The AI race continues: China-made processors cut training costs by 20%

A Chinese fintech giant claims using domestic chips can cut AI training costs by 20%, challenging detractors and boosting technological self-reliance.

Just after the recent searing incident in which China-based DeepSeek caused a global stir in America’s AI aspirations, a Chinese tech giant has just leaked another piece of unwelcome news for AI players in the US.

Fintech giant Ant Group has claimed to have reduced AI training costs by 20% using exclusively Chinese-made semiconductors. The approach involves using chips produced by Chinese tech giants Alibaba and Huawei in a “Mixture of Experts” (MoE) machine learning technique that has reportedly yielded results comparable to those obtained using Nvidia’s H800 GPUs “at times”. The MoE technique divides tasks into smaller sets of data, very much like having a team of specialists that each focus on a segment of a job, making the process more efficient.

The Chinese chips used in this claimed breakthrough include various semiconductor technologies that perform at approximately 80% of Nvidia’s H100 AI chipset capabilities. Alibaba’s contribution stems from using an earlier chip (12nm fabrication technology with 17bn transistors): when running the popular ResNet-50 neural network, Alibaba claims this chip could process 78,563 images per second.

While Ant Group continues to use Nvidia hardware for some AI development tasks, it is increasingly relying on alternatives, including AMD chips and those from Chinese manufacturers, for its latest models.

This latest achievement in competitive AI training development occurs against the backdrop of tightening US export controls on advanced semiconductors to China. If the performance of these domestic chips proves consistent, it could potentially impact Nvidia’s strong position in the AI chip market. The achievement could also vindicate countries that are against unfair trade practices in the AI race, while highlighting the inefficiencies of western technological mores to date.

However, industry analysts caution that despite notable progress, Chinese semiconductor manufacturers still face challenges in matching the full ecosystem, software support, and long-term competitiveness offered by their Western counterparts.