In a bold move to reduce costs and lessen dependence on U.S. technology, Ant Group AI models are now being trained using Chinese-made semiconductors. The Alibaba-affiliated fintech giant is leveraging domestic chips from suppliers tied to Alibaba and Huawei, experimenting with alternative hardware to Nvidia’s powerful but restricted H800 GPUs, according to sources familiar with the matter.
While Ant Group AI models continue to use Nvidia chips for certain AI developments, the company is increasingly exploring cost-effective alternatives from AMD and local Chinese chipmakers. Reports indicate that the Mixture of Experts (MoE) training method used with domestic chips has delivered results comparable to those achieved with Nvidia’s hardware.
China’s AI Race and Ant Group’s Strategy
The shift signals Ant Group’s deeper involvement in the AI competition between China and the U.S., particularly as Chinese firms face export restrictions on high-performance GPUs. The company’s push toward domestic AI infrastructure reflects a broader trend in China, where businesses are working to develop AI capabilities without relying on Western technology.
According to a research paper published by Ant, some of its AI models have outperformed those developed by Meta in select tests. However, Bloomberg News, which first reported the development, has not independently verified these claims. If accurate, this could mark a major step forward in China’s ability to reduce the cost of AI training while maintaining competitive performance.
What Makes Ant Group AI Models Unique?
Ant Group’s MoE models break down complex AI tasks into smaller components, making training more efficient. This technique, also used by Google and DeepSeek, allows AI models to be trained using lower-specification GPUs, reducing overall computing costs.
Training AI models is typically expensive, with high-performance GPUs being a significant cost barrier. Ant Group’s research suggests that training one trillion tokens using conventional hardware costs 6.35 million yuan ($880,000). However, by optimizing its training process with domestic chips, the company slashed costs to 5.1 million yuan—a significant reduction in AI model training expenses.
AI Applications in Healthcare and Finance
Ant Group plans to deploy its AI models—Ling-Plus (290 billion parameters) and Ling-Lite (16.8 billion parameters)—in industrial applications like healthcare and finance. The company’s recent acquisition of Haodf.com, a leading Chinese online medical platform, underscores its ambition to expand AI-driven solutions in the healthcare industry. Ant also operates other AI-powered services, such as Zhixiaobao, a virtual assistant, and Maxiaocai, a financial advisory platform.
How This Challenges Nvidia’s AI Strategy
Nvidia’s CEO, Jensen Huang, has emphasized that demand for high-performance GPUs will continue to grow, despite more efficient AI models like DeepSeek’s R1. However, Ant’s approach—reducing training costs by using less expensive chips—contrasts sharply with Nvidia’s strategy of building ever-more powerful GPUs.
Ant’s research paper acknowledges that training AI models with alternative hardware comes with challenges. The company noted that even minor adjustments to hardware or model structures during training can lead to instability and performance fluctuations, including spikes in error rates.
China’s AI Future: A Step Toward Self-Sufficiency?
Ant’s decision to make its AI models open-source could further accelerate AI innovation in China. While Ling-Plus and Ling-Lite are significantly smaller than GPT-4.5 (estimated at 1.8 trillion parameters, according to MIT Technology Review), their success with domestic chips signals China’s growing ability to train competitive AI models without relying on Western hardware.
As China’s AI ecosystem continues to evolve, Ant Group AI models could play a crucial role in shaping a self-reliant tech industry, challenging Nvidia’s dominance and offering cost-effective AI solutions for businesses worldwide.
Get the Latest AI News on AI Content Minds Blog