The Chinese government is considering measures to restrict overseas access to its advanced artificial intelligence (AI) models. This move is seen as a counter to U.S. export controls, as the competition between the two nations escalates from semiconductors to AI model restrictions. South Korea, lacking a globally competitive independent frontier model, faces challenges due to its dual dependency structure.
According to reports from the IT industry and foreign media on July 13, the Chinese Ministry of Commerce recently held discussions with major domestic tech companies, including Alibaba and ByteDance, regarding restrictions on overseas access to cutting-edge AI models, including those not yet released. The proposed restrictions would apply to both closed-source and open-source models, with discussions also addressing the classification of AI technology leakage and theft as crimes under national security laws. Additionally, there were talks about limiting the entities that can fund Chinese AI startups.
This initiative appears to be a direct response to U.S. control measures. On June 12, the U.S. export control guidelines led to the shutdown of Anthropic's frontier models, 'Preamble 5' and 'Mythos 5,' marking the first global AI model shutdown mandated by government order. Following this, Alibaba prohibited its employees from using 'Claude Code' for work, and Meta ceased operations and data sharing with the Chinese AI startup Manus under directives from Beijing, resulting in a series of retaliatory actions between the two countries.
However, a significant issue is that developers worldwide have become heavily reliant on low-cost Chinese models. An analysis by Bloomberg of data from OpenRouter, the largest AI model brokerage platform, revealed that as of the last week of June, Chinese models accounted for 48% of token processing share, more than double that of the U.S. at 20%. This marks a complete reversal from a year ago when the U.S. held 74% and China 20%. OpenRouter serves as a key indicator of actual model selection in development, providing access to over 400 AI models through a single API. Chinese open-source models are priced 60% to 90% lower than leading U.S. models.
South Korea's situation is more complex, as it relies on high-performance closed models from the U.S. and low-cost open-source models from China. Many domestic startups utilize open-source models from DeepSeek and Q1, meaning that if China restricts access to open-source models, the very foundation of their services could be jeopardized.
Despite this, domestic alternatives still have a long way to go. The government's initiative to develop an 'independent AI foundation model' is currently assessed to be at a global level of around 20th place, according to evaluations by the Ministry of Science and ICT.
LG AI Research, the leading domestic entity, achieved a score of 32 in the global evaluation by Artificial Analysis Intelligence Index, ranking 7th in the world based on open-weight models. However, this is still less than half of the scores for frontier models like Google's 'Gemini 3 Pro' and OpenAI's 'GPT 5.0,' both scoring 73. The actual usage metrics are even less impressive, with the domestic model Upstage's 'Solar Pro 3' processing only a few billion tokens compared to OpenRouter's weekly processing of 25 trillion tokens, indicating a virtually nonexistent presence.
In this context, South Korea's potential card lies in manufacturing data-based physical AI. Physical AI refers to AI that interacts with the physical world, such as robots and equipment. However, this approach faces challenges, including the issue of siloed manufacturing data that is not shared for learning purposes and a reliance on foundational models and GPUs from the U.S. and China, creating a cyclical dependency.
* This article has been translated by AI.
Copyright ⓒ Aju Press All rights reserved.