Chinese artificial intelligence lab DeepSeek plans to deploy more domestically produced chips for training AI models, founder Liang Wenfeng said, with Huawei expected to deliver chips as early as the fourth quarter of 2026 or by the first quarter of next year, according to The Information.
Liang described the use of Huawei and other domestically produced chips for AI training as one of the company's biggest strategic bets, adding that the move must succeed.
He also assessed that Huawei's AI chip performance will catch up with that of Nvidia within a few years, so the company decided to adopt hardware outside the Nvidia-centric ecosystem.
Meanwhile, the leading AI developer is training a model with 2 trillion parameters and plans to develop an 8-trillion-parameter model in the future, Liang told investors.
The market estimated the 2-trillion-parameter model to be the upcoming DeepSeek V4.1 Pro, which could be unveiled in the middle or at the end of October.
Training a 2-trillion-parameter model needs 50 trillion to 60 trillion tokens, equivalent to the computing power of 30,000 Nvidia H100/H200 chips, or 15,000 B200 chips, according to mainland media's calculation.
If the training is powered by Huawei Ascend 910C or Ascend 950 series, the number of required chips will be around 60,000 and 30,000.
