Jack Ma-backed Ant Group has reportedly achieved a significant milestone in artificial intelligence by training models with Chinese-made semiconductors, slashing costs by 20%. Sources familiar with the development revealed that the company utilized domestic chips from Alibaba Group and Huawei Technologies to implement a Mixture of Experts (MoE) machine learning approach. This technique delivered results comparable to Nvidia’s H800 chips, despite the ongoing U.S. export restrictions on such advanced hardware.
While Ant continues to use Nvidia GPUs for AI development, the company is now leaning more on alternatives, including Advanced Micro Devices (AMD) chips and Chinese processors, for its latest models. This shift reflects a broader effort by Chinese companies to reduce dependency on American technology, especially as the AI race between China and the U.S. heats up.
The breakthrough highlights China’s growing focus on leveraging local semiconductor capabilities. Ant’s models reportedly achieved comparable performance levels while cutting training expenses significantly — a promising development amid soaring costs for AI model training. According to Ant, training one trillion tokens cost 6.35 million yuan ($880,000) with high-performance hardware, but the optimized method lowered the cost to 5.1 million yuan using less powerful chips.
This progress was detailed in a research paper published by Ant this month. The paper claimed that in some benchmarks, its models even outperformed Meta’s Llama models, though these results have not been independently verified. If accurate, Ant’s achievements could accelerate China’s AI ambitions by offering a more cost-effective path to large-scale model training and deployment.
The company’s Ling-Plus and Ling-Lite large language models (LLMs) form the core of this advancement. Ling-Plus boasts 290 billion parameters, while Ling-Lite operates with 16.8 billion parameters. For comparison, OpenAI’s GPT-4.5 reportedly runs on 1.8 trillion parameters. Despite their smaller size, Ant’s models outperformed DeepSeek’s equivalents on Chinese-language tasks and even surpassed Meta’s Llama model on English-language benchmarks in specific tests.
MoE models like Ant’s have gained traction in the AI community for their efficiency. By dividing tasks into smaller data sets — much like assigning tasks to specialized teams — MoE models reduce computational demands while maintaining strong performance. Google and Chinese startup DeepSeek have both embraced this approach, which is becoming popular as companies seek to balance AI capabilities with mounting costs.
Ant’s research pushes back against Nvidia CEO Jensen Huang’s prediction that demand for powerful GPUs will continue rising. Huang argues that as AI models grow more efficient, businesses will still need better chips to unlock new revenue streams rather than just seeking cheaper alternatives. Yet, Ant’s progress suggests otherwise — efficient models using domestic chips may offer a viable path forward.
Bloomberg Intelligence analysts view Ant’s breakthrough as a sign of accelerating innovation in China’s AI sector. If the claims hold, this could move China closer to AI self-sufficiency, especially as it navigates U.S. export restrictions.
Ant is already applying its AI technology across industries. Earlier this year, it acquired health platform Haodf.com to bolster its AI healthcare services. The company also developed an AI Doctor Assistant designed to help Haodf’s 290,000 doctors manage medical records and other tasks.
Additionally, Ant has launched several AI-powered services. Its AI “life assistant” app Zhixiaobao and the financial advisory platform Maxiaocai are already live. In the healthcare sector, Ant rolled out two medical AI agents — Angel, now supporting over 1,000 medical facilities, and Yibaoer, which provides assistance with medical insurance. Last year, Ant introduced an AI Healthcare Manager within its Alipay app, expanding its AI footprint in digital health.
Ant further revealed that it has built large healthcare-specific models used by seven hospitals and healthcare providers in cities like Beijing and Shanghai. These models integrate DeepSeek R1, Alibaba’s Qwen, and Ant’s proprietary LLM to deliver medical consultancy services. However, the company acknowledged challenges during development, particularly around model stability, where even minor hardware or structural changes led to error spikes.
Ant’s open-source release of the Ling models signals a push to make AI development more accessible while positioning itself as a leading player in China’s fast-growing AI industry. This strategy could unlock new applications in healthcare, finance, and beyond — all while reducing reliance on high-cost, foreign-made chips.
