DeepSeek plans to pretrain V5 on 160,000 Huawei Ascend chips

DeepSeek has optimized its software for non-NVIDIA hardware before, but until now Ascend has mostly handled inference. Pretraining V5 on 160,000 Ascend 950DT chips would be the largest known frontier training run on Chinese silicon.

The significance is strategic. As analyst @kyleichan noted on X, a strong V5 pretrained entirely on Ascend would be a major milestone for China since it was cut off from TSMC in 2020. It would show that export controls slowed China's frontier labs without stopping them. The reported $15.9B raise at a $90B valuation gives DeepSeek the capital for a cluster of this size. DeepSeek's current V4.1-Flash, the first model in a new architecture family, already performs strongly, with a reported 90.6% on Terminal-Bench 2.1.

There is also a link to Mistral CEO Arthur Mensch's warning this week that frontier AI is being treated as weapons technology. A self-sufficient Chinese training stack makes that framing more likely.

Skeptics point to Ascend's software maturity, interconnect performance at scale, and yields. All three have constrained Huawei before. The training run is planned for 2027, so for now this is a plan, not a result.