Ant Group Taps Chinese Chips to Train AI Models, Cutting Costs and Reducing US Tech Dependence
Ant Group, the financial technology giant backed by Alibaba, is increasingly turning to Chinese-made semiconductors to power its artificial intelligence efforts, as reported by AINews.
The company has reportedly utilized chips from domestic suppliers, some affiliated with its parent Alibaba, as well as from Huawei Technologies to train its large language models (LLMs) using the Mixture of Experts (MoE) framework.
Insiders claim that the results achieved using Chinese chips were comparable to those trained on Nvidia’s H800 GPUs. While Ant still employs Nvidia hardware for certain projects, it is said to be pivoting toward options from AMD and local Chinese manufacturers for newer models.
Navigating Export Barriers
Chinese companies, like Ant, are innovating to bypass US chip export restrictions by experimenting with local chips. This effort allows them to continue AI development despite limited access to advanced GPUs like Nvidia’s H800, which, though not the most powerful, is one of the few high-end chips legally available in China.
According to AINews, in a recently published research paper, Ant detailed its training approach using less advanced chips, claiming performance in some cases exceeded that of Meta’s AI models. though Bloomberg initially reported the findings, the results have not been independently verified.
Using MoE to Lower Training Costs
Ant’s AI models use the Mixture of Experts technique, which segments tasks for efficiency, lowering computing demands. This method, adopted by companies like Google and DeepSeek, reduced the cost of processing one trillion tokens from $880,000 to $660,000 by using lower-spec chips without sacrificing performance.
Ant is deploying its AI models, Ling-Plus (290B parameters) and Ling-Lite (16.8B parameters), in healthcare and finance. It acquired Haodf.com to strengthen its healthcare AI and offers services like Zhixiaobao and Maxiaocai. Both models are open-sourced, with Ling-Plus smaller than OpenAI's GPT-4.5 (1.8 trillion parameters).
Despite its cost-effective approach, Ant acknowledged ongoing challenges in training AI models. Minor changes to hardware or model architecture during the training process occasionally resulted in instability, including fluctuations in error rates.
Also read: Alibaba Set to Launch Qwen 3 AI Model Amid Intensifying Competition with DeepSeek
Competing Philosophies on AI Hardware
Ant’s emphasis on cost-effective training contrasts with Nvidia’s strategy. CEO Jensen Huang believes demand for high-performance chips will continue to grow, driven by companies seeking to scale AI capabilities for revenue generation. Nvidia remains focused on creating increasingly powerful chips, investing in more cores, memory, and transistors.
