DeepSeek bets on Huawei training chips, Liang Wenfeng says 'must succeed'

Beating AI News Flash: The Information, citing people familiar with the matter, reported that DeepSeek CEO Liang Wenfeng, at a recent closed-door investor meeting, listed training large models on domestic chips such as Huawei's as one of the company's biggest bets, and said this matter "must succeed."

DeepSeek expects to receive a new batch of Huawei chips usable for training as early as the fourth quarter of this year, with the specific models and quantities not yet disclosed. Liang Wenfeng believes that Huawei chips are expected to catch up with Nvidia in just a few years, and hopes DeepSeek will adapt in advance to hardware beyond Nvidia.

DeepSeek is still using Nvidia chips to train its models. In May this year, Liang Wenfeng estimated that training a top-tier large model requires about 50,000 Nvidia GB300 chips, while switching to Huawei Ascend 950 would require about 200,000. At the time, the biggest problem was Huawei's insufficient supply.