated 2026-11-03

2026-11-03 Coding NVIDIA

💻 DeepSeek to Train V5 on Huawei Chips Instead of NVIDIA

According to Chinese media outlets Caiwen News and Star Market Daily (as reported by TrendForce), the V5 model will be trained primarily on Huawei Ascend 950 accelerators, with a release expected in 2027. Previously, Ascend was used mainly for inference, but pre-training is now being migrated to it as well. A cluster in Ulanqab consisting of at least 160,000 Ascend 950DT chips has been announced.

🌍 Huawei will gain a flagship client and confirmation that the Ascend stack is ready for model training, while NVIDIA loses its largest Chinese buyer.

👤 If the plan succeeds, DeepSeek will be able to release advanced models without a critical dependence on US export restrictions. The open Ascend stack—TileLang and libraries—is already available for testing.

Source 1: https://www.trendforce.com/news/2026/11/03/news-deepseek-v5-could-complete-training-on-huawei-ascend-950-in-2027-moving-beyond-inference/ Source 2: https://www.chaincatcher.com/en/article/2301282

Sources