Professor Kim Jung-ho of the School of Electrical and Electronic Engineering at the Korea Advanced Institute of Science and Technology, hailed as the 'Father of HBM' in the industry, points out that the dominance of AI computing is accelerating its shift from GPUs to memory. As AI progresses from the generative era to the age of agents, memory has become a critical bottleneck restricting development. According to Kim, this shift stems from the rise of 'context engineering,' as AI needs to simultaneously process vast amounts of documents, videos, and multimodal data. To support this trend, memory bandwidth and capacity need to increase by up to 1000 times. Currently, HBM products, which achieve ultra-high bandwidth through stacked DRAM, may struggle to meet the demands of the agent era. Kim is optimistic about HBF (High Bandwidth Flash) technology. HBF replaces DRAM with stacked NAND flash to build a long-term storage system, achieving a leap in capacity by orders of magnitude. Kim expects that HBF engineering samples will emerge around 2027, with companies like Google, NVIDIA, or AMD potentially adopting the technology as early as 2028. In the future, AI computing architectures will be built around ultra-large-capacity memory, with GPUs and CPUs becoming embedded components within them.