[Special Column] Beyond the AI Factory: The Era of the AI Memory Factory and the Significance of Jensen Huang's Visit to Korea

AI Factories Require GPUs and HBM
Korea Emerges as a Key Alternative Market to China
Securing HBM from SK Hynix and Samsung Electronics Is Top Priority
Construction of AI Memory Factories Is Essential
Korea, with a Complete Ecosystem, Is an At

Until now, factories have symbolized industrial and economic growth. The representative model that emerged with the Industrial Revolution is the "Ford automobile factory." Henry Ford, born in 1863, was an American engineer and entrepreneur, and the founder of Ford Motor Company. As a business leader, he pioneered "standardization" by repeating the same production process, implemented "division of labor" by having each worker handle only their assigned tasks, and successfully established "specialization" by ensuring everyone performed their roles perfectly. Modern factories operate according to these three principles.

In 1912, while visiting a slaughterhouse, Ford witnessed workers using a monorail to quickly move carcasses hanging from hooks to the next worker. He immediately applied this to his automobile factory. This was the starting point of Ford's conveyor system-a pivotal event that changed the history of modern manufacturing.

Now, in the age of artificial intelligence (AI), the definition of a factory is changing. The emergence of the AI factory is a testament to this shift. An AI factory connects at least tens of thousands of graphics processing units (GPUs) and hundreds of thousands of high-bandwidth memory (HBM) modules via optical communication. To minimize the time required for digital electrical signals to be transmitted between GPUs for AI generation, and to enable them to operate in real time as a single unit, an AI factory must be housed in a building the size of a gymnasium. It requires gigawatt-level power supply facilities and water for cooling-either river water or seawater will suffice.

This is where artificial intelligence is mass-produced and distributed. It is the most fundamental infrastructure for AI. At the same time, millions of people can issue prompts and use it simultaneously. In this way, the AI factory becomes the hub for the development, production, and dissemination of artificial intelligence. Without an AI factory, there can be no AI services, no technology development, and no workforce training. For agentic AI, which goes beyond generative AI, such facilities are even more essential. It is a critical infrastructure for enhancing industrial competitiveness by integrating AI not only into traditional service sectors but also into education, healthcare, advertising, finance, law, and manufacturing.

To build an AI factory, GPUs and HBM are the first necessities. Recently, during the 2025 Asia-Pacific Economic Cooperation (APEC) summit, Jensen Huang, CEO of Nvidia, who visited Korea, promised to supply 260,000 GPUs to Korea. Nvidia announced plans to supply 50,000 of its highest-performing Blackwell GPUs to the Korean government, 50,000 each to Samsung, SK, and Hyundai Motor, and 60,000 to Naver Cloud-a total of 260,000 units. This is the minimum foundational quantity required to launch an AI factory. GPUs are needed to embed AI functionality into the artificial intelligence products and services of Samsung Electronics, SK, Hyundai Motor, and Naver. They are used not only for training generative AI models for these products but also for inference. The transformation of the AI industry cannot happen without AI factories.

Now, Korea has the opportunity to secure its own AI models and make a leap to become a true AI G3 powerhouse. In the future, this scale must be expanded to at least one million units. With this goal in mind, Jensen Huang, Samsung Electronics Chairman Lee Jae-yong, and Hyundai Motor Group Chairman Chung Euisun held a surprise "chimaek" (chicken and beer) meeting at a Kanbu Chicken restaurant near Samseong Station in Seoul.

Why did Jensen Huang decide to sell 260,000 GPUs to us and attend the chimaek gathering? First, it is to secure a foothold in the Korean market ahead of GPU competitors such as AMD, Google, Broadcom, and Intel. By supplying GPUs and software stacks first, Nvidia can lock in customers. Once customers use their products, it becomes difficult to switch to other GPUs. This is known as the lock-in effect.

Secondly, there is the substitution effect for the Chinese market. Due to U.S. export control policies, it is difficult to sell Blackwell GPUs to China. As a result, alternative markets include the Middle East, Europe, and Asian countries. Korea is competitive in AI technology, industry, talent, and government support. It is a market with growth potential. To expand and secure these markets, Jensen Huang is traveling around the world.

Third, there is a need to cooperate with Samsung Electronics and SK Hynix to secure supplies of memory semiconductors, including HBM. GPUs cannot function without HBM. If the future market for GPUs reaches 10 million units, about 80 million units of HBM will be needed. Even if Samsung Electronics and SK Hynix supply all their HBM, it would still not be enough. Collaboration is inevitable, with each party depending on the other.

In the future, the era of the "AI memory factory" will arrive, surpassing the AI factory. The performance of generative AI training and inference will be determined more by memory, such as HBM, than by GPUs. The number of characters or images generated simultaneously, and the number of concurrent users, will be determined by the bandwidth and capacity of memory semiconductors.

Especially in the future, AI will use the Retrieval-Augmented Generation (RAG) technique to read and respond to even more data in real time. High-bandwidth memory that stores large amounts of data next to the GPU will be essential. The results generated must be stored nearby and reused. For this, large-capacity memory semiconductors are needed. "High-bandwidth flash memory (HBF)" is being developed as such a semiconductor memory. Flash memory semiconductors will be stacked alongside GPUs and HBM. Therefore, in the future, demand for complex memory combining HBM and HBF is expected to increase as much as for GPUs. These memory semiconductor integration facilities will be built on the scale of factories, called "AI memory factories," and will be constructed on a large scale right next to AI factories.

In this way, investments in AI infrastructure will focus on memory. This is why Jensen Huang is expected to visit Korea even more frequently in the future. Negotiation and cooperation require bargaining chips. Korea's semiconductor industry possesses the entire industrial ecosystem, including design, foundry, packaging, and memory. By adding software, algorithms, and AI servers, Korea can become a true AI powerhouse. It has been a delightful week for both the eyes and ears.

Kim Joungho, Professor of Electrical and Electronic Engineering, KAIST