Industry Sources Say DeepSeek Plans V5 Training on Huawei Ascend Chips

Multiple people in the chip supply chain say DeepSeek plans to train its next artificial intelligence model, DeepSeek V5, on Huawei Ascend chips, which would expand their role beyond inference.

Ascend processors have mainly been used for inference, when an already trained model produces responses. Using the chips to train V5 would put them to work in building the model itself.

DeepSeek's earlier attempt to train its R2 model on Huawei chips ran into stability, communication and software problems. It completed that training on NVIDIA hardware.

In late September, DeepSeek released open-source components for Huawei's Ascend platform, including TileLang compiler tools, computing libraries and distributed communication software. Those components match tools it had released earlier for NVIDIA's platform.