AI BOARD
AI Board
AI tools, policy moves, and public decisions are tracked in plain language.
[Kim Kyung-jin's AI Strategy Note] <22> Korea's AI Infrastructure: Scattered It Dies, Concentrated It Lives
Author
김 경진
Date
2026-02-24 23:25
Views
26
https://n.news.naver.com/mnews/article/030/0003401583?sid=100
Electronic Times
[Kim Kyung-jin's AI Strategy Note] <22> Korea's AI Infrastructure: Scattered It Dies, Concentrated It Lives
Entered
2026.02.24. 4:01 PM
The total number of graphics processing units (GPUs) Korea has secured from NVIDIA is 260,000. Naver has 60,000, Samsung 50,000, SK 50,000, Hyundai Motor 50,000, and the government 50,000. Looking only at the numbers, it seems acceptable.
Look overseas. Meta is operating 1.3 million GPUs as of the end of 2025. xAI built a supercluster in Memphis called Colossus, where 200,000 GPUs are running, with plans to expand to 1 million. Anthropic signed a contract worth tens of trillions of won with Google Cloud and secured access to up to 1 million TPU chips. Microsoft set aside more than KRW 190 trillion for its 2026 capital expenditure budget, Google Alphabet KRW 244 trillion, and Amazon KRW 270 trillion. The combined 2026 AI infrastructure investment of the five big tech companies is about KRW 880 trillion. That is close to half of Korea's GDP.
In this situation, what can 260,000 GPUs do? The problem is that even these 260,000 are divided among five places. Naver uses them for HyperCLOVA, Samsung for Gauss, and SK and Hyundai Motor for their own businesses. The government's 50,000 are separate as well.
When xAI trained Grok-4, it put 200,000 GPUs, mixing H100, H200, and GB200, into a single cluster. It is known that more than 50,000 GPUs were used to train OpenAI's GPT-5 as well. No Korean company can reach this scale alone. Whether 50,000 or 60,000, if used separately, they are far too insufficient for training frontier models.
There is one answer. They must be pooled.
Integration needs two axes: physical concentration and operational integration. To train large-scale AI models, tens of thousands of GPUs must be in one place, connected through ultra-high-speed dedicated lines. This is why xAI concentrated 200,000 GPUs in one location in Memphis. All GPUs must be gathered in one place to build a dedicated large-scale training cluster.
GPU use should be coordinated by a central system that allocates resources. In the end, a special-purpose entity must be created to coordinate among companies. Each company has different peak times for GPU demand and different training cycles. If operations such as "Naver is not running large-scale training today, so let us allocate idle GPUs to startups" or "Let us support university research teams with Samsung's spare resources" become possible, integrated scheduling alone can greatly raise overall utilization.
The only actor that can do this is the government. Naver cannot suggest to Samsung, "Let us combine our GPUs." Private companies in competition do not voluntarily share core assets. The government must establish a special-purpose corporation or a joint-stock company, something like a "National AI Computing Corporation," and create a structure in which each company contributes GPUs in kind or entrusts usage rights. Companies should be guaranteed priority usage rights and revenue distribution rights according to their contribution ratios, while idle resources are allocated to startups and research institutions. Only then do economies of scale begin to work.
Even while unable to obtain cutting-edge GPUs because of U.S. export controls, China is building more than 250 intelligent computing centers and growing its own chip ecosystem. The state stepped in directly to use scarce resources as efficiently as possible. Korea can buy NVIDIA chips. Being able to buy them and using them properly are different matters.
If the 260,000 GPUs are divided by company and used separately, they are trapped in silos. It is like going to a battlefield with five rifles and firing them one by one separately. They must be pooled. Training resources should be physically in one place, and the rest should be unified through an operating system. Only the state can do that. And only then will the Republic of Korea avoid falling behind in the war for AI dominance.
Kim Kyung-jin, former National Assembly member
Kim Kyung-jin, former National Assembly member
Kim Kyung-jin, former National Assembly member 2016kimkj@gmail.com
Electronic Times
[Kim Kyung-jin's AI Strategy Note] <22> Korea's AI Infrastructure: Scattered It Dies, Concentrated It Lives
Entered
2026.02.24. 4:01 PM
The total number of graphics processing units (GPUs) Korea has secured from NVIDIA is 260,000. Naver has 60,000, Samsung 50,000, SK 50,000, Hyundai Motor 50,000, and the government 50,000. Looking only at the numbers, it seems acceptable.
Look overseas. Meta is operating 1.3 million GPUs as of the end of 2025. xAI built a supercluster in Memphis called Colossus, where 200,000 GPUs are running, with plans to expand to 1 million. Anthropic signed a contract worth tens of trillions of won with Google Cloud and secured access to up to 1 million TPU chips. Microsoft set aside more than KRW 190 trillion for its 2026 capital expenditure budget, Google Alphabet KRW 244 trillion, and Amazon KRW 270 trillion. The combined 2026 AI infrastructure investment of the five big tech companies is about KRW 880 trillion. That is close to half of Korea's GDP.
In this situation, what can 260,000 GPUs do? The problem is that even these 260,000 are divided among five places. Naver uses them for HyperCLOVA, Samsung for Gauss, and SK and Hyundai Motor for their own businesses. The government's 50,000 are separate as well.
When xAI trained Grok-4, it put 200,000 GPUs, mixing H100, H200, and GB200, into a single cluster. It is known that more than 50,000 GPUs were used to train OpenAI's GPT-5 as well. No Korean company can reach this scale alone. Whether 50,000 or 60,000, if used separately, they are far too insufficient for training frontier models.
There is one answer. They must be pooled.
Integration needs two axes: physical concentration and operational integration. To train large-scale AI models, tens of thousands of GPUs must be in one place, connected through ultra-high-speed dedicated lines. This is why xAI concentrated 200,000 GPUs in one location in Memphis. All GPUs must be gathered in one place to build a dedicated large-scale training cluster.
GPU use should be coordinated by a central system that allocates resources. In the end, a special-purpose entity must be created to coordinate among companies. Each company has different peak times for GPU demand and different training cycles. If operations such as "Naver is not running large-scale training today, so let us allocate idle GPUs to startups" or "Let us support university research teams with Samsung's spare resources" become possible, integrated scheduling alone can greatly raise overall utilization.
The only actor that can do this is the government. Naver cannot suggest to Samsung, "Let us combine our GPUs." Private companies in competition do not voluntarily share core assets. The government must establish a special-purpose corporation or a joint-stock company, something like a "National AI Computing Corporation," and create a structure in which each company contributes GPUs in kind or entrusts usage rights. Companies should be guaranteed priority usage rights and revenue distribution rights according to their contribution ratios, while idle resources are allocated to startups and research institutions. Only then do economies of scale begin to work.
Even while unable to obtain cutting-edge GPUs because of U.S. export controls, China is building more than 250 intelligent computing centers and growing its own chip ecosystem. The state stepped in directly to use scarce resources as efficiently as possible. Korea can buy NVIDIA chips. Being able to buy them and using them properly are different matters.
If the 260,000 GPUs are divided by company and used separately, they are trapped in silos. It is like going to a battlefield with five rifles and firing them one by one separately. They must be pooled. Training resources should be physically in one place, and the rest should be unified through an operating system. Only the state can do that. And only then will the Republic of Korea avoid falling behind in the war for AI dominance.
Kim Kyung-jin, former National Assembly member
Kim Kyung-jin, former National Assembly member
Kim Kyung-jin, former National Assembly member 2016kimkj@gmail.com