Ant Group, the financial technology company backed by Jack Ma, has achieved a significant breakthrough in reducing artificial
intelligence training costs through the use of Chinese-manufactured semiconductors. The company has successfully implemented a combination of domestic and American chips to train AI models, resulting in a 20% reduction in costs while maintaining efficiency levels comparable to Nvidia’s H800 chips.
The Hangzhou-based company utilized semiconductors produced by Chinese tech giants Alibaba Group Holding and Huawei Technologies, alongside their existing Nvidia infrastructure. While Ant Group continues to employ Nvidia components for AI development, it has shifted
predominantly toward alternatives, including chips from Advanced Micro Devices Inc. and Chinese manufacturers for its latest model
iterations.
Central to this cost-reduction achievement is Ant’s implementation of the “Mixture of Experts” (MoE) machine learning methodology. This approach divides complex tasks into smaller, more manageable data sets, similar to assigning specialized teams to handle specific aspects of a larger project. The technique has gained prominence in the AI industry, with major players like Google and Hangzhou-based startup DeepSeek adopting similar strategies.
Ant Group recently published a research paper titled “EVERY FLOP COUNTS: SCALING A 300B MIXTURE-OF-EXPERTS LING LLM WITHOUT PREMIUM GPUS,” which details their progress in addressing cost inefficiencies and resource constraints in AI model training. This development follows China’s recent launch of DeepSeek, a domestic alternative to Western language models like GPT and Llama 3.1, which has demonstrated significantly reduced training and computing power requirements.
The advancement is particularly noteworthy given the context of U.S. restrictions on China’s access to AI chip procurement. Bloomberg Intelligence analyst Robert Lea emphasized the significance of Ant’s achievement, suggesting it demonstrates China’s progress toward technological self-sufficiency in the AI sector. The country’s ability to develop computationally efficient models serves as a strategic response to export controls on Nvidia chips.
Building on this momentum, Ant Group has announced substantial improvements to its healthcare AI offerings, leveraging its
cost-effective AI model. This development highlights China’s growing capability to produce AI models at a fraction of the investment required by Western tech giants like OpenAI and Alphabet, which have poured tens of billions into their AI initiatives.
The use of domestic semiconductors for AI training represents a significant step forward in China’s technological independence, particularly as the country faces increasing restrictions on access to advanced Western technology. The success of Ant Group’s approach, combining domestic and international components while implementing innovative training methodologies, suggests a viable path forward for Chinese AI development despite external constraints.
The MoE technique’s effectiveness in reducing computing power requirements while maintaining performance standards comparable to industry-leading chips demonstrates the potential for alternative approaches to AI model training. This development could have far-reaching implications for the global AI industry, potentially offering more cost-effective solutions for AI development while challenging the current paradigm of resource-intensive training methods.
As companies worldwide continue to invest heavily in AI development, Ant Group’s achievement in cost reduction while maintaining
performance standards could influence future approaches to AI model training and development. The success of this initiative also highlights the growing technological capabilities of Chinese firms in the face of international restrictions, suggesting a shift in the global AI development landscape.