"Father of HBM" predicts AI architecture revolution: Memory will replace GPU as the core

When AI moves from "generation" to "autonomous action," the bottleneck in computing power may shift from GPUs to memory. According to local media reports from South Korea, including Asia Economy, Professor Joungho Kim from KAIST, hailed as the "Father of HBM" by the industry, recently forecasted: The current GPU-centric AI architecture dominated by NVIDIA will eventually be replaced by a new memory-centric architecture. Behind this judgment lies a fundamental transformation in AI application form. As generative AI advances towards agentic AI, systems will need to simultaneously process massive volumes of documents, videos, and multimodal data—Kim refers to this trend as the rise of "context engineering." He pointed out that to ensure speed and accuracy, memory bandwidth and capacity must increase by up to 1000 times. Even more astonishing are the demand-side numbers: According to Money Today Broadcasting's earlier citation of Kim, if input scale expands by 100 to 1000 times, memory demand could surge exponentially, with total capacity swelling by up to 1 million times. HBM Will Hit Its Limit, HBF Takes Over Kim made it clear that existing HBM technology—which achieves ultra-high-speed transmission by vertically stacking DRAM and currently dominates the AI accelerator memory market—will find it difficult to sustain in the era of agentic AI. His next-generation solution is HBF (High Bandwidth Flash): Not DRAM but stacked NAND, to build a "giant bookshelf" for long-term memory, with capacity far exceeding current limits. By comparison, HBM is more like sticky notes on a desktop—fast but limited in capacity—while HBF is like an entire wall of books, holding vastly more information. On the architectural side, SK Hynix has already proposed an "H3" architecture in an IEEE paper—according to an article in Korea Economic Daily from February, this architecture deploys HBM and HBF side-by-side next to the GPU, rather than only having HBM adjacent to the processor as in current designs. This means the GPU’s role will shift from "lead actor" to "supporting actor," with computing units embedded inside a system centered around memory. The timeline is gradually becoming clear. According to Kim’s predictions, HBF engineering samples are expected to emerge around 2027, and Google, NVIDIA or AMD may be the first to adopt this technology as early as 2028. This pace closely mirrors the route HBM took from lab to mass commercial use, signaling that the window of opportunity for the industry has opened. SK Hynix and Samsung: Head-to-Head Again Kim also noted that competition in the HBF space will replicate the script from the HBM era—SK Hynix and Samsung Electronics will once again become the protagonists. Currently, SK Hynix formed an HBF standardization alliance with SanDisk in February this year, aiming to seize dominance over the ecosystem. At the same time, according to Aju News, Samsung continues pursuing next-generation HBM products like HBM4E, and is also investing in developing NAND architectures consistent with the HBF concept. The two giants have different development routes, but aim at the same track. Whoever first accomplishes the closed loop from standards setting to mass production will largely determine the next round of AI memory market dynamics. Risk Warning and Disclaimer The market involves risks; investors should exercise caution. This article does not constitute personal investment advice and does not consider any individual user's specific investment goals, financial situation, or needs. Users should consider whether any opinions, views, or conclusions in the article are suitable for their particular circumstances. Any investment made based on this article is at your own risk.