DeepSeek Bets on Huawei Training Chips, Liang Wenfeng Says "Must Succeed"

Beating AI News Flash, citing sources from *The Information*, DeepSeek CEO Liang Wenfeng stated at the latest closed-door investor meeting that using Huawei and other domestic chips to train large models is one of the company's biggest bets, emphasizing that this endeavor "must succeed." DeepSeek expects to receive a new batch of Huawei chips suitable for training as early as the fourth quarter of this year, though specific models and quantities have not yet been disclosed. Liang Wenfeng believes Huawei chips could catch up to Nvidia within just a few years and hopes DeepSeek can adapt to hardware beyond Nvidia in advance. Currently, DeepSeek is still using Nvidia chips to train models. Back in May, Liang Wenfeng estimated that training top-tier large models would require approximately 50,000 Nvidia GB300 chips, whereas switching to Huawei Ascend 950 chips would require around 200,000 units. At that time, the biggest challenge was Huawei's insufficient supply. [Original Link]