Ant Group Adopts Domestic Chips for AI Training to Reduce Costs and Dependence on US Tech

In a cost-saving measure for artificial intelligence (AI) model training, Ant Group, the fintech arm of Alibaba, has increasingly turned to domestic Chinese semiconductors to reduce or minimize its reliance on restricted US technology. According to sources familiar with the matter, the company was successful in training large language models (LLMs) on domestic chips sourced from Huawei and in-house suppliers of Alibaba at a performance level comparable to that of LLMs trained using Nvidia H800 GPUs. 

Given that advanced chips have seen the US tighten export controls on such products, these shifts may reflect an ever-growing position of Ant Group within the intensifying AI race of Chinese tech firms against US ones, especially for companies seeking cost-efficient alternatives.

Ant’s Shift to Domestic Chips and the MoE Approach

Ant Group has been experimenting with the Mixture of Experts (MoE) method, an AI training technique that divides tasks into smaller datasets handled by specialized sub-models. This approach, also used by Google and Chinese AI startup DeepSeek, improves efficiency by allowing different components to focus on specific aspects of a problem—much like a team of specialists working together.

The company’s research paper, titled “Scaling Models Without Premium GPUs,” details how Ant optimized its AI training process using lower-specification domestic chips while maintaining performance levels similar to those achieved with Nvidia’s high-end H800 GPUs.

Key Findings from Ant’s Research:

Training one trillion tokens (the basic data units for AI learning) cost 6.35 million yuan (~$880,000) using conventional high-performance hardware.

By optimizing its training methods, Ant reduced costs to 5.1 million yuan (~$700,000)—a 20% reduction—using more affordable domestic chips.

The company’s models, Ling-Lite (16.8 billion parameters) and Ling-Plus (290 billion parameters), demonstrated performance comparable to Meta’s models in some benchmarks.

Why This Move Matters for China’s AI Industry

The US has imposed strict export controls on advanced AI chips, including Nvidia’s A100 and H100 GPUs, forcing Chinese tech firms to seek alternatives. Although the H800 (a downgraded version of the H100) is still available in China, its supply remains uncertain due to geopolitical tensions.

Ant’s success in using domestic chips for AI training could signal a broader shift in China’s tech ecosystem:

• Reduced Dependence on US Tech: By proving that Chinese chips can deliver competitive results, Ant is paving the way for other firms to follow suit.
• Lower Costs for AI Development: Smaller companies that cannot afford Nvidia’s premium GPUs may benefit from cost-effective domestic alternatives.
• Encouraging Local Semiconductor Innovation: Huawei’s Ascend chips and Alibaba’s in-house Hanguang processors are emerging as viable substitutes, boosting China’s self-reliance in AI hardware.

Challenges and Limitations

Despite the progress, Ant’s research paper acknowledges that training AI models on domestic hardware is not without challenges:

• Performance Instability: Small adjustments in hardware or model structure sometimes led to erratic performance, including sudden spikes in error rates.
• Scalability Concerns: While MoE improves efficiency, scaling these models to match the capabilities of industry leaders like OpenAI’s GPT-4.5 (estimated at 1.8 trillion parameters) remains difficult.
• Ongoing Reliance on Nvidia: Ant still uses Nvidia GPUs for some AI tasks, indicating that a full transition to domestic chips may take time.

Ant’s AI Ambitions: Healthcare, Finance, and Beyond

Ant Group is not just focusing on cutting costs—it’s also expanding its AI applications in real-world industries:

• Healthcare: Earlier this year, Ant acquired Haodf.com, a leading Chinese online medical platform, to integrate AI-driven diagnostics and patient care solutions.
• Finance: The company operates AI-powered financial services, including Zhi Xiao Bao (a virtual assistant) and Maxiaocai (a financial advisory platform).
• Open-Source AI Models: By making Ling-Lite and Ling-Plus open-source, Ant aims to foster collaboration and accelerate AI adoption across industries.

Nvidia’s Counter-Argument: The Need for More Power

Ant, while pretty much talking about cost-effective solutions, Jensen Huang, CEO of Nvidia, affirms the ever-increasing demand for computing power irrespective of any efficiency improvements. He states that businesses will go for the more powerful chips (like Nvidia’s incoming Blackwell GPUs) to make money, rather than compromise on cheaper setups that just don’t cut it anymore.

This divergence in strategy highlights a key debate in AI development:

• Ant’s View: Optimize models and hardware to reduce costs and reliance on foreign tech.
• Nvidia’s View: Push for more advanced GPUs to handle increasingly complex AI workloads.

What’s Next for China’s AI Chip Industry?

Ant Group’s experiment with domestic chips is just one part of China’s broader push for semiconductor independence. Other key developments include:

• Huawei’s Ascend AI Chips: Already being used by major firms like iFlytek and Baidu.
• Biren Technology and Moore Threads: Chinese GPU startups working on competitive alternatives to Nvidia.
• Government Support: Beijing’s “Big Fund” continues to invest heavily in domestic chip manufacturing.

Expert Insight

“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,” said Robin Yu, CTO of Beijing-based AI firm Shengshang Tech. This philosophy aligns with Ant’s strategy of focusing on practical, cost-effective AI solutions rather than purely chasing raw computing power.

Conclusion

First and foremost, the move by Ant Group to develop AI chips on the home turf underscores the importance this development will have in China’s momentum toward technological self-sufficiency. As cost-efficient and performance-proficient as the company has succeeded in making its products, this phenomenon may also encourage adoption across more Chinese tech entities. Significantly, innovations such as the one presented by Ant in its MoE-based training would prove vital in shaping future development in global AI, especially as tensions rouse between China and the USA.