South Korea is preparing to spend billions of dollars to develop a frontier artificial intelligence model, betting that access to the world’s most powerful AI systems may eventually be determined by governments rather than customers.
The plan reflects a growing concern in Seoul that the country’s dependence on American and Chinese technology could become an economic and national security vulnerability. South Korean officials argue that foreign AI services could be restricted, withdrawn or made prohibitively expensive as competition between Washington and Beijing intensifies.
Bae Kyung-hoon, South Korea’s deputy prime minister and minister of science and information and communications technology, said the country needs an AI model capable of competing with systems developed by leading American companies such as Anthropic.
The objective is not to replace American AI services that are already widely used in South Korea. It is to give the country a technological insurance policy if access to those systems becomes subject to export controls, security reviews or political disputes.
The shift shows how the AI race is expanding beyond the United States and China. South Korea, one of the world’s largest semiconductor producers and a major American security ally, increasingly views advanced AI as infrastructure that it cannot afford to import indefinitely.
Bae said the United States and China were beginning to control high performance AI as though it were a strategic weapon. A country without its own models, he warned, could find itself unable to respond if an overseas provider suddenly suspended service or raised prices dramatically.
South Korea plans to continue funding foundation models designed for domestic industries while pursuing a separate system that can compete at the technological frontier. The government has cited Anthropic’s latest model as a benchmark for the proposed project.
Reaching that level would require far more computing power than South Korea’s current national AI program provides. Bae estimated that the project would need about 10,000 of Nvidia’s next generation Vera Rubin graphics processing units. The chips alone would cost roughly $2.3 billion, excluding data centers, networking equipment, electricity and other infrastructure.
Companies participating in South Korea’s existing foundation model program currently receive access to about 500 to 735 Nvidia B200 GPUs. That gap illustrates the scale of the government’s ambition and the difficulty of turning it into a commercially competitive system.
The government has acknowledged that it cannot simply hand billions of dollars in computing resources to one company. It is considering equity investments, special purpose companies and other investment based research programs that would allow public money to support the project without relying entirely on conventional grants.
Officials are discussing the funding plan with other government agencies as they prepare next year’s budget.
South Korea’s effort could also create a new challenge for American AI companies. The country has long been a large customer of American software and semiconductor technology. A successful domestic model would give Korean companies and government agencies an alternative to relying exclusively on platforms developed by OpenAI, Anthropic and other foreign providers.
The strategy could also give South Korea a role in countries seeking AI systems outside the two dominant technology blocs. Bae said governments and institutions in Europe, the Middle East and Southeast Asia had shown interest in cooperation that was not centered entirely on the United States or China.
Seoul is considering offering those partners a full technology package that would combine Korean AI models, services and neural processing units. The proposal would build on South Korea’s strength in semiconductors and its experience exporting communications networks and digital infrastructure.
The plan faces a significant obstacle. Korean AI models will need to perform across many languages if they are to compete outside the domestic market. Bae said publicly available web data was insufficient and that training data from partner countries would need to represent a meaningful share of future datasets.
South Korea is also trying to avoid choosing a single side in the widening AI divide. American AI services are deeply embedded in the country, while Korean companies also use open models developed in China. Bae said the government would therefore take a cautious approach toward joining AI organizations led by Beijing.
That balancing act mirrors South Korea’s position in the semiconductor industry. The country relies on the United States for security and critical chipmaking technology while maintaining extensive commercial ties with China. AI is now creating a similar strategic problem, but one that affects software, data and computing infrastructure at the same time.
The government’s broader program includes a cybersecurity AI model that it plans to introduce first in the public sector. It also wants to develop physical AI systems that can operate robots and other machines in real world environments.
For physical AI, South Korea plans to focus heavily on synthetic data created through computer simulations. Bae said such data could account for about 80% of the information used to train future systems. The government intends to support world models, physical AI foundation models and common data standards, describing the next three years as a critical period.
