DeepSeek Reverts to Nvidia for R2 AI Training After Huawei Chip Failure

DeepSeek’s ambitious plan to train its new artificial intelligence model, R2, on Huawei’s Ascend chips has faltered, forcing the Chinese AI firm to revert to Nvidia’s technology and delaying the model’s launch. This setback highlights the unyielding technical realities that can challenge Beijing’s push for technological self-sufficiency.

Following the successful launch of its R1 model in January, DeepSeek found itself under considerable pressure to champion the national cause by adopting domestic hardware. According to three individuals with direct knowledge of the situation, speaking to the Financial Times, the directive was unambiguous: prioritize Huawei’s chips over Nvidia’s.

However, when DeepSeek commenced the actual training of its new R2 model, the company reportedly encountered “persistent technical issues” with Huawei’s AI chips. These problems proved so fundamental that the project ground to a halt. A source familiar with the situation indicated that these insurmountable challenges were the primary reason for scrapping the model’s planned May launch, leaving DeepSeek at a disadvantage in a rapidly evolving market.

To grasp the significance of this hurdle, it’s crucial to distinguish between AI training and inference. Training is the profoundly demanding phase, akin to years of intensive university-level learning, requiring immense computational power and unwavering stability. Inference, by contrast, is the comparatively less strenuous task, like asking a graduate a question – it involves applying a trained model to new data. DeepSeek discovered that while Huawei’s chips might be adequate for the final exam (inference), they were not yet capable of handling the rigorous demands of the university course (training). Consequently, the company had no recourse but to switch back to Nvidia’s more robust systems for the critical training stage. Sources indicate that DeepSeek’s team is still attempting to optimize the R2 model for the less demanding inference stage using Huawei chips.

The severity of the issue was underscored by Huawei’s direct intervention. Two sources confirmed that Huawei dispatched its own team of engineers to DeepSeek’s offices to assist in getting the R2 model operational on their chips. Yet, even with these expert resources on-site, a successful training run remained elusive.

Industry observers widely acknowledge that this outcome is not entirely surprising. Earlier this year, Huawei CEO Ren Zhengfei himself conceded that the US had “exaggerated Huawei’s achievements” and that the company “is not that great yet,” admitting that its best chips still lag a generation behind leading-edge alternatives.

Despite these technical limitations, Beijing continues to actively encourage its technology giants to favor local hardware. The Financial Times has reported that Chinese firms are now compelled to justify orders for Nvidia’s export-compliant H20 chip, which is a less powerful variant approved for sale in China. This strategy aims to cultivate domestic champions but can inadvertently force companies into technically suboptimal choices, potentially hindering their global competitiveness.

Beyond the challenges posed by Huawei’s chips, DeepSeek founder Liang Wenfeng has reportedly expressed dissatisfaction with the overall progress of the R2 model, urging his team to aim higher and develop a product that can secure the company’s position among the AI industry leaders.

Ultimately, DeepSeek’s experience serves as a potent reminder that in the global race for AI supremacy, engineering principles and performance realities often trump top-down directives and national pride. While China plays the long game in its pursuit of technological independence, for the foreseeable future, the performance crown in AI hardware remains firmly with Nvidia.