The global artificial intelligence (AI) power struggle is reaching its peak. The United States, led by companies like NVIDIA and AMD, has secured AI chipsets and weaponized them to dominate infrastructure, while China is enhancing AI efficiency based on its vast data resources.
According to recent analyses from Bloomberg Intelligence and others, the benchmark score gap between China's top AI model and the U.S. model has narrowed to within 3% (81.1 points vs. 83.4 points). This gap has decreased from 15% at the beginning of the year and 9% in the first half of the year. If this trend continues, forecasts suggest that by next year, Chinese AI models may achieve performance levels comparable to their U.S. counterparts.
However, it is not accurate to conclude that China has fully caught up with the U.S. based solely on the 3% figure. The score difference based on live benchmarks is only 2.8%, while the Artificial Analysis Intelligence Index (v4.3) shows a significant gap of about 16%, with the U.S. scoring 53.4 and China 44.9. In expert-level hacking tasks conducted by the National Institute of Standards and Technology (NIST) under CAISI, U.S. models achieved a success rate of 71%, compared to just 32% for Chinese models, indicating a performance difference of more than double.
In preference voting or simple evaluations, the differences appear small. However, in complex reasoning and high-level areas such as cybersecurity attacks, the gap remains significant. When various assessments, including simple benchmarks, are considered, China still lags behind the U.S. by more than six months. Notably, the dynamics of competition are striking. When the U.S. takes a step forward, China leaps five steps ahead. As the U.S. begins to run, China accelerates into a full sprint. While gaps still exist, the distance between the two countries continues to narrow.
In terms of infrastructure, China is undergoing significant changes. Following U.S. semiconductor export controls, China has been training AI models using older chipsets like the NVIDIA H800 that it had previously secured. There have also been efforts to develop domestic alternatives, such as the Huawei Ascend 950 PR/DT chips. To compensate for insufficient computing power, China has utilized quantity, bundling thousands of chips together and providing ample power to achieve computational capabilities similar to the latest models.
On the software front, China has moved away from NVIDIA's AI chipset platform, CUDA. While AI companies worldwide have few alternatives to NVIDIA, China is creating its own infrastructure ecosystem without it, which could be economically advantageous.
The rise of China's AI is also driven by price disruption and the expansion of open-source ecosystems. The U.S. sells high-performance closed models like GPT and Claude for $25 to $50 per million tokens, while China offers models like GLM-5.3 for $4.40 and DeepSeek V4.1 Flash for $1.20, at less than one-tenth the price. Given that not all tasks delegated to AI require high computational power, the structure favors the widespread use of Chinese models.
This is reflected in the fact that six out of the top ten models on the global AI intermediary platform OpenRouter are Chinese. DeepSeek's weekly request share has reached 21.8%, surpassing Google and OpenAI. Although there are concerns that China's AI industry may struggle to turn a profit by 2030, its near-free subsidy strategy is rapidly attracting global users and data.
The U.S. monopolizes 'cutting-edge infrastructure' with advanced semiconductors (NVIDIA, AMD), supercomputing data centers, and the capital power of global tech giants. Meanwhile, China is leveraging its 1.4 billion population and state-led data collection to pursue vast learning data and an open-source ecosystem, despite supply constraints.
South Korea struggles to match the absolute scale of these two giants in both infrastructure and data. It lacks the capital and computing resources to compete in the U.S.-style GPU cluster construction race, and its data volume is inherently limited compared to China's large-scale data collection structure and single-language population base. Competing on the same level as the U.S. and China is realistically challenging.
So where should South Korea seek answers? As the paradigm of AI competition shifts from basic technology development to 'real-world application and optimization (fine-tuning),' South Korea's competitive edge is expected to become more visible.
South Korea boasts a 99.97% internet access rate per household, a 94.5% internet usage rate among citizens aged three and older, and a smartphone penetration rate exceeding 97%. It is a prime example of a highly advanced information society. The population is not only connected but also familiar with digital services based on smartphones and PCs. The experience rate of generative AI services like ChatGPT is also rapidly increasing.
All sectors, including manufacturing, finance, healthcare, education, and public administration, are highly digitized. High-quality, high-density 'real-life private data' is continuously generated. While China's data is characterized by its 'quantitative vastness,' South Korea's data is generated by a digitally literate public, resulting in 'high-quality optimized data.' There are few testing grounds globally that surpass South Korea in precise field testing and industry-specific specialized AI implementation.
ChatGPT and Claude are already part of daily life. In this context, some argue that a national initiative for free AI services for all, dubbed 'Everyone's AI,' is unnecessary. However, when considering the competitive advantages we possess, the narrative changes. South Korea is combining private sector vitality with decisive national readiness. The government plans to distribute the 'Everyone's AI' free service to all citizens this year to prevent the AI utilization gap from leading to economic and social disparities. The initiative is concrete, focusing on welfare and administrative services closely tied to daily life and industries, including health, education, finance, and taxation.
This will create a strong policy foundation for all citizens to integrate AI as a daily tool and serve as a driving force to transform the entire nation into a vast 'AI application ecosystem.'
South Korea's path to establishing itself as a global AI power is not about imitating the U.S. or China. It is about maximizing the use of its 'high-quality digital infrastructure' and 'world-class user agility.'
First, South Korea should build its own AI models, focusing on 'specialized AI' and 'physical AI' that combine its strengths in manufacturing, healthcare, finance, and public data, rather than getting caught up in a one-to-one performance battle. It is also essential to solidify its position as a 'super player' by integrating hardware with customized sovereign AI services based on its world-leading HBM memory semiconductor supply chain. Most importantly, through the national 'Everyone's AI' policy, South Korea must cultivate a highly skilled population adept at utilizing AI flexibly.
Even years from now, the U.S. will likely continue to monopolize infrastructure. China will also maintain its strategy of overwhelming competitors with a flood of quantitative data. South Korea must find the optimal path amid this deluge of infrastructure and data. By leveraging its unique field application capabilities and digital receptiveness to develop a high-quality application ecosystem, South Korea can confidently establish itself as one of the 'global AI powers' amid the fierce currents of the U.S.-China power struggle.
* This article has been translated by AI.
Copyright ⓒ Aju Press All rights reserved.