South Korea’s finance ministry is tallying an unexpected windfall from the semiconductor boom, and a growing faction of officials wants to spend part of it on artificial intelligence. The government is considering using roughly 5 trillion won, about $3.6 billion at recent exchange rates, in incremental tax revenue to fund the development of homegrown AI models, the Korea Economic Daily reported Thursday, citing unnamed government officials.
The Ministry of Science and ICT has discussed buying about 10,000 Nvidia Vera Rubin GPU modules and recruiting AI talent with the money, the newspaper said. To compete with global models, the ministry is weighing whether to assign the full cluster of roughly 10,000 GPUs to a single team building a frontier model, rather than splitting resources across research institutes.
The government pushed back on the specifics. In a statement, it said it has not assessed a supplementary budget and has made no decision on whether to buy GPUs on a large scale to advance frontier-model development. The ministry’s deliberations, the statement said, remain at an early stage.
The discussion reflects a familiar anxiety in Seoul. South Korea dominates memory chips, with Samsung Electronics and SK Hynix supplying the high-bandwidth memory that Nvidia’s accelerators require, yet the country has no frontier AI lab of its own. Its largest companies have trained models, but none ranks with the systems built by OpenAI, Google and China’s leading labs.
The windfall gives the government room to act. Corporate tax revenue has swelled as memory prices climbed and chipmakers posted record profits, and officials have said some of the surplus should fund the country’s next industrial bet. A supplementary budget for AI would be the clearest sign yet that the government sees models, not just chips, as strategic infrastructure.
The 10,000-GPU question is the heart of the debate. Officials who favor concentrating the cluster in one team argue that frontier models need a critical mass of compute to be credible, and that splitting hardware across institutes would produce several mediocre systems rather than one competitive one. Others warn that a single failed bet would leave the country further behind, and that public money should spread across research groups with different approaches.
Vera Rubin, named for the astronomer, is Nvidia’s next major computing platform, the successor to the Blackwell generation that data centers have been installing over the past two years. A purchase of 10,000 modules would rank among the largest public commitments to AI compute outside the United States and China, analysts said.
The talent question is as hard as the hardware one. Korean researchers have been recruited aggressively by American and Chinese labs, and the country’s universities produce fewer AI researchers than its industry wants. Officials have discussed salary supplements and new graduate programs, the Korea Economic Daily reported, though those proposals face resistance from budget officials who question whether public money can compete with private-sector pay.
The plan is politically delicate. Opposition lawmakers have asked whether taxpayers should fund compute that chaebol groups could buy themselves, and whether a government-run model lab would compete with private companies. The ruling party has been more receptive, arguing that the country’s memory-chip dominance shows what state-backed industrial policy can achieve when it is aimed at the right technology.
Timing also matters politically. The ruling party faces an election cycle in which economic competitiveness is likely to be a central issue, and a visible commitment to AI could help the government argue that it is preparing the country for the next decade of technology competition. The opposition has countered that officials should first show how earlier industrial subsidies have performed, pointing to mixed results in areas such as biotech and robotics.
The scale of the proposed purchase would also test Korea’s electricity and data-center capacity. A cluster of 10,000 GPU modules would consume power on the order of a small city, and Korea has already faced criticism over the pace of grid expansion for its chip factories. Officials would need to line up sites, power and cooling before the hardware could be put to work, which could stretch the timeline by a year or more.
South Korea is not alone in this calculus. Japan, India and Saudi Arabia have all announced national AI programs built around large GPU purchases, and each has wrestled with the same choice between concentrating compute in one flagship effort and spreading it across institutions. The outcomes have been mixed, which makes the Korean debate a test case for how smaller economies approach the frontier.
For now, the ministry’s discussions remain internal, and the government has been careful to leave the door open without committing. The semiconductor windfall is real, and so is the pressure to use it. Whether the money flows into a single 10,000-GPU bet or into a wider portfolio of research will say a lot about how Seoul intends to compete.
