Jack Ma-backed group cuts AI training costs 20%—using Chinese-made chips
What's the story
Ant Group, the Jack Ma-backed fintech giant, has made a major breakthrough in AI model training.
The company says it has cut training costs by 20% with Chinese-made semiconductors. The innovative method was developed with chips from Alibaba Group Holding Ltd and Huawei Technologies Co.
The new technique builds on the Mixture of Experts machine learning method, which breaks down tasks into smaller data sets for efficient processing.
AI development
Ant Group's AI models rival NVIDIA's performance
Ant Group has claimed that its AI models, built on Chinese semiconductors, have yielded results similar to those produced with NVIDIA chips such as the H800.
Ant still uses NVIDIA for AI development but is now depending on alternatives like Advanced Micro Devices (AMD), and Chinese chips for its latest models.
This highlights a growing trend among Chinese companies who prefer local alternatives to advanced foreign semiconductors.
Performance comparison
Outperforming Meta Platforms
In a recent research paper, Ant Group claimed that its AI models sometimes outperformed those of Meta Platforms Inc. on certain benchmarks.
This claim has not been independently verified, says Bloomberg News.
If substantiated, it could mark a major step forward in Chinese artificial intelligence development and possibly lower the cost of supporting AI services.
Cost reduction
Cost-effective approach to AI training
Ant Group has disclosed that training one trillion tokens cost around 6.35 million yuan ($880,000) on high-performance hardware.
However, their optimized method could bring this down to 5.1 million yuan by utilizing lower-specification hardware.
Tokens are basically units of information that a model consumes to learn about the world and provide useful responses to user queries.
Model performance
Ling-Plus and Ling-Lite models outperform competitors
Ant Group intends to capitalize on the latest advancements in its LLMs, Ling-Plus and Ling-Lite, for AI applications in industries like healthcare and finance.
The company also runs an AI "life assistant" app, Zhixiaobao, and a financial advisory AI service, Maxiaocai.
In English language comprehension, Ant claimed that the Ling-Lite model outperformed one of Meta's Llama models on a key benchmark.
Training issues
Parameter comparison and challenges in AI training
Ant Group's Ling-Lite model has 16.8 billion parameters, while Ling-Plus has 290 billion.
For context, experts estimate ChatGPT's GPT-4.5 has 1.8 trillion parameters (according to MIT Technology Review), while DeepSeek-R1 has 671 billion.
Ant struggled during training due to stability issues, where even minor changes in hardware or model structure could lead to increased error rates in the models.