Despite Huawei’s progress, Nvidia continues to dominate AI chips market in China

Huawei is ramping up its AI chip production but will be unable to meet the strong demand of China’s booming AI sector any time soon. Chip behemoth Nvidia continues to dominate the Chinese AI chip market, selling one million of its H20 chips in 2024. H20 are downgraded AI chips, specifically designed for the Chinese market to match US export restrictions. In the same year, Huawei only sold 200,000 units of its competing Ascend 910B chips. The Chinese company plans to increase output to 300,000 in 2025, while also commencing the fabrication of the newer 910C chips, targeting 100,000 pieces annually.

Huawei’s slow progress reveals structural issues in China’s semiconductor industry. US and Dutch export restrictions on advanced lithography technologies have led to inefficient production and low yield rates. Many chips are defective. While Huawei is making strides technologically, reportedly increasing the yield rate of its newest 910C chip from 20 to 40 percent in 2025, global industry leaders are closer to 60 percent for large AI chips. 

Huawei and [its production partner] SMIC are also unable to easily expand production capacity. Rumors even suggest that significant parts have been manufactured in secret by TSMC. To progress technologically, Huawei and SMIC would need breakthroughs, as no company has ever used the equipment available to create smaller, faster chips below 7nm. 

Moreover, manufacturing limitations also impact the functionality of Huawei chips, with the 910B reportedly struggling with inter-chip connectivity and memory speeds. While the 910B seems good enough for many inference workloads, i.e., for running a large language model that has already been trained, its limitations make it a poor choice for model training itself. 

Even the downgraded H20 is preferred for many applications. In addition, Huawei’s software lags behind Nvidia’s, with developers in China reluctant to adopt the chip for training most models, as training runs are often costly and long (DeepSeek, which became famous partly because of its quick and cheap training run, spent between 3 and 8 weeks on its final training run for R1). 

Antonia Hmaidi, Senior Analyst at MERICS: “Nvidia will likely maintain its dominance over China’s AI chip market in the foreseeable future. The demand for AI chips will only increase after legislators reaffirmed support for AI development at the National People’s Congress in March. However, Huawei cannot meet rising demand with its current production capacity due to US and Dutch export controls.”