DeepSeek delays new models due to problems with Huawei chips

Chinese artificial intelligence startup DeepSeek is delaying the release of new models due to difficulties in training them caused by ongoing technical problems with Huawei's AI accelerators, Reuters reports, citing the FT.

DeepSeek uses Huawei's Ascend to train its large language models, but due to ongoing technical issues, the Chinese startup has started using NVIDIA chips for training. At the same time, it was the problems with Ascend that forced the company to postpone the release of the R2 model, which was supposed to take place back in May.

The Ascend chips will instead be used to enable models to generate responses to user queries, a process that typically requires less computing power than training.

The DeepSeek situation highlights the general difficulties Chinese companies have with using domestic chips. Despite the Chinese government actively promoting Huawei accelerators, most AI developers continue to turn to NVIDIA.

At the same time, NVIDIA is not an ideal option either, as the world's most valuable company faces pressure from both the US and China. While the Chinese government is concerned about potential security risks from using Western chips, the US government has sought to restrict exports of leading chips to China.

Despite this, US President Donald Trump recently lifted restrictions on the export of H20 AI chips to China after meeting with NVIDIA CEO Jensen Huang. It later emerged that in return for Trump's move, NVIDIA, and now AMD, must pay the US government 15% of the revenue from chips sold to China.

DeepSeek gained global popularity in early 2025 with the release of its flagship model, the DeepSeek-R1. At the time, it was claimed to be as good as Western counterparts, but much cheaper to train. DeepSeek-R2 was supposed to be the Chinese startup's next flagship model, but Ascend's technical problems have so far prevented this from happening.