Chinese AI Takeover? Half of South Korea's New Models Built on Chinese Tech
A new study reveals nearly 50% of South Korean AI model derivatives are built on Chinese technology, primarily Alibaba's Qwen.
South Korea's ambition to lead the global artificial intelligence race is facing a surprising reality check. A recent study has revealed that nearly half of the AI models released by South Korean developers this year are actually built upon Chinese technology, sparking a heated debate over the country's "sovereign AI" goals.
The findings, published by the private think tank Rockbridge Network Korea and authored by Hanyang University business professor Choi Seong-jin, highlight a dramatic shift in the technological landscape. Between January and August of this year, the global model-sharing platform Hugging Face saw 1,845 AI model derivatives published by Korea-linked accounts. A staggering 49.3 per cent of these were based on Chinese foundation models.
The Rapid Rise of Alibaba's Qwen
The surge in Chinese AI adoption has been meteoric. In 2023, Chinese models held a mere 1.2 per cent share of the market, which grew to 22 per cent last year, before skyrocketing to nearly 50 per cent in 2024. This growth comes at the expense of US dominance, which plummeted from a 79.5 per cent share in 2023.
The primary driver of this trend is Alibaba's Qwen. This single model accounted for 864 derivatives—roughly 46.8 per cent of the total. To put this in perspective, Qwen far outpaced major US competitors like Google's Gemma (14.7 per cent) and Meta's Llama (13.8 per cent).
How Korean Homegrown Models Compare
While South Korea has its own tech giants, their foundation models are struggling to keep pace with the adoption of foreign tech. Korean-developed models accounted for only 14.9 per cent of new releases. Among them:
- LG AI Research's Exaone: Ranked fourth with 171 derivatives (9.3%).
- Kakao's Kanana: Ranked sixth with 49 derivatives (2.7%).
Combined, these homegrown efforts represent less than one-third of the popularity of Alibaba's Qwen alone.
The "Sovereign AI" Struggle
These statistics cast a shadow over the South Korean government's sovereign AI program. The initiative has poured approximately 200 billion won (USD 151 million) into developing homegrown alternatives to ensure technological independence.
The stakes are high, and the rules are strict. Earlier this year, Naver Cloud was eliminated from the program after it was discovered that the company had used components from Alibaba's Qwen model, failing the program's independence requirements. Currently, LG AI Research, SK Telecom, and the startup Upstage have moved forward to the next stage of the government-backed project.
Why Developers are Choosing Chinese AI
The attraction to Chinese models isn't just about performance; it's about the bottom line and flexibility. An official from a Korean AI startup noted that Chinese models like Qwen and DeepSeek are the preferred choice for services with high user traffic due to lower operating costs.
Beyond price, the report identifies several key advantages of Qwen:
- Versatility: A wide array of model sizes to suit different needs.
- Language Support: Compatibility with 119 different languages.
- Developer Friendly: Readily available tools that simplify the customization process.
The Hidden Integration and Security Fears
Interestingly, this trend is largely invisible in standard consumer data. In July, no Chinese chatbot appeared in the top 10 most-used AI apps in Korea. Furthermore, a scan of 1,900 public GitHub repositories from 36 major Korean firms showed no mention of Chinese models.
However, the report warns that this is a misleading picture. Developers can download these models and run them locally without relying on the Chinese companies' online services, making the usage "invisible" to traditional tracking.
This "hidden" usage brings significant concerns to the forefront. In a survey of 102 small and medium-sized businesses, 80.2 per cent expressed fear over potential data leaks, while 31.3 per cent were concerned about the actual quality and performance of the models.
"For Korean open models to gain traction among developers, performance alone is not enough," wrote Professor Choi. "They also need commercially viable licenses, a range of model sizes and the tools developers need to customize them."
The report concludes with a warning: over-reliance on a single foreign provider leaves Korean developers vulnerable to sudden changes in licensing terms or the cessation of model updates, potentially crippling the local AI ecosystem.