Beijing, China – Chinese artificial intelligence startup DeepSeek has delayed the release of its highly anticipated new AI model after encountering significant technical failures while trying to use Huawei’s domestically produced Ascend chips for training, the Financial Times reported on Thursday.
The setback, which has forced the company to switch back to using chips from U.S.-based Nvidia, underscores the immense challenges Chinese tech firms face in their government-backed push for technological self-sufficiency.
Technical Hurdles Stall R2 Model
According to the report, which cited three sources familiar with the matter, DeepSeek ran into “persistent technical issues” when training its new R2 model on Huawei’s Ascend chips. These problems were the primary reason the model’s launch was postponed from its original May timeline.
As a result, DeepSeek has reportedly altered its strategy. It will now use Nvidia chips for the computationally intensive training phase—the process of teaching an AI model using vast amounts of data. Huawei’s Ascend chips will instead be used for inference, the less demanding task of running the already-trained model to generate responses.
A Setback for China’s Tech Ambitions
The incident highlights a critical vulnerability for China’s booming AI sector. Despite a strong push from Beijing to adopt homegrown technology, particularly in the face of U.S. export restrictions, Chinese AI developers continue to rely heavily on Nvidia’s market-leading chips.
While the U.S. recently allowed Nvidia to resume sales of its H20 chips in China, Beijing has flagged security concerns over the use of American hardware in critical infrastructure. However, Nvidia’s chips remain the most popular choice for major Chinese AI players, including ByteDance, Tencent, and Alibaba. DeepSeek’s previous model, the R1, which made a significant impact on the AI sector earlier this year, was also largely developed using Nvidia’s H20 chips.
In a sign of growing official pressure, China reportedly began asking local developers this week to provide justifications for their orders of Nvidia chips.
Despite the delay, Chinese media reports suggest that the R2 model is expected to be released in the coming weeks, as DeepSeek races to keep pace with a series of high-profile releases from its competitors.
