In a setback for China’s push toward technological self-reliance, artificial intelligence company DeepSeek has abandoned plans to train its upcoming R2 model on Huawei’s Ascend chips, reverting instead to Nvidia’s hardware after repeated technical failures.
The move follows months of pressure from Beijing for Chinese tech leaders to prioritize domestic solutions over U.S. technology. DeepSeek’s first model, R1, launched successfully in January and positioned the company as one of the nation’s AI frontrunners. But according to multiple people familiar with the matter, the government made clear it wanted the company’s next-generation model trained on local hardware—specifically Huawei’s chips.
That plan quickly ran into trouble. Sources say the Ascend processors struggled with the demanding task of AI training, a process that requires vast computing power and rock-solid stability. While the chips may be capable of handling “inference”—the lighter, day-to-day operation of a trained model—they fell short during the intense training phase. Attempts to resolve the issues, including sending Huawei engineers to work alongside DeepSeek’s team, failed to produce a working result.
As a result, DeepSeek’s R2 launch—originally targeted for May—was postponed, forcing the company to fall back on Nvidia’s export-compliant H20 chips to keep development on track. DeepSeek is still exploring ways to use Huawei hardware for inference once the model is complete, but for now, training will be done using Nvidia’s proven technology.
Industry experts say the outcome is unsurprising. Huawei’s own CEO, Ren Zhengfei, has acknowledged that the company’s best chips remain at least a generation behind top U.S. offerings. Despite that gap, Chinese regulators continue to push companies toward homegrown solutions, requiring them to justify any purchase of Nvidia chips.
L_104026
For DeepSeek founder Liang Wenfeng, the chip issue is just one concern. Insiders say he has urged his team to accelerate progress and aim for higher performance to keep pace with global AI leaders.
The episode underscores a broader reality: while national ambitions can set the direction, the unforgiving demands of engineering determine the pace. For now, Nvidia retains its edge in high-performance AI training—leaving China’s quest for chip independence a work in progress.
