Ant Group, the fintech company backed by Jack Ma, has achieved an important advancement in artificial intelligence (AI) model training by using Chinese-made semiconductors (chips). According to a Bloomberg report, by incorporating chips from Alibaba Group and Huawei Technologies, Ant Group has developed a method that reduces AI training costs by around 20%.
In fact, despite using locally produced chips instead of relying on costly, high-end imported technology and hardware, the Chinese company is still capable of delivering high-performance results and remains competitive.
Generally, training large AI models requires expensive and high-performance chips, making it challenging for smaller companies to compete. But Ant Group’s latest approach leverages domestic semiconductors and a machine learning technique known as Mixture of Experts (MoE). Speaking of working, this method divides tasks into specialized subsets, enhancing efficiency and reducing dependency on costly hardware. Notably, Ant Group’s models have outperformed those from larger international tech companies (like OpenAI and others) in certain benchmarks.
In terms of significance, by reducing training costs, Ant Group’s strategy could democratize access to advanced AI tools, enabling smaller businesses to participate more actively in AI development. The company has reported that training an AI model on one trillion tokens typically costs around 6.35 million yuan (~ $880,000) when using high-performance hardware. However, with its optimized method (using locally produced low-cost semiconductors), the company have managed to reduce these costs by around 20%, bringing the expense down to about 5.1 million yuan.
Interestingly so far, Ant Group has developed two large language models (LLMs) – Ling-Plus and Ling-Lite – to enhance AI applications in sectors like healthcare (a life assistant app, Zhixiaobao) and finance (a financial advisory service, Maxiaocai). Ling-Plus is a 290-billion-parameter model, while Ling-Lite has 16.8 billion parameters.
Meanwhile, in a recent research paper, the company also claims that its artificial intelligence (AI) models are performing as well as or better than those developed by leading organizations like DeepSeek, Mistral, and Meta’s Llama 3.1. This is a noteworthy claim as earlier Hangzhou-headquartered startup – DeepSeek gained attention for its open-source model, DeepSeek V3, which has outperformed prominent models like Llama 3.1 and OpenAI’s GPT-4o in various benchmarks.
The development aligns with a broader trend among Chinese tech firms to seek local chip solutions, especially at a time when China is facing US export restrictions on advanced AI chips. However, despite such restrictions, Chinese companies like Alibaba, Tencent, and DeepSeek have made significant progress in AI research and development.
In another example of an ongoing intensified AI race in China, tech giant Tencent launched its latest ‘T1 reasoning model.’ This model is optimized for handling long documents quickly and accurately. As per the company’s claim, it processes extended text efficiently, responding faster while keeping the content logical and clear. Notably, China’s AI market alone is projected to hit $206.6 billion by 2030, growing at a much faster rate (44.4% CAGR) than the global average.
