Multiple media reported that DeepSeek's next-generation V5 model is planned to be pretrained on Huawei Ascend chips, moving from using domestic accelerators mainly for inference to pretraining - a key step for "domestic chips training domestic models."
The report comes as global AI chip controls tighten; Nvidia has reportedly guided cloud customers that at least 51% of new compute services must be sold to non-Chinese firms, and Oracle canceled a major chip-leasing contract with Tencent. Analysts say domestic compute substitution is no longer optional but a matter of timeline.
Source: Sina Tech / Weibo Tech Daily
