DeepSeek AI model fails Huawei chip training, delays launch

Chinese artificial intelligence firm DeepSeek has reportedly postponed the launch of its new R2 model, a setback attributed to persistent technical difficulties encountered while attempting to train it using Huawei’s Ascend processors. This delay underscores the significant hurdles Beijing faces in its ambitious drive to replace foreign, particularly US, technology with domestic alternatives.

Sources familiar with the situation indicate that Chinese authorities had encouraged DeepSeek to adopt Huawei’s Ascend chips for its R2 model, following the release of its R1 model in January, steering it away from prevalent Nvidia systems. However, DeepSeek quickly ran into issues during the intensive training phase of its R2 model with Ascend chips. Consequently, the company resorted to using Nvidia chips for the crucial training process, reserving Huawei’s chips for the less demanding inference stage—where a trained model applies its knowledge to generate predictions or responses, such as a chatbot query.

The technical snags with Huawei’s hardware were the primary cause for the R2 model’s launch being pushed back from May, a delay that has allowed rivals to gain ground. Industry insiders point to a range of performance disparities, noting that Chinese chips, including Huawei’s, still lag behind their US counterparts in critical areas. These deficiencies reportedly include stability issues, slower inter-chip connectivity, and less robust accompanying software compared to Nvidia’s offerings.

Despite Huawei dispatching a team of engineers to DeepSeek’s offices to assist with the R2 model’s development on Ascend chips, the collaborative efforts did not yield a successful training run. DeepSeek is reportedly still working with Huawei to ensure the model’s compatibility with Ascend for inference, indicating a continued commitment to domestic hardware where feasible.

The challenges have not gone unnoticed internally. DeepSeek founder Liang Wenfeng has expressed dissatisfaction with the R2’s progress, reportedly pushing for more time and resources to develop an advanced model capable of maintaining the company’s competitive edge in the rapidly evolving AI landscape. Compounding the chip-related delays, the R2 launch was also impacted by longer-than-expected data labeling requirements for its updated model. Despite these setbacks, Chinese media reports suggest the model could still be released in the coming weeks.

The broader context for DeepSeek’s struggles is Beijing’s intensified push for technological self-reliance. The Financial Times recently reported that Chinese tech companies are now required to justify their orders of Nvidia’s H20 chips, a move explicitly aimed at promoting domestic alternatives from companies like Huawei and Cambricon. This pressure highlights the strategic importance placed on homegrown technology in the ongoing geopolitical competition.

Ritwik Gupta, an AI researcher at the University of California, Berkeley, commented on the dynamic nature of the AI market, noting that “models are commodities that can be easily swapped out.” He cited Alibaba’s Qwen3 as a powerful and flexible alternative that has successfully integrated DeepSeek’s core concepts, such as its reasoning-capable training algorithm, making them more efficient. Gupta, who closely monitors Huawei’s AI ecosystem, acknowledged that the company is experiencing “growing pains” with Ascend chips for training purposes. However, he remains optimistic about Huawei’s long-term prospects, stating, “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.”

Meanwhile, Nvidia, a central player in the US-China tech rivalry, recently agreed to a revenue-sharing arrangement with the US government to resume sales of its H20 chips to China. The chipmaker has emphasized the importance of developers in building a winning AI ecosystem, arguing that “surrendering entire markets and developers would only hurt American economic and national security.” Neither DeepSeek nor Huawei responded to requests for comment on the matter.

[China’s AI ambitions hit a chip wall: DeepSeek’s struggle with Huawei highlights the tough road to tech independence.]