Ant Group, the fintech powerhouse backed by Jack Ma, has made significant strides in artificial intelligence by dramatically reducing the cost of training AI models—by nearly 20%—using Chinese-made chips. This breakthrough is seen as a bold step toward technological self-sufficiency amid U.S. sanctions on advanced semiconductor exports.
Homegrown Chips Replace Nvidia’s Powerhouse
Instead of relying solely on Nvidia’s H800 GPUs, which face export restrictions to China, Ant tapped into chips developed by domestic giants Huawei and Alibaba (an Ant affiliate). These alternative chips, paired with innovative AI modeling strategies, deliver comparable performance at a fraction of the cost.
Ant’s training process for 1 trillion AI tokens—a basic unit of data processing—normally costs about 6.35 million yuan ($880,000). With the new system, that figure drops to 5.1 million yuan ($710,000), signaling a cost-saving strategy that could reshape how AI is developed across the globe.
The MoE Strategy: Division of Intelligence
The cost reduction is largely credited to the Mixture of Experts (MoE) machine learning approach. MoE breaks complex AI tasks into smaller, specialized components—similar to assigning various experts to specific problems, rather than overburdening a single large system.
Robin Yu, CTO of Shengshang Tech Co., described the approach metaphorically:
“If you find one point of attack to beat the world’s best kung fu master, you can still say you beat them. That’s why real-world application is key.”
Ling-Lite and Ling-Plus: China’s New AI Contenders
Ant has unveiled two homegrown language models:
- Ling-Lite – A 16.8 billion parameter model that reportedly surpasses Meta’s Llama on certain English tasks
- Ling-Plus – A 290 billion parameter model that performs competitively on Chinese-language benchmarks
Though smaller in scale compared to GPT-4.5’s estimated 1.8 trillion parameters or DeepSeek-R1’s 671 billion, these models have shown powerful performance, especially in Chinese-language tasks. Importantly, Ant has open-sourced both models, allowing broader community testing and feedback.
Shifting Toward Chinese and AMD Chips
While Nvidia GPUs remain part of Ant’s AI infrastructure, the company is pivoting to more cost-effective and unrestricted options like AMD and domestic chips. This shift reflects a growing trend among Chinese firms prioritizing national production in response to foreign tech restrictions.
Bloomberg Intelligence’s Robert Lea remarked:
“If these claims hold, China is well on the way to becoming self-sufficient in AI by developing computationally efficient models that circumvent U.S. export rules.”
Real-World Deployments: From Finance to Healthcare
Ant’s innovations aren’t confined to labs. Their AI is already in practical use across several industries:
- Healthcare: Ant acquired Haodf.com, an online healthcare platform, and launched an AI Doctor Assistant used by over 290,000 doctors for managing records.
- Finance: A digital advisory service named “Maxiaocai” guides users on investments and budgeting.
- Consumer Tech: The company’s “Zhixiaobao” app acts as a smart life assistant for day-to-day queries.
AI systems built by Ant are operational in seven major hospitals in Beijing, Shanghai, and other key cities, reflecting real-world impact beyond prototype stages.
Nvidia’s Model Questioned
Ant’s success presents a quiet challenge to Nvidia’s strategy. CEO Jensen Huang has maintained that AI models will continue demanding more computing power, arguing for increasingly advanced (and expensive) chips.
Ant’s success using lower-spec chips and efficient architecture raises questions: is more power always better—or can smarter, leaner models deliver more value?
Still, Ant acknowledges hurdles. Training stability remains a concern; small changes in hardware or model configuration have caused unexpected spikes in error rates. But the progress shows that efficiency and domestic innovation can carve a competitive path forward in AI.
