Jack Ma's Ant Group Demoes Groundbreaking AI Chips as China Eyes Local AI Supremacy

Jack Ma-backed Ant Group has unveiled chips that can support the training of foundational models at 20 per cent less cost than what global competitors such as OpenAI and Alphabet's Google are using. According to Bloomberg, the new AI breakthrough uses Chinese-made semiconductors to train models based on the Mixture of Experts machine learning approach. Not only does it reduce costs, but it could also be a major milestone in China's bid to bolster its position in the AI landscape.

Ant Group has used locally developed chips, sourced from Jack Ma-led Alibaba Group and Huawei Technologies — one of China's biggest technology companies. Techniques developed using these chips allow for the training of AI models at significantly lower costs to provide results similar to those from Nvidia chips such as H800. Although Ant Group is still relying on Nvidia chips to develop these techniques, the report said it is also using alternatives such as semiconductors from Advanced Micro Chips, better known as AMD, and other Chinese-origin chips.

With the latest development, Ant has jumped on the bandwagon, especially at a time when China is putting out all the stops to counter the US in the AI race. Its claim that the local AI chips can reduce costs by 20 per cent puts more scrutiny on more capable models, which OpenAI and Google have poured in billions of dollars to train.

Its entry into the AI race also emphasises the need for cost-effectiveness in AI development, sparked after the Chinese startup DeepSeek took the world by storm with its AI chatbot. DeepSeek claimed its AI model was trained at less than half the cost that industry leaders spent on their AI models. The key to this alternative is the use of MoE models, popularised by DeepSeek. These models involve the division of tasks into smaller sets of data, putting less pressure on the hardware and still getting the job done. However, their accuracy and speed have been questioned against the more capable models, such as OpenAI's GPT and Google's Gemini.

According to the report, Ant published a research paper earlier this month, claiming that its AI models scored better than Meta's models in certain benchmarks.

Read more: Realme’s New P3 Is a Waterproof Phone At Under ₹17,000