On October 20, 2026, a private think tank called Rockbridge Network Korea published a report titled How Chinese Artificial Intelligence Is Being Used in South Korea. The finding that traveled fastest was a single percentage: of the 1,845 AI models published on Hugging Face by 1,993 South Korean accounts between January and August 2026, 49.3 percent were built on Chinese foundation models.
The breakdown was more striking than the headline. Alibaba's Qwen alone accounted for 864 models, or 46.8 percent of the total. Google's Gemma followed at 271 models (14.7 percent), Meta's Llama at 254 (13.8 percent). Chinese models overall — Qwen plus DeepSeek and others — accounted for nearly half. American models accounted for 35.9 percent. South Korean models accounted for 14.9 percent.
The trend line was steeper than the snapshot. Chinese foundation models accounted for 1.2 percent of Korean derivatives in 2023, 7.2 percent in 2024, and 22.0 percent in 2025. In eight months of 2026, that figure reached 49.3 percent. Over the same period, the American share fell from 79.5 percent to 35.9 percent.
South Korea has a "sovereign AI" initiative. It has a national project to build an independent frontier model. Its science minister has said frontier AI models could become a strategic asset "more like a nuclear weapon." And yet, in the work that Korean developers actually publish, the dominant foundation is Chinese.
The Developers' Choice
The Rockbridge report contains a second finding that complicates the first. When the same researchers surveyed 102 small and mid-sized Korean companies about whether they used Chinese AI services, not a single company said yes. In the rankings of AI chatbot apps by monthly active users in Korea as of July 2026, no Chinese service appeared in the top ten.
The 49.3 percent figure describes Korean developers building on Chinese foundations, not Korean consumers using Chinese chatbots.
The reasons are practical. Qwen offers a range of model sizes, strong multilingual support, and a mature toolchain. Korean developers fine-tune it with Korean-language corpora and deploy the result in chatbots, industry Q&A systems, content generation, and enterprise knowledge bases.
The applications are not marginal. SK Telecom, one of Korea's three major telecommunications carriers, released its A.X 4.0 large language model built on Qwen. ThakiCloud, a Korean AI company, released a 27-billion-parameter model called MerchantDesk on a Qwen base, designed to handle customer inquiries for small restaurants. It also released KoToolAct-4B, a 4-billion-parameter model built for tool execution in Korean enterprise workflows.
The pattern is consistent with how open-weight models diffuse globally. Developers choose the foundation with the best capability per unit of cost and effort. In 2026, that foundation is frequently Chinese.
The Sovereign AI Paradox
The tension between the developer data and the government's stated policy is visible in the government's own infrastructure.
South Korea's "sovereign AI" push includes a project called the Government-wide Ultra-large AI Common Infrastructure, which allows central and local agencies to share AI models and GPU resources. It operates on two platforms: Samsung SDS's "Fabrics" and Naver Cloud's "Clova Studio for Gov."
According to data submitted to the National Assembly, as of September 2026, of the 31 AI models introduced on the government platform, only 12 were domestic and 19 were foreign. Of the 11 models already deployed, seven were foreign.
One of those foreign models was Bge-M3, developed by the Beijing Academy of Artificial Intelligence (BAAI). BAAI was added to the U.S. Department of Commerce's Entity List in March 2025. The model was running on a South Korean government platform built to reduce dependence on foreign AI.
The Ministry of Science and ICT acknowledged that there were no verification criteria to assess the countries or institutions behind the AI models during the platform's development. Each platform provider selected models based on "universality, search performance, multilingual support, and usability." The government said it did not restrict models from specific countries.
The same pattern appears outside the government. Korea Investment Corporation (KIC), a sovereign wealth fund managing 339 trillion won, adopted Chinese-made models as core components of its internal AI, according to reporting by Edaily. The models included an Alibaba language model and a search model from a Chinese research institute on the U.S. export control list.
The sovereign AI initiative has its own contradictions. The Independent AI Foundation Model project — known in Korean as Dokpamo — selected LG AI Research, SK Telecom, and Upstage as the three finalists in its competition to build a homegrown frontier model. Naver Cloud was eliminated in the first round in January 2026, in part because its model used a vision encoder from Alibaba's Qwen. The evaluation criteria required originality, and Naver's use of the Qwen weights did not meet the standard.
Naver, the company that built Korea's most widely used domestic search and AI stack, was excluded from the sovereign AI project for using a Chinese component. The government's own platform was running a model from an entity on the U.S. Entity List.
Washington's Message
The United States has been watching the Korean AI ecosystem with increasing attention, and the message it is sending has become more explicit.
In June 2026, Korean media reported that the U.S. decision to impose export controls on Anthropic's Fable 5 and Mythos 5 models was partly driven by concern that China was attempting to access American frontier models through South Korea. The White House believed Beijing had made the acquisition of U.S. AI model information a national task, and that small and mid-sized Korean AI infrastructure companies were targets.
In August 2026, Reuters reported that the U.S. State Department was preparing to send letters to 35 countries, including South Korea, that had signed the Pax Silica joint statement on AI supply chains. The draft letter stated that "seeking to belong everywhere is the same as belonging nowhere," and that recipients "cannot simultaneously maintain membership in initiatives that conflict with our expectations." A senior U.S. official put it more directly: "You cannot benefit from both sides."
China had launched its own World Artificial Intelligence Cooperation Organization (WAICO) in July 2026, and Seoul was reportedly under pressure from Beijing to join. The U.S. letter was designed to make that choice explicit.
The pressure extends beyond models to cloud access. Korean media reported that if Washington imposes controls on cloud access to advanced AI compute, it could require allies like South Korea to participate in the export control regime. For a country whose AI developers depend on both American and Chinese foundations, that requirement would force a choice the government has been trying to avoid.
The Double Dependency
South Korea's science minister, Bae Kyung-hoon, has described the country's position in unusually direct terms. Speaking at a conference in July 2026, he said frontier AI models are becoming "more like a nuclear weapon," and that "the tendency for certain countries to seek control of AI models of this caliber will only grow stronger."
He also acknowledged the structural problem. Korea relies on the United States for security and advanced semiconductors, and on China for a growing share of the AI foundations its developers use. Korean media has described this as a "double dependency" — a position that becomes more difficult to maintain as Washington and Beijing both demand alignment.
Bae said China was "strongly hoping" South Korea would join its AI cooperation network, but that Seoul had to consider its "many relationships with the U.S. government." He said Korea needed to "strengthen independent AI capabilities" while finding the best approach. The government's plan is to build a frontier model through the Dokpamo project and to become a "full-stack" AI provider to other countries, offering models, services, and semiconductors as a package.
The 49.3 percent figure measures how far that plan has to travel. The developer data shows where the market has already gone.
What the Numbers Mean
The Korean case illustrates the gap between what governments say and what developers do, rather than a country being forced to use Chinese AI.
The Korean government has built a sovereign AI initiative, a frontier model competition, and a government platform intended to reduce foreign dependence. The developers in that same country have built nearly half of their new models on Chinese foundations. The government's own platform runs foreign models, including one from a U.S.-blacklisted Chinese institute. The sovereign wealth fund uses Chinese models for internal AI.
Washington is now asking Seoul to choose. The Pax Silica letter, the cloud control discussions, and the concerns about model leakage through Korean infrastructure all point in the same direction: the United States wants its allies to treat Chinese AI as a supply-chain risk, not a development tool.
The Korean developers who built 864 Qwen derivatives made a different calculation. They chose the foundation that worked.
The tension between those two positions is the central fact of South Korea's AI policy in 2026. The government can build a sovereign model. It can sign Pax Silica. It can issue guidance against foreign components. It cannot easily tell thousands of developers to stop using the foundation that gives them the best results at the lowest cost — especially when the government's own platforms are running the same kinds of models.
The 49.3 percent measures how open-weight diffusion works in practice. The foundation model is not a product that a government can procure and control; it is a tool that developers choose. In 2026, a growing number of Korean developers are choosing Chinese tools.
Sources: Yonhap News (October 20, 2026); Korea Herald (October 21, 2026); Korea Times (October 25, 2026); Chosun Ilbo (October 20, 2026); Chosun Ilbo English (October 5–6, 2026); Yonhap News English (July 20, 2026); Seoul Economic Daily English (July 20, 2026; August 15, 2026); Asia Economy (June 21, 2026; July 13, 2026); Kyunghyang Shinmun (August 20, 2026); Edaily (October 6, 2026; October 19, 2026); KBS News (January 15, 2026); UPI (August 18, 2026); CCIDnet (October 26, 2026); ThakiCloud (September 28, 2026; October 19, 2026); Adnkronos (January 21, 2026); Global Times (October 22, 2026); Sina Finance (October 21, 2026).
Disclaimer
The information provided in this article is for general informational and educational purposes only. It does not constitute legal, financial, or professional advice. The author and publisher are not responsible for any actions taken based on the content of this article. Readers should consult qualified professionals for advice specific to their situation. All trademarks and references to third-party products, services, or organizations are the property of their respective owners. The performance data and benchmarks discussed are based on specific research studies and may not generalize to all use cases or environments. As of the publication date, the AI landscape continues to evolve rapidly, and readers should verify current information independently.
Limitations
The dataset counts models published on Hugging Face by Korean accounts from January to August 2026, not all Korean AI activity.
A model built on a Chinese foundation may be heavily fine-tuned; base-model origin does not describe the finished system.
The report was published by a private think tank, and its selection criteria for Korean accounts are not fully disclosed.
Korean government procurement rules described here were still being drafted at the time of writing.
Reporting standards differ between Chinese and Western outlets, and figures cited here reflect the primary sources as published.
Market size and shipment estimates vary by methodology; where estimates conflict, the source used is named in the text.
The sectors discussed are changing quickly, and some figures may be revised after publication.
Sources
- Yonhap News (October 20, 2026)
- Korea Herald (October 21, 2026)
- Korea Times (October 25, 2026)
- Chosun Ilbo (October 20, 2026)
- Chosun Ilbo English (October 5–6, 2026)
- Yonhap News English (July 20, 2026)
- Seoul Economic Daily English (July 20, 2026; August 15, 2026)
- Asia Economy (June 21, 2026; July 13, 2026)
- Kyunghyang Shinmun (August 20, 2026)
- Edaily (October 6, 2026; October 19, 2026)
- KBS News (January 15, 2026)
- UPI (August 18, 2026)
- CCIDnet (October 26, 2026)
- ThakiCloud (September 28, 2026; October 19, 2026)
- Adnkronos (January 21, 2026)
- Global Times (October 22, 2026)
- Sina Finance (October 21, 2026).
The information provided in this article is for general informational and educational purposes only. It does not constitute legal, financial, or professional advice. The author and publisher are not responsible for any actions taken based on the content of this article. Readers should consult qualified professionals for advice specific to their situation. All trademarks and references to third-party products, services, or organizations are the property of their respective owners. The performance data and benchmarks discussed are based on specific research studies and may not generalize to all use cases or environments. As of the publication date, the AI landscape continues to evolve rapidly, and readers should verify current information independently.
Limitations: The dataset counts models published on Hugging Face by Korean accounts from January to August 2026, not all Korean AI activity.; A model built on a Chinese foundation may be heavily fine-tuned; base-model origin does not describe the finished system.; The report was published by a private think tank, and its selection criteria for Korean accounts are not fully disclosed.; Korean government procurement rules described here were still being drafted at the time of writing.; Reporting standards differ between Chinese and Western outlets, and figures cited here reflect the primary sources as published.; Market size and shipment estimates vary by methodology; where estimates conflict, the source used is named in the text.; The sectors discussed are changing quickly, and some figures may be revised after publication.