As the AI arms race between China and the U.S. continues to heat up, Ant Group has made a bold move that could shake things up – Yes, the one backed by Jack Ma. According to insiders, the company has been working on training large AI models using Chinese-made semiconductors, and the results? Pretty impressive.
Instead of solely relying on NVIDIA’s powerful (and now export-restricted) H800 GPUs, Ant has been experimenting with chips from Alibaba (its affiliate) and Huawei, applying a technique known as Mixture of Experts (MoE) to train AI models more efficiently. MoE essentially breaks down complex AI tasks into smaller, specialized subtasks – kind of like assembling a team of pros, each focused on a single task to get the job done faster and cheaper.

According to sources familiar with the matter, Ant’s locally trained models achieved results similar to what NVIDIA’s high-end GPUs can deliver, but with a reported 20% cost reduction. And while Ant still uses NVIDIA hardware in its labs, the company is increasingly leaning on AMD and domestic chips for its latest work.
Ant’s new AI models, Ling-Plus and Ling-Lite were trained using this optimized setup. The company claims that in certain English-language benchmarks, Ling-Lite outperformed Meta’s LLaMA model, and both Ling models beat DeepSeek’s offerings on Chinese-language tasks. It’s worth noting that these results haven’t been independently verified yet, but if true, it’s a big win for China’s homegrown AI capabilities.
Ant also open-sourced the Ling models. Ling-Lite comes in at 16.8 billion parameters, while Ling-Plus is significantly larger at 290 billion. For comparison, DeepSeek-R1 boasts 671 billion, and GPT-4.5 is estimated to be rocking a mind-blowing 1.8 trillion parameters.
Cheaper Training, Same Performance?
The research paper released by Ant claims that training a model on 1 trillion tokens cost them ¥6.35 million (around US$880,000) using high-end hardware. But using their optimized, lower-spec setup? That cost drops to ¥5.1 million, which is a massive saving, especially for startups or companies without infinite AI budgets.
Still, it wasn’t all smooth sailing. Ant admitted to facing challenges with stability, where even minor tweaks to the model or hardware could cause error rates to spike.
Ant isn’t just developing these models for bragging rights. It plans to deploy them in industrial applications, particularly in finance and health care. The company even acquired Chinese medical platform Haodf.com earlier this year to boost its health AI offerings. Meanwhile, Ant’s AI chatbot Zhixiaobao and finance advisor bot Maxiaocai are already helping users navigate day-to-day tasks and investment decisions.
A Step Closer to Local AI Independence?
The shift to local chips and cost-efficient training techniques shows how serious Chinese tech firms are about reducing their dependence on U.S. semiconductors, especially with current trade restrictions. While NVIDIA CEO Jensen Huang believes better chips will always be in demand, Ant’s approach seems to be challenging that notion by focusing on efficiency over brute force.
In the end, whether Ant’s approach becomes a new industry standard or just an interesting experiment, one thing’s for sure is that the AI race just got a lot more interesting.
Source: Bloomberg



