[Market Pulse] Why a Korean AI-Dedicated Supercomputer Must Be a National Design Project

World-Class Semiconductor Competitiveness
Yet to Be Linked to System Design
Government-Led Consortium Is Essential

Youngsun Park, former Minister of SMEs and Startups

Is it possible for South Korea to realize a domestically designed artificial intelligence (AI) supercomputer? While this may seem like a technical question, in reality, it is a matter of national policy choices. Japan and Taiwan have already taken steps in this direction, or are currently doing so. Only South Korea has yet to begin.

Japan has defined supercomputers as a national scientific infrastructure. The supercomputer Fugaku, which Japan takes pride in, was conceived not simply to achieve the highest performance, but to accumulate the capability for Japan to independently design, operate, and improve computing architectures. As a result, this initiative has propelled the Fugaku supercomputer to the ranks of the world's top performers.

Taiwan has chosen a different path. Rather than insisting on design sovereignty, Taiwan aims to become the country best at manufacturing and integrating global AI systems. Even within the Nvidia graphics processing unit (GPU)-centered ecosystem, Taiwan has secured an irreplaceable position in packaging, power efficiency, and system integration. Currently, Taiwan is collaborating with Nvidia to build an AI supercomputer based on 10,000 GPUs, with the goal of becoming the world's leading AI manufacturing nation. If all goes according to plan, this AI supercomputer will be operational this year for use in AI factories.

The Issue for South Korea Is Not 'Capability' but 'Connection'

South Korea possesses world-class competitiveness in memory semiconductors, including high-bandwidth memory (HBM). In AI supercomputers, memory is not a secondary component, but a critical factor that determines both performance and power efficiency. Nevertheless, South Korea has failed to connect this strength to system design. This is not due to a lack of technology, but rather a lack of policy.

South Korea's industrial policy has succeeded in strengthening competitiveness at the component level, but it has failed to address HBM, neural processing units (NPU), systems, and software as a unified design challenge. For South Korea to become one of the world's top three AI powers, it must develop a "Korean AI supercomputer." There is a significant difference in industrial impact between countries that have experience in designing their own supercomputers and those that do not.

Therefore, the government must adopt the development of purpose-built AI supercomputers for national strategic industries as a national project. Rather than focusing solely on computation, there should be AI supercomputers specialized in large language model (LLM) training, manufacturing and process AI, and defense and energy simulation, featuring architectures optimized for memory and data movement. This is the most direct way for South Korea to convert its HBM competitiveness into system sovereignty. It is also an excellent opportunity to provide global reference cases for domestic fabless NPU companies such as FuriosaAI, which are just beginning to blossom.

AI Supercomputers Must Be a National Project

An AI supercomputer is not a project that a single company can undertake. The government should take the lead in forming a permanent design consortium. The participating entities should include memory, fabless, and system companies, government-funded research institutes, universities, compiler and AI framework researchers, and more. This cannot be managed with a short-term performance evaluation system. It must be administered as a long-term national infrastructure project on a 10-year scale.

And most importantly, there must be a national project that links HBM, NPU, and AI supercomputers. This is the decisive differentiating factor unique to South Korea. South Korea is almost the only country capable of simultaneously possessing and nurturing both HBM and next-generation NPUs. Only by doing so can South Korea move beyond being a mere "parts supplier" and become a "nation that designs AI computing architectures."

In this context, it is also extremely important to design a budget structure that assumes the possibility of failure. Designing an AI supercomputer may involve failures. The problem lies in the system that does not tolerate failure. The audit criteria, which are currently defined solely by "results," must also be changed. Without such institutional changes, the birth of a "domestically developed supercomputer" will not be easy. The reason a South Korean-designed AI supercomputer has yet to be realized is not due to technological difficulty, but rather fear of failure, systems and mindsets that do not tolerate it, and the issues of technological integration and decisive action.

Will South Korea Become a Design Nation or Remain a Parts Supplier?

Japan has chosen design; Taiwan has chosen optimization. South Korea must now decide whether to become a design nation by integrating HBM, NPU, and systems, or remain a country that supplies the world's best components.

An AI supercomputer is not just a calculator. It is the blueprint that determines the industrial structure of a nation in the AI era. If the Ministry of Science and ICT cannot make this decision, the National AI Strategy Committee must set this long-term agenda. The National AI Strategy Committee should take the lead as the flagship organization guiding the entire AI industry, creating a symbolic agenda for the Lee Jaemyung administration.

Youngsun Park, former Minister of SMEs and Startups