DeepSeek plans to train V5 on 160,000 Huawei Ascend 950DT chips

Moving a frontier pretraining run onto 160,000 Ascend 950DT accelerators would be the largest public bet yet on Huawei silicon for training, not just inference. It reflects tightening US export controls: this same week, NVIDIA told cloud clients that at least 51% of their service sales must go to non-Chinese companies.

The technical risk is substantial. Ascend's software stack and interconnect have historically lagged CUDA for large distributed training, and efficiency at this scale is unproven. If it works, it would show that China's frontier labs can decouple from NVIDIA.

The community is cautious. Rumors about V5 continue, but nothing has shipped, and timelines remain unclear. Watch for any V5 preview and for disclosed training-efficiency figures.