The Ministry of Trade, Industry and Energy will sign agreements next month with public institutions under its supervision to expand purchases of domestically made artificial intelligence chips. Of 191 AI accelerators bought over the past two years or so by public institutions overseen by the National Assembly's Trade, Industry, Energy, SMEs and Startups Committee, only one was a domestic product, and the government aims to create early demand for locally built neural processing units (NPUs) starting with the public sector.
Trade, Industry and Energy Minister Kim Jung-kwan said at a parliamentary audit session on the 22nd that the ministry had agreed to sign a domestic semiconductor agreement with public institutions during November.
Public institutions' purchases of AI accelerators have in practice gone almost entirely to Nvidia. Rep. Kim Jong-min, an independent lawmaker, obtained and analyzed purchase records from January 2024 through August of this year from 42 organizations, including the trade ministry, the Ministry of SMEs and Startups, the Ministry of Intellectual Property and public institutions under their supervision. The data showed that 21 institutions bought 191 units worth 10.29 billion won ($7.2 million). Nvidia products accounted for 190 units worth 10.285 billion won, or 99.98% of the total by value.
The only domestic purchase was a single NPU card from FuriosaAI, bought for 2.5 million won by Korea Venture Investment Corp. for a chatbot service pilot. For its separate AI infrastructure, Korea Venture Investment Corp. installed four Nvidia H200 units at a cost of 297 million won. By institution, Korea Hydro & Nuclear Power spent the most at 3.964 billion won, followed by Korea Gas Corporation at 1.0918 billion won, Kangwon Land at 955 million won, KEPCO at 871 million won and Korea Mine Rehabilitation and Mineral Resources Corp. at 781 million won.
The tilt toward Nvidia was not limited to the high-performance GPUs needed for large-scale AI training. Nvidia products were also used extensively in inference, the market domestic NPU makers are targeting as their main battleground. Nvidia purchases with a clear inference purpose — running generative AI services, chatbots and call centers, document recognition, detecting and masking personal data, cybersecurity monitoring and AI agents — came to 18 cases worth 4.8521 billion won, or 47.2% of total spending. Of 19 purchases with a clear inference purpose, only one went to a domestic NPU.
Kim pointed to the burden of accountability on public institutions as one reason adoption of domestic NPUs has been slow to pick up. "I found that public institutions feel a heavy burden in using NPUs," the minister said. "There is a concern that if they use one, they could later be called to account and asked, 'Why did you use this?'" The remarks reflect concerns that purchasing officials would be held responsible if problems arose after choosing a domestic product over Nvidia hardware with a proven performance record and accumulated operating experience.
The ministry is pursuing separate agreements with affiliated public institutions in order to lower that burden. "Rather than simply issuing administrative guidance, the ministry and the public institutions are putting together an agreement that sets out these points," the minister said. Instead of leaving individual institutions to bear responsibility for choosing domestic products, the ministry intends to set a joint policy on expanding purchases and generate early demand.
Rep. Kim said the public sector should serve as a testing ground for domestic AI chips, asking, "If Korean institutions themselves won't use them, why would foreign companies?" He urged the government to increase purchases starting with inference, where domestic NPUs can be competitive, and to make active use of approaches that deploy GPUs and NPUs together.