The South Korean government pushed back against reports that it was scrapping its sovereign AI foundation model program, saying the project will continue. But officials concluded that a single goal of reaching 95% of the performance of the world's top artificial intelligence models is not enough to keep pace with fast-moving global AI competition, and decided to restructure the program into two tracks — frontier AI development and industrial deployment — once the third round of evaluation now under way is completed.
Ryu Je-myung, second vice minister of science and ICT, said at a press briefing on the 2nd at the Seoul office of the National Information Society Agency (NIA) in Jung-gu, Seoul, that the government will "develop top-tier models on one side and broaden both the vertical and horizontal axes so companies can use sovereign AI models directly on industrial sites on the other." As a result, the original plan to select two teams in the third evaluation in February and extend them into a fourth round of support will be revised.
Discussions on the overhaul began in April, when Anthropic unveiled "Mythos Preview." Ryu said the emergence of an overwhelmingly powerful tool in the cybersecurity field "was an unexpected event, so much so that the U.S. government was alarmed enough to halt its release, and the shock was large enough to require lifting the entire target we had set." He added that "the government, experts and companies agreed that the sovereign AI model project cannot stay on its current trajectory while global AI technology advances exponentially."
Under the revision, sovereign AI model development will be expanded into a frontier AI development program, and a new industrial deployment track will be created to rapidly supply the AI models that industry needs to the field.
The practice of allocating 1,000 graphics processing units (GPUs) each to participating companies will also change. The government previously said it would channel 4.7 trillion won ($3.3 billion) of next year's budget into frontier AI model development and concentrate support on 10,000 Nvidia Vera Rubin GPUs along with data.
Ryu said the government will "select core companies to develop frontier AI through an open call and evaluation, while designing the program structure around equity investment so that the results flow back to the public." He added that the program "is not meant only for large companies with deep pockets; it aims for a structure in which multiple companies, universities and research teams take part together, with the likelihood of success as the top priority." The government plans to open the call for participants around the end of the year, when next year's budget bill clears the National Assembly, and to begin frontier AI development as early as March.
The government is also considering incentives for the two companies to be selected in the third evaluation. Ryu said there "could be debate over whether it is fair to give incentives to only two companies, so we will look for a point where enough consensus can be reached."
Selected companies that do not take part in the frontier AI development program will receive separate government support under the industrial deployment track. Ryu said that "when developing and spreading AI models, the most valuable data and the final bottleneck is actual field data, so spreading sovereign AI models is also urgent," adding that the industrial deployment track "will support a range of companies, including those selected in the third evaluation."
