Ant Group, backed by Jack Ma, has developed innovative techniques for training AI models using Chinese-made semiconductors that could slash costs by 20%, according to recent reports.
The company leveraged chips from Alibaba (an Ant affiliate) and Huawei to achieve results comparable to those produced by Nvidia’s H800 processors, which are currently restricted from sale to China under US export controls.
The MoE Approach
Ant’s breakthrough is heavily credited to the “Mixture of Experts” (MoE) machine learning approach. This technique divides complex AI tasks into smaller, more manageable datasets, similar to assembling a team of specialists, each focusing on their area of expertise.
Robin Yu, chief technology officer at Shengshang Tech Co., compared the achievement to martial arts: “If you find one point of attack to beat the world’s best kung fu master, you can still say you beat them, which is why real-world application is important.”
The financial advantage is significant. Ant reports that training 1 trillion tokens (the basic units of information AI models process) costs about 6.35 million yuan ($880,000) using high-performance hardware.
Their optimized approach reduces this to 5.1 million yuan with lower-specification hardware—representing substantial savings in an industry where training costs can run into millions.
New Models Challenging Global Tech Giants
Ant has developed two notable language models, namely:
- Ling-Lite: A 16.8 billion parameter model that reportedly outperforms Meta’s Llama in certain English-language benchmarks
- Ling-Plus: A more robust 290 billion parameter model, considered relatively large in the realm of language models
For context, experts estimate ChatGPT’s GPT-4.5 has approximately 1.8 trillion parameters, while DeepSeek-R1 has 671 billion.
Both Ling models have been made open source and supposedly outperform DeepSeek’s equivalents on Chinese-language benchmarks.
Shift in AI Chip Usage
While Ant continues to use Nvidia chips for some AI development, the company is increasingly relying on alternatives, including AMD processors and Chinese chips, for its latest models.
This pivot indicates a broader trend among Chinese tech companies seeking self-sufficiency in the face of US export restrictions on advanced semiconductors.
Robert Lea, a senior analyst at Bloomberg Intelligence, noted that Ant’s claims, if confirmed, suggest “China was well on the way to becoming self-sufficient in AI as the country turned to lower-cost, computationally efficient models to work around the export controls on Nvidia chips.”
Real-World Applications Already Emerging
Ant isn’t limiting these AI innovations to theoretical research. The company has already deployed AI-powered solutions in several sectors. These include:
- Healthcare: Ant purchased online platform Haodf.com to strengthen its AI healthcare services and developed an “AI Doctor Assistant” supporting 290,000 doctors with tasks like medical record management
- Finance: The company offers a financial advisory AI service called “Maxiaocai”
- Consumer Applications: Ant has launched “Zhixiaobao,” an AI “life assistant” app
The company has also built healthcare-focused large model machines currently used by seven hospitals and healthcare providers in major Chinese cities, including Beijing and Shanghai.
Challenging Nvidia’s Perspective
Interestingly, this development runs counter to Nvidia CEO Jensen Huang’s strategy.
Huang has argued that computation demands will continue growing even with more efficient models, believing companies will need better (not cheaper) chips to generate more revenue. Nvidia has pursued building increasingly powerful GPUs with more processing cores, transistors, and memory capacity.
Despite its achievements, Ant acknowledged challenges in the training process. Even small changes in hardware or model structure led to stability issues, including unexpected jumps in error rates.
