Ant Group is turning to domestic semiconductor suppliers to power its next-generation artificial intelligence models, aiming to reduce operational costs and reduce reliance on U.S. technologies.
Backed by Alibaba, Ant Group has reportedly used chips from Chinese manufacturers—including some affiliated with Alibaba itself and Huawei Technologies—to train advanced large language models. These models were developed using the Mixture of Experts (MoE) framework, a technique that divides tasks among several specialized sub-models for efficiency. According to insiders, the performance of these domestically trained models rivals those trained using Nvidia’s H800 chips.
Although Ant still incorporates Nvidia hardware into certain AI workflows, the company is increasingly exploring alternatives from AMD and Chinese chipmakers to push the boundaries of cost-effective model training.
China’s Strategic Shift in AI Hardware
The move is part of a broader national effort to circumvent U.S. export restrictions that limit access to high-end chips. While Nvidia’s H800 is not the most advanced GPU globally, it remains one of the leading options available to Chinese tech firms due to geopolitical constraints.
Ant Group’s deeper engagement in AI development underscores China’s ambition to become a self-reliant technology powerhouse. In fact, a recent partnership between Pythian and GigaOm also echoes the urgency for ethical and scalable AI infrastructure across enterprises, highlighting the global shift toward responsible innovation.
MoE Models: Efficient and Scalable
MoE-based models are gaining traction among developers and researchers due to their ability to handle vast datasets more efficiently. Ant Group described its methodology in a research paper, stating that their models outperformed Meta’s in some benchmark tests. Bloomberg reported on these claims, although the results have yet to be independently verified.
MoE architecture functions like a team of experts, where different components handle specific parts of a task. This approach allows for greater scalability and reduced computational load. Google and Chinese AI startup DeepSeek have also adopted this model for their AI systems.
Cost Savings and Real-World Applications
Training AI models is notoriously expensive, especially when it involves processing trillions of tokens—basic units of data used in machine learning. Ant’s optimized training method reportedly reduced costs from approximately 6.35 million yuan ($880,000) to around 5.1 million yuan, thanks to the use of less expensive hardware.
Ant plans to implement these models—named Ling-Plus and Ling-Lite—in sectors such as finance and healthcare. Earlier this year, the company acquired Haodf.com, a leading online medical platform in China, to further integrate AI into medical services. The group also offers other AI-driven products like the virtual assistant app Zhixiaobao and financial advisory tool Maxiaocai.
Open-Source Contributions and Technical Challenges
In a move to foster transparency and collaboration, Ant has made its models open-source. Ling-Lite features 16.8 billion parameters, while Ling-Plus boasts a massive 290 billion. For context, it’s estimated that GPT-4.5, a closed-source model, contains around 1.8 trillion parameters.
Despite these advancements, Ant Group acknowledged challenges in training stability. Minor changes in hardware configurations or model architecture occasionally led to spikes in error rates, making consistency a hurdle in large-scale deployments.
The Future of AI Development in China
Ant’s strategic pivot toward domestic chips marks a significant milestone in China’s AI ambitions. By reducing the financial and geopolitical costs of AI development, the company is helping to carve out a more independent and sustainable path forward.
As the global AI race intensifies, cost-effective and locally sourced innovation may become the deciding factor in who leads the next wave of technological transformation.
