Ant Group Uses Chinese Chips to Cut AI Training Costs by 20%

Ant Group, backed by billionaire Jack Ma, has reportedly developed a cost-saving method to train large AI models using Chinese-made chips—achieving comparable results to Nvidia hardware while reducing expenses by up to 20%. The move positions Ant more firmly in China's race to build sovereign AI capabilities amid U.S. chip restrictions.

According to sources familiar with the effort, Ant trained its latest large language models (LLMs) using domestic semiconductors from companies like Alibaba and Huawei, employing the Mixture of Experts (MoE) architecture. While still using some Nvidia GPUs, Ant is now primarily relying on AMD and Chinese chips for development.

Local Chips, Global Stakes

This advancement reflects China’s accelerating push for chip independence, as U.S. export controls continue to block access to high-end Nvidia processors like the H800. While the H800 isn't Nvidia's most advanced chip, it remains a powerful GPU—and it is now off-limits to Chinese buyers.

As global AI investment intensifies, Mixture of Experts (MoE) models have gained traction, adopted by leaders like Google and DeepSeek. This technique divides tasks into smaller components and routes them to specialized “expert” subnetworks—much like assigning tasks to a team of specialists. The result is greater efficiency, lower computational demands, and improved scalability—even on lower-spec hardware. Ant’s models are built on this MoE architecture, enabling their cost-effective performance.

This stands in contrast to Nvidia’s strategy. CEO Jensen Huang has maintained that compute demand will continue to grow—even with the rise of more efficient models like DeepSeek’s R1—arguing that companies will require increasingly powerful chips to drive revenue, not cheaper alternatives to cut costs. Nvidia has stayed focused on building larger GPUs with more cores, transistors, and expanded memory. Ant’s approach, by contrast, prioritizes optimization and affordability over raw power.

Despite using lower-cost hardware, Ant’s models reportedly deliver results comparable to Nvidia-based systems. According to the company, its optimized training methods reduce the cost of processing 1 trillion tokens from 6.35 million yuan ($880,000) to 5.1 million yuan ($707,000)—a potential breakthrough for companies constrained by compute budgets.

Ant acknowledged technical hurdles, including training instability caused by small changes in hardware or model architecture. Still, the outcome signals growing maturity in China’s AI engineering capabilities. Ling Models and Real-World Application

Ant's efforts have yielded two key models:

• Ling-Plus: 290 billion parameters

• Ling-Lite: 16.8 billion parameters

For context, experts estimate ChatGPT’s GPT-4.5 has around 1.8 trillion parameters, according to MIT Technology Review, while DeepSeek-R1 comes in at 671 billion.

Both models reportedly outperform DeepSeek’s equivalents in Chinese-language benchmarks, while Ling-Lite also beat a Meta LLaMA model on English-language tasks in at least one key benchmark, according to Ant’s paper.

“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.” — Robin Yu, CTO, Shengshang Tech

The Ling models are open source and are already being applied across sectors. In healthcare, Ant is deploying its LLMs to support hospitals and doctors, including through:

• AI Doctor Assistant for medical record support

• AI "Life Assistant" app, Zhixiaobao

• Maxiaocai, an AI financial advisory service

• Angel and Yibaoer, two healthcare AI agents that specialize in medical facilities and medical insurance services, respectively

• AI Healthcare Manager, launched via Alipay

• Partnerships with hospitals in Beijing, Shanghai, and beyond

What This Means

Ant Group's AI milestone represents more than a technical win—it signals a deepening shift in AI development strategy. By proving that high-quality models can be trained on domestic hardware, Ant is helping China move toward AI self-sufficiency, challenging the West’s dominance in compute-heavy AI systems.

As open-source models like Ling-Plus and Ling-Lite enter real-world deployment in health care and finance, Ant is quietly building a robust AI ecosystem designed to thrive without access to restricted hardware. That may reshape the future landscape of enterprise AI—where cost efficiency, adaptability, and local infrastructure matter just as much as raw model size.

While OpenAI continues to pursue ever-larger models trained on elite hardware behind closed APIs, Ant's strategy highlights a more decentralized, accessible path forward—one that could prove just as influential in scaling AI across real-world industries.

Editor’s Note: This article was created by Alicia Shapiro, CMO of AiNews.com, with writing, image, and idea-generation support from ChatGPT, an AI assistant. However, the final perspective and editorial choices are solely Alicia Shapiro’s. Special thanks to ChatGPT for assistance with research and editorial support in crafting this article.