The past year was a time when we safeguarded Korea's artificial intelligence (AI) sovereignty and laid the groundwork for an AI great transition that will reshape the foundations of industry. Starting with the early securing of advanced graphics processing units (GPUs), we focused on simultaneously shoring up the key pillars of the national AI ecosystem, from a domestically made AI Semiconductor and physical AI to an independent AI foundation model project.
Park Yun-kyu, 60, president of the National IT Industry Promotion Agency (NIPA), shared these reflections on his first year in office in an interview with ChosunBiz on Apr. 8. Park is an information and communications technology (ICT) policy expert who served as the 2nd vice minister at the Ministry of Science and ICT and took office as the sixth president of NIPA at the end of March last year. During his tenure at the ministry, he oversaw a broad range of communications, software, and AI policies, including as head of network policy, director-general for AI-based policy, and director-general for information and communications industry policy.
NIPA is the implementing agency that supports the development of the national information and communications technology (ICT) and AI industries. It serves as a bridge between policy and industry sites, from building AI infrastructure to corporate validation, commercialization, and overseas expansion support. Recently, through the National AI Model Project aimed at developing Korea's independent AI foundation model project, it is bolstering the competitiveness of the domestic AI ecosystem and accelerating its spread to industrial sites.
Park said, This year we will shift the center of gravity of AI policy from building infrastructure to delivering real results in industrial sites and people's daily lives, such as factories, hospitals, logistics, and administration, adding, What matters now is not how many corporations we supported, but whether the supported corporations achieved results in the actual market. Park noted, In the end, the remaining keywords are execution and collaboration. The following is a Q&A with Park.
◇ We must deliver results that industry and the public can feel… speeding up to preempt the agent AI market
―What are the achievements so far and the tasks ahead?
The past year was the time when we did the groundwork for the AI expressway. The biggest achievement was securing advanced GPUs early and expanding the supply base to startups, small and midsize firms, universities, and research institutes. This opened the way for AI development capabilities to spread across the ecosystem rather than staying with a few large corporations. Elevating the domestically made AI Semiconductor as a central pillar of the national strategic industry, and initiating a framework for physical AI and independent and specialized AI foundation models are also important advances. The tasks are clear. We must achieve the goal of securing 50,000 GPUs by 2028 without setbacks, and we need many more battle-tested success models that go beyond technology development to transform industrial productivity and quality of life. If the past year was groundwork, from now on we must deliver both industrial outcomes and results that the public can feel along that road.
―What is the top priority this year?
Until now, AI policy and projects have centered on building infrastructure, developing models, and discovering pilot projects, but now AI must enter sites such as factories, hospitals, logistics, and administration to raise productivity, cut expense, and create new revenue. AI must no longer remain in the lab. It must go to the market, go to industry, and enter people's lives. To this end, we will spread the secured independent AI models and industry-specific models across manufacturing, healthcare, finance, and the public sector, and we will also speed up efforts to preempt the agent AI market.
◇ Investing budget in AX service development… securing infrastructure and scaling services at the same time
―What are the plans to secure 50,000 GPUs and the allocation criteria?
The volume the government secures directly totals about 28,000 units, combining 13,000 units procured through last year's extra budget and 15,000 units reflected in this year's budget. In addition, the Korea Institute of Science and Technology Information (KISTI) plans to secure about 9,000 more units this year, and the special purpose company (SPC) promoting the National AI Computing Center also plans to secure 15,000 units by 2028. In total, that comes to about 52,000 units. The allocation criteria are relatively clear. Of the 13,000 units secured last year, 30% were allocated to national AI projects such as independent and specialized AI foundation models, 30% to national core projects, and 40% to tasks among industry-academia-research that the state deems important. After the National AI Model Project wraps up in the first half of next year, those volumes will be allocated to the other two areas.
―How does the government secure GPUs?
The government provides budget support to private corporations to secure GPUs. For example, if we select corporations such as Naver Cloud, NHN Cloud, and Kakao, they bring in GPUs in the quantities they propose. We can split the allocation by corporation, such as 3,000 or 7,000 units each. The GPUs secured in this way are operated by the respective corporations in a cloud format. Ownership rests with the government. Corporations will bear only about 5% to 10% of the GPU market price, and we plan to provide them free of charge to universities and research institutes in industry-academia-research.
―NIPA's budget has more than quadrupled in a year.
When I took office a year ago, the budget was in the 700 billion to 800 billion won range, and this year it has grown to 3.1 trillion won. It has more than quadrupled. This does not simply mean the budget increased; it should be seen as a sign that GPU procurement will be ramped up at the national level. More than 2 trillion won of the total budget will be invested in purchasing GPUs. But it does not end with securing GPUs. About 1 trillion won of the remaining budget can be invested not only in AI model development but also in developing AX (AI transformation) services in the public institutional sector that the public can feel. In the end, it means we will push forward with securing infrastructure and scaling services at the same time.
◇ Korea can play offense in agent and physical AI
―Where does Korea's global AI competitiveness stand now?
Korea is at an inflection point to leap into the world's top tier. To differentiate, in general-purpose model development we are in a trailing, defensive role, but this should be seen as a process of securing a safety pin in terms of technical capability and security. The National AI Model Project targets performance at 95% or higher of top global models, and participating consortia say they will do even better. Korea can play offense in agentic AI and physical AI. Compared with absolute powerhouses like the United States and China, there is a gap in resources and capital. But Korea is a rare country that simultaneously possesses high digital receptivity, a strong manufacturing base, and world-class semiconductor capabilities. The key is that we are not merely a country that uses AI well; we are a country that can industrialize by combining AI models, semiconductors, manufacturing, and services. The world is watching precisely that point.
―What is most needed for Korea to rise as "AI G3"?
Rather than facing off head-on in the same way as the United States and China, we need a Korea-style AI strategy suited to our industrial structure and technical capabilities. We must combine AI with industries where we have strengths—such as manufacturing, shipbuilding, automobiles, healthcare, and logistics—to deliver results on the ground. Going forward, competition in general-purpose AI models alone has limits, so we must secure a competitive edge in areas such as industry-specific models based on the independent AI foundation model project, physical AI, and On-device AI. Ultimately, what we need is not a Goliath strategy that leads with capital and scale, but a David strategy based on our strengths.
―What is the relationship between the National AI Model Project and "AI for all"?
The independent AI foundation model project is the brain of Korea's AI, and "AI for all" is the services and ecosystem through which the public and industry actually use that brain. The idea is to make it easily accessible to anyone. Four elite teams are currently participating, and after evaluations every six months, we plan to select two final national elite teams by the end of this year. It is true that many people already use ChatGPT or Gemini. However, when there is an assessment that other options (Korean AI models) are usable if expense is offset, we can play the role of providing them for easy use. This is about breaking away from a structure where user experience accumulates only with foreign models and ensuring it circulates back into the competitiveness of our models and industrial ecosystem. Of course, this presumes that the national AI model reaches a top global level. "AI for all" is not a simple distribution policy but an ecosystem strategy to build independent AI competitiveness.
―Some say the National AI Model Project is a waste of budget.
Just as Korea would not stop growing rice because Vietnam produces a lot of it, losing competitiveness in the technology ecosystem may not be immediately visible but will ultimately become a heavy burden. In an era when wars are fought with AI, if we only use others' products, our potential applications will also diminish. If there is no alternative when supply fails, we face the worst situation. AI should not be seen as a piecemeal technology but as a core strategic asset. Korea did not master defense or nuclear power from the start. Securing initial technological competitiveness is crucial. In the AI ecosystem, our corporations must be able to collaborate with confidence based on technological competitiveness.