Ant Group reduces AI inference costs by 20% via deployment of Chinese-manufactured semiconductor chips.

Ant Group, backed by Alibaba cofounder Jack Ma, has achieved a significant breakthrough in AI model training by integrating Chinese-made semiconductors to cut computing costs by 20%. According to Bloomberg, the Hangzhou-based company used domestic chips from Alibaba and Huawei to train models using the Mixture of Experts (MoE) approach, which enables training with substantially reduced computational resources.

By combining Chinese and U.S.-made semiconductors in MoE training, Ant Group has reduced computing costs while limiting dependence on single-chip suppliers like NVIDIA. Sources familiar with the initiative—speaking on condition of anonymity—said the company achieved results comparable to those produced using NVIDIA's H800 chips.

The shift reflects Ant Group's strategy as the AI race between U.S. and Chinese companies intensifies. While still using NVIDIA chips, Ant Group is increasingly relying on alternative semiconductors for its latest models, demonstrating how Chinese developers can build advanced AI without exclusive dependence on U.S. suppliers. This is particularly significant given that the H800 is currently restricted under U.S. export controls aimed at limiting China's access to cutting-edge hardware essential for AI development.

Ant Group recently published research claiming its models have at times outperformed Meta's in internal benchmark tests. The company also suggested its training strategy could reduce inference costs—the expense of delivering real-time AI services—making advanced capabilities more affordable. If validated, Ant Group's cost-efficient techniques could represent a significant milestone in China's AI development.

Ant Group is not alone in this effort. Chinese startup DeepSeek released its R1 model earlier this year, adding momentum to the demonstration that powerful AI models can be trained at lower cost. Ant Group has also open-sourced its Ling models—Ling-Lite and Ling-Plus—further encouraging AI development across the region.

MoE has become a preferred approach in AI training, dividing tasks into smaller datasets to optimize performance and efficiency. Ant Group's cost-conscious methods could broaden AI access by reducing developers' reliance on premium, high-performing chips.

Yet NVIDIA CEO Jensen Huang offered a contrasting perspective at the company's recent GTC conference, arguing that companies seeking to maximize revenue will require more powerful chips, not cheaper ones—positioning performance rather than price as the future of AI infrastructure.