BEIJING – DeepSeek has reportedly delayed the release of its next-generation model, the R2, after encountering significant difficulties training the system on domestically produced chips. This development highlights the ongoing challenges facing China’s ambitious goal of achieving technological self-sufficiency and reducing its dependence on U.S. chip giant Nvidia.
According to sources familiar with the situation, DeepSeek was encouraged by Chinese authorities to use Huawei’s Ascend processors for its R2 model, following the release of its successful R1 model in January. However, the company faced persistent technical issues during the training phase using the Ascend chips. This ultimately led DeepSeek to revert to using Nvidia chips for training while retaining Huawei’s for inference, the process of using a trained model to generate responses.
The training difficulties with the Ascend chips were the primary reason the R2’s launch was postponed from May, a delay that has caused the company to lose ground to its rivals. The setback underscores how Chinese-made chips still lag behind their American counterparts, particularly for the demanding task of training large language models.
The incident with DeepSeek reveals the challenges facing China’s push to replace U.S. technology in critical areas, and it highlights the enduring reliance on companies like Nvidia.
The difficulties experienced by DeepSeek are not isolated. Industry insiders have reported that Chinese chips, including those from Huawei, suffer from stability issues and slower inter-chip connectivity compared to Nvidia’s products. They also note that the software ecosystem for these domestic chips is less mature.
Even with a dedicated team of Huawei engineers working on-site at DeepSeek’s offices, the startup was reportedly unable to conduct a successful training run on the Ascend chip. DeepSeek is, however, still collaborating with Huawei to ensure the R2 model is compatible with the Ascend chip for inference tasks.
The delay in the R2 launch has also reportedly caused internal friction. DeepSeek founder Liang Wenfeng has expressed dissatisfaction with the model’s progress, pushing his team to spend more time developing an advanced system that can maintain the company’s competitive edge in the AI field. Another person with knowledge of the matter added that the launch was also delayed by longer-than-expected data labeling for the updated model.
The situation also comes amid a broader push from Beijing to promote domestic alternatives. The Financial Times recently reported that Chinese tech companies are now required to justify their orders of Nvidia’s H20 chips, a move seen as a way to pressure them into adopting products from companies like Huawei and Cambricon.
The fierce competition and the geopolitical pressures are shaping the global AI landscape. Ritwik Gupta, an AI researcher at the University of California, Berkeley, noted that “models are commodities that can be easily swapped out.” He pointed to Alibaba‘s Qwen3 as a powerful alternative, noting it adopted core concepts from DeepSeek, such as its training algorithm, and made them more efficient.
Gupta, who monitors Huawei’s AI ecosystem, described the company’s challenges with using Ascend for training as “growing pains.” However, he remains optimistic about the future. “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,” he said.
Nvidia, a central player in the geopolitical tech battle, recently agreed to share a portion of its Chinese revenue with the U.S. government to resume sales of its H20 chips to the country. The company has stressed the importance of its role in the market, stating that “developers will play a crucial role in building the winning AI ecosystem.” In a statement, Nvidia added, “Surrendering entire markets and developers would only hurt American economic and national security.”
DeepSeek and Huawei have not responded to requests for comment regarding the situation.
- With reporting from Victor P.
