Ant Group, the fintech giant backed by Jack Ma, is revolutionizing AI development by training large language models using domestically produced chips, significantly reducing costs and challenging the dominance of Western hardware giants like Nvidia.
According to new reports, Ant has achieved performance on par with Nvidia’s H800 processors by utilizing semiconductors from Alibaba and Huawei. This innovation slashes AI model training expenses by up to 20%, a substantial margin in an industry where such costs can reach millions.
Mixture of Experts: The Core Strategy
At the heart of Ant’s breakthrough is the Mixture of Experts (MoE) approach, a machine learning technique that splits complex tasks into smaller segments handled by specialized sub-models. This not only boosts efficiency but also optimizes resource use.
Robin Yu, CTO of Shengshang Tech Co., illustrated the significance with a martial arts analogy:
“If you find one point of attack to beat the world’s best kung fu master, you can still say you beat them… real-world application is important.”
By training 1 trillion tokens at 6.35 million yuan ($880,000) using top-tier hardware and reducing the cost to 5.1 million yuan with Chinese chips, Ant demonstrates the real-world financial benefits of its technique.
Introducing Ling-Lite and Ling-Plus
Ant’s innovations go beyond cost savings. It has released two new language models:
- Ling-Lite: A 16.8 billion parameter model outperforming Meta’s Llama in some English benchmarks
- Ling-Plus: A robust 290 billion parameter model, among the larger models in the industry
Both models are open-source and reportedly surpass counterparts from DeepSeek on Chinese-language performance, reinforcing China’s growing capability in AI.
For context, OpenAI’s GPT-4.5 has an estimated 1.8 trillion parameters, and DeepSeek-R1 has 671 billion—highlighting the scale of Ling-Plus.
China’s Strategic Shift in Chip Dependency
Though Nvidia chips remain in use, Ant is increasingly relying on AMD and Chinese-designed processors, signaling a strategic move toward self-reliance amidst strict US export controls.
Robert Lea from Bloomberg Intelligence observed:
“China was well on the way to becoming self-sufficient in AI as the country turned to lower-cost, computationally efficient models.”
From Labs to Real Life: AI in Action
Ant is already putting its AI to work:
- Healthcare: Acquisition of Haodf.com and deployment of an “AI Doctor Assistant” to aid over 290,000 doctors
- Finance: Introduction of “Maxiaocai,” an AI financial advisor
- Consumer Tech: Launch of “Zhixiaobao,” an AI-driven life assistant
Large model healthcare machines are now operating in seven hospitals across China’s biggest cities.
Nvidia’s Challenge
This development contrasts Nvidia CEO Jensen Huang’s vision, which leans on ever-increasing computational power to meet AI demands. While Huang pushes for higher-end chips, Ant’s approach shows that smarter, leaner models can be equally transformative.
Still, the journey hasn’t been without obstacles. Ant admits training stability issues, with minor changes in hardware or model structure causing unexpected error rate fluctuations.
Nevertheless, Ant’s AI strategy marks a bold step in China’s technological independence and a challenge to the global AI status quo.
