Ant Group, the fintech affiliate supported by Alibaba founder Jack Ma, is reported to be leveraging domestically manufactured chips in the development and training of its artificial intelligence (AI) models—a move that purportedly cuts costs by 20 percent and marks the firm’s entry into the intensifying AI competition between Chinese and American companies.
According to Bloomberg on Monday (March 24), sources disclosed that Ant Group is employing Chinese-made chips from companies including Alibaba and Huawei, and is utilizing Mixture of Experts (MoE) technology—originally adopted by the domestic startup DeepSeek—for model training. The performance of these models is said to be on par with those developed using processors such as Nvidia’s H800.
The MoE model, which has gained traction in the AI industry after its adoption by firms like Google and DeepSeek, divides tasks into smaller datasets much like a team of specialists each handling a particular segment, thereby enhancing overall efficiency.
A source familiar with the matter noted that while Ant Group continues to use Nvidia chips for AI development, its latest models primarily rely on alternatives, including products from American semiconductor company AMD as well as Chinese chipmakers.
Since DeepSeek demonstrated how to train powerful models at a fraction of the cost incurred by companies such as OpenAI, competition between Chinese and American firms in the AI arena has intensified. Ant Group’s strategy not only thrusts it into the competitive landscape but also underscores China’s broader efforts to replace advanced Nvidia chips with domestic products.
Although Nvidia’s H800 is not the most advanced processor available, it remains a relatively powerful chip that is currently banned from export to China.
Earlier this month, Ant Group published a research paper claiming that its model outperformed benchmarks set by Meta, Facebook’s parent company, on certain tests. Bloomberg industry research analyst Robert Lea remarked that, if substantiated, Ant’s assertions would highlight China’s progress toward AI self-sufficiency, as the nation shifts toward more cost-effective and computationally efficient models to circumvent export controls on Nvidia chips.
Meanwhile, Chinese tech giant Tencent has upgraded its AI inference model, the Mixing T1, claiming that its performance is comparable to DeepSeek’s R1, OpenAI’s GPT 4.5, and o1, thus maintaining its competitive edge in the AI race.