DeepSeek CEO Liang Wenfeng (량원펑) has identified training AI models with domestically made chips as the top priority, a report said.
According to a recent report by The Information, Liang said at a private investor meeting that Huawei could start supplying DeepSeek with training chips as early as the fourth quarter of this year.
He called training models with Chinese-made chips such as Huawei’s one of DeepSeek’s biggest bets and said, "It must succeed."
DeepSeek is set to receive Huawei chips for training its latest model in the fourth quarter of this year or the first quarter of next year. Liang did not mention how many chips Huawei would supply.
Training AI models requires more powerful chips than running them. The Information said DeepSeek’s increased use of Huawei chips for training shows cooperation between the two companies to move beyond U.S. government chip export controls.
Liang believes Huawei chips could become as good as Nvidia’s within a few years. He said the company therefore needs to adapt first to semiconductors other than Nvidia’s.
DeepSeek is currently training a model with 2 trillion parameters. Its current flagship model, V4, has 1.4 trillion parameters. DeepSeek also plans to develop a model with 8 trillion parameters.
Chi-gyu Hwang
delight@d-today.co.kr