DeepSeek’s struggle with Huawei chips delays R2 release

Chinese AI start-up DeepSeek has delayed the release of its new R2 model after facing difficulties training it on Huawei’s Ascend chips, underscoring the limitations of China’s push to replace US technology.

Following encouragement from Chinese authorities, DeepSeek shifted from using Nvidia chips—considered the industry standard—to Huawei’s Ascend processors after launching its R1 model in January, according to three people familiar with the matter. However, persistent technical issues during the training phase forced the company to revert to Nvidia’s chips for training, while retaining Huawei’s chips for inference, Caliber.Az reports, citing foreign media.

The training setbacks became a key reason behind the delayed launch of R2, which was originally scheduled for May, said one person with knowledge of the situation. The delay has caused DeepSeek to fall behind rivals in China’s competitive AI race.

Training allows a model to learn from large datasets, while inference is the process of generating results, such as chatbot responses, from a trained model. Industry experts say Huawei’s chips lag behind Nvidia’s due to stability problems, slower inter-chip connectivity, and less mature software.

Despite support from a dedicated Huawei engineering team stationed at DeepSeek’s office, the start-up was unable to complete a successful training run on the Ascend hardware. “Just because we’re not seeing leading models trained on Huawei today doesn’t mean it won’t happen in the future. It’s a matter of time,” said Ritwik Gupta, an AI researcher at the University of California, Berkeley.

Gupta also noted that rival models, such as Alibaba’s Qwen3, have adopted and optimized DeepSeek’s innovations. “Models are commodities that can be easily swapped out,” he said. “A lot of developers are using Alibaba’s Qwen3, which is powerful and flexible.”

DeepSeek is still working with Huawei to make R2 compatible with Ascend chips for inference. Internally, founder Liang Wenfeng has expressed dissatisfaction with R2’s progress and is pushing for a more advanced version to maintain the company’s competitive edge.

The launch was also delayed due to longer-than-expected data labelling. Chinese media suggest the model may be released in the coming weeks. DeepSeek and Huawei declined to comment.

By Vafa Guliyeva