According to Bloomberg, Ant Group, a major fintech group supported by Jack Ma, utilizes semiconductor technology produced in China such as Alibaba and Huawei, and employs a mixture-of-experts (MoE) architecture to reduce training costs by 20% instead of using solely NVIDIA H800 chips.
Ant stated that the cost of training Token exceeds 1 trillion units, utilizing high-performance hardware costing around 6.35 million yuan (880,000 USD), but by using Chinese chips, the cost reduces to 5.1 million yuan.
With the United States restricting the export of high-level chips to China, Big Tech China has turned to use their own technology. According to Ant’s latest research, AI models like Ling-Plus and Ling-Lite outperform Meta and DeepSeek models in some aspects, particularly in Chinese language processing, without the need for consistently expensive hardware.
Moreover, Ant has opened up the use of open-source AI models for developers to innovate further. The Ling-Lite model with 1.68 trillion parameters, and Ling-Plus with over 2.90 trillion, smaller than OpenAI’s GPT-4.5 with 1.8 trillion parameters.
Nevertheless, Ant faces challenges, especially in the stability of the models due to slight hardware changes or model structures that could lead to errors.
However, Ant’s AI models are being utilized in various hospitals in major cities like Beijing and Shanghai, integrating their own technology with other Chinese AI models like DeepSeek R1 and Alibaba’s Qwen to assist in medical consultations and healthcare services.
TLDR: Ant Group, supported by Jack Ma, uses advanced semiconductor technology and a mixture-of-experts architecture to reduce training costs. Their AI models outperform others in certain aspects and are being used in hospitals for medical consultations.
Leave a Comment