Professor Jung-Ho Kim from the Department of Electrical Engineering at the Korea Advanced Institute of Science and Technology (KAIST), hailed as the 'Father of HBM,' points out that the core competitiveness of AI is shifting from GPUs to memory. Kim believes that the essence of AI lies in memory, as GPU utilization during AI inference processes remains far below theoretical levels. In the past, GPUs were central to AI training, but as we enter the era of inference, memory capacity will become the key factor determining AI performance. Currently, HBM is the memory technology best suited for AI, but as AI evolves towards multimodality and Agentic AI, the need to store vast amounts of cold data will make HBF technology—which stacks NAND flash memory—likely to become the future mainstream, with its market demand expected to surpass that of HBM in 10 years. Looking further ahead, HBS technology may emerge, utilizing SRAM with read and write speeds 1,000 times faster than DRAM and envisioning the deployment of SRAM on 12-inch wafers to increase capacity to approximately 1,600GB. Kim also envisions the future form of AI computers, which will become massive 'three-dimensional structures' where HBM serves as the shopping mall, HBF layers act as residential areas, and HBS functions as high-speed cache. Various memory forms will combine to supply data to GPUs, forming a 3D composite architecture with approximately 100 layers. Kim has long been dedicated to research on high-bandwidth memory (HBM) and 2.5D/3D integrated packaging, collaborating with SK Hynix to develop HBM1 since 2010 and participating in the planning of the long-term development roadmap for HBM4-HBM8 last year.