On November 3, the Chinese financial outlet STAR Market Daily said DeepSeek plans to pre-train V5, its next model, on Huawei's Ascend chips. Its sources are several people in the supply chain. Pre-training is the first big stage, where a model learns from a huge body of data. The report says DeepSeek has so far used Ascend mostly for inference, meaning running a finished model. Bloomberg wrote in early September that DeepSeek's training had run on Nvidia hardware up to then.
The original article is no longer online, and the source does not say why. Sina Finance carried the report. The plan is unconfirmed. DeepSeek's V4 technical report states its data size, 32 to 33 trillion tokens, but leaves the chips unnamed. Hello China Tech, which recapped the story, says V5's technical report would be the proof to wait for.
Training is a heavier hardware load than inference. It runs for a long stretch across many chips that must swap results constantly, while inference just serves one finished model. So the claim that matters here is about training. A buyer reading a vendor's chip claim should ask which of the two jobs it covers. If V5 did train on Ascend, that would speak to the harder job.