DeepSeek V5 Planned for Pre-training on Huawei Ascend Chips
Industry reports indicate that DeepSeek plans to use Huawei Ascend chips for the pre-training of its next-generation DeepSeek V5 model, marking a significant step beyond previous inference-only deployments. DeepSeek has also open-sourced infrastructure components tailored for the Ascend platform, such as TileLang and DeepEP, to align with the domestic hardware stack. This development represents a major shift toward domestic Chinese hardware for heavy-duty AI model pre-training, potentially bypassing US export restrictions on advanced GPUs. Successfully utilizing Huawei Ascend chips for full pre-training could dramatically alter the competitive landscape for foundational AI infrastructure. While previous deployments focused on inference or post-training of large models like DeepSeek-V4-Pro, transitioning the entire pre-training workload to Ascend requires migrating complex software ecosystems and networking solutions like UBL128. DeepSeek anticipates further cost reductions and performance gains once upcoming Ascend super-node series become mass-market available.
## BACKGROUND
Huawei Ascend is a series of AI accelerators developed by HiSilicon, a subsidiary of Huawei, designed to handle heavy machine learning workloads amid US trade restrictions. LLM pre-training is the foundational and most computationally intensive phase of artificial intelligence development, where a model learns language patterns from massive text datasets.