Jack Ma-backed Ant Group has reportedly achieved a significant AI breakthrough using Chinese chips, developing techniques that reduce AI model training costs by 20%. The move marks a pivotal step in China’s push for AI self-sufficiency, as companies seek alternatives to Nvidia’s powerful but restricted semiconductor technology.
According to sources familiar with the development, Ant Group successfully trained AI models using domestically produced chips, including those from Alibaba Group and Huawei Technologies. The company leveraged the Mixture of Experts (MoE) machine learning approach, an advanced technique that optimizes computing power while maintaining performance.
Despite these advancements, Ant continues to use Nvidia chips, such as the H800, for AI research. However, the firm is increasingly relying on Chinese and AMD-produced alternatives to power its latest models, a move driven in part by US restrictions on exporting high-end Nvidia semiconductors to China.
Competing in the Global AI Race Without Nvidia
As AI research demands massive computational power, companies worldwide are searching for cost-effective training methods. Ant’s successful implementation of MoE-based AI training highlights a growing competition between Chinese and US tech firms, particularly as companies like DeepSeek explore similar strategies to reduce dependence on expensive hardware.
The H800, while not Nvidia’s most advanced processor, remains a powerful AI training chip. However, due to US restrictions, Chinese companies are aggressively working on domestic alternatives that can perform similar tasks at a fraction of the cost.
Ant’s recent research claims that its locally trained models can match or even surpass those built using Nvidia’s hardware. The company’s latest AI breakthroughs could signal a major shift in global AI development, as China continues to challenge the dominance of US chipmakers.
AI Models That Work “Without Premium GPUs”
A key highlight of Ant’s AI research is its focus on scaling large language models (LLMs) without premium GPUs. The company’s research paper—published earlier this month—suggests that its cost-efficient training method outperforms traditional approaches in certain benchmarks.
MoE models have gained traction due to their ability to divide tasks into smaller, specialized data sets, similar to a team of experts working on different sections of a project. Tech giants like Google and China’s DeepSeek have embraced this strategy to optimize AI processing power.
However, MoE-based training typically requires high-performance chips, like Nvidia’s powerful GPUs. Ant Group’s innovation aims to bypass this reliance, offering a more accessible, cost-effective solution for companies working with limited AI budgets.
Ant Group estimates that training one trillion tokens—a core unit of AI learning—using high-performance GPUs costs approximately ¥6.35 million yuan ($880,000). However, by using optimized lower-specification Chinese chips, this cost could drop to ¥5.1 million yuan ($700,000), marking a 20% reduction in expenses.
Ant’s Growing AI Ecosystem
Ant Group is actively expanding its AI ecosystem, integrating its Ling-Plus and Ling-Lite models into various industries, including healthcare and finance. The company’s latest acquisition of Haodf.com, a Chinese online medical platform, reflects its ambition to strengthen AI-driven healthcare services.
Additionally, Ant has developed AI-powered assistants, such as:
- Zhixiaobao – A digital life assistant
- Maxiaocai – A financial advisory AI
Ant’s Ling-Lite model has also demonstrated strong English-language comprehension, outperforming Meta’s Llama models in specific benchmarks. Both Ling-Plus and Ling-Lite models have reportedly outperformed DeepSeek’s equivalents in Chinese-language tasks.
Can China’s AI Challenge the Global Tech Giants?
With Ling-Plus boasting 290 billion parameters, it ranks among the larger AI models currently in development—although it still falls short of DeepSeek-R1’s 671 billion parameters and OpenAI’s rumored 1.8 trillion parameters for GPT-4.5.
Despite these advancements, Ant’s AI training process has faced hurdles, particularly in stability issues. Even minor modifications to the hardware or model structure resulted in fluctuations in error rates. However, these challenges are common in large-scale AI development and are actively being addressed.
Robin Yu, CTO of Shengshang Tech, compared the AI race to martial arts, stating:
“If you find a single weak point to defeat the world’s best kung fu master, you still win. Real-world applications matter more than theoretical superiority.”
The Future of AI Without Nvidia?
Ant Group’s AI breakthrough using Chinese chips is not just a cost-saving measure—it’s a strategic move in the global AI competition. With US-China tech tensions rising, Ant’s success in training powerful AI models without full reliance on Nvidia could serve as a blueprint for future AI development in China and beyond.
As AI continues to evolve, the real question is not whether Chinese chips can replace Nvidia, but how quickly they can close the gap. Ant’s latest innovation brings that future one step closer.
Get the Latest AI News on AI Content Minds Blog