I'll rewrite this article for you. Here's a more concise and restructured version:
Chinese Semiconductors Help Ant Group Cut AI Training Costs by 20%
Ant Group has successfully leveraged Chinese-made semiconductors to develop cost-effective AI training techniques, according to sources familiar with the matter. Using domestic chips from Alibaba Group and Huawei, the company implemented the Mixture of Experts (MoE) machine learning approach, achieving results comparable to those obtained with Nvidia's H800 chips.
While Ant continues to use Nvidia processors for some AI development, it has increasingly shifted to alternatives from AMD and Chinese manufacturers for its latest models. This move reflects the ongoing competition between Chinese and US companies to develop capable AI models at lower costs, following DeepSeek's demonstration that powerful models can be trained for significantly less than the billions invested by OpenAI and Google.
Cost Efficiency Breakthrough
Ant recently published research claiming its models sometimes outperform Meta's on certain benchmarks. If these claims prove accurate, Ant's platforms could represent a significant advancement for Chinese AI by reducing inferencing costs and enhancing AI service support.
MoE models have gained popularity among companies like Google and DeepSeek because they divide tasks into smaller data sets—similar to assigning specialists to different segments of a job—making the process more efficient. However, training these models typically requires high-performance chips like Nvidia's GPUs, making the process prohibitively expensive for smaller firms.
According to Ant's paper, traditional methods cost approximately 6.35 million yuan ($880,000) to train 1 trillion tokens using high-performance hardware. Their optimized approach reduces this to 5.1 million yuan using lower-specification hardware.
Applications and Performance
Ant plans to apply its Ling-Plus and Ling-Lite language models to industrial AI solutions in healthcare and finance. The company recently acquired Chinese online platform Haodf.com to strengthen its healthcare AI services and has developed an AI Doctor Assistant to support 290,000 doctors with tasks like medical record management.
The company claims its Ling-Lite model outperforms Meta's Llama model on English-language understanding benchmarks, while both Ling-Lite and Ling-Plus surpass DeepSeek's equivalents on Chinese-language benchmarks.
Industry Perspective
Bloomberg Intelligence analyst Robert Lea notes that Ant's paper highlights China's accelerating innovation in AI and suggests the country is progressing toward self-sufficiency in artificial intelligence, developing computationally efficient models to work around export controls on Nvidia chips.
This development contrasts with Nvidia CEO Jensen Huang's strategy, who argues that computational demand will continue to grow even with more efficient models, maintaining that companies will need better chips to generate more revenue rather than cheaper ones to reduce costs.