在首尔，散户投资者如今像半导体分析师一样追踪先进记忆体议题，从 HBM4 yield rates 到下一波被称为 HBF（high-bandwidth flash）的概念。韩国半导体先驱、常被称为“father of HBM”的 Kim Joungho 在 Semicon Korea 展出一份具野心的 HBF 路线图，吸引学术界、工程学生与投资者。 他认为若 HBM 点燃了 AI 热潮，HBF 将使其持续燃烧，并指出过去六个月 HBM 超级循环已让全球前三大记忆体供应商增加约 US$1.1 兆市值。

Kim 将现有 AI 记忆体架构比作家庭工作间：HBM 像放在 AI accelerator 旁的书架，速度极快但容量有限，导致超大模型资料必须溢出到较慢、类似 SSD 的储存。HBF 则将 NAND flash（USB 随身碟中的材料）垂直堆叠，类似 HBM 的就近存取方式；但 HBF 速度较低，容量却高出 10 到 20 倍，且成本更低。此设计可让系统以明显快于现行储存的速度存取 TB 级资料，符合 AI 从问答转向 24/7 personal agent、需要掌握大量个人化资料的趋势。Kim 指出 HBF 可降低 Microsoft、Meta 等大型资料中心的运行成本，并预测最终每个人可能需要约 100 TB 记忆体；在此下 HBM 不易大规模经济化，HBF 则较有可能，且他预估到 2038 年 HBF 市场可能超越 HBM。

Kim 进一步表示，HBM（速度）与 HBF（容量）仍将形成 AI 伺服器的双记忆体标准，而非互相取代。产业路线图显示 HBF 样品约在 2027 年进入市场，且因可沿用 HBM 的技术底盘，商用化速度或许更快。SK Hynix 正在推进 AIN（AI NAND），Samsung 则加码 V-NAND（vertical NAND）以避免重演 HBM 的错失。韩国目前情绪呈现 Silicon fever：KOSPI 不断创高，中央银行（Bank of Korea）上修 GDP 预期，消费者信心也创下三个月高点；若 HBF 计划如期推进，这波机会周期可能持续很久。

In Seoul, retail investors are now tracking advanced-memory developments as if they were analysts, from HBM4 yield rates to the next wave called HBF (high-bandwidth flash). Professor Kim Joungho, a leading Korean semiconductor pioneer often described as the “father of HBM,” presented an ambitious HBF roadmap at Semicon Korea, attracting academics, students, and investors. He argues that if HBM ignited the AI boom, HBF will keep it burning, and he says the HBM supercycle has added about US$1.1 trillion to the valuation of the top three memory makers over the last six months.

Kim likens the current AI memory stack to a home workspace: HBM is the bookshelf beside the AI accelerator—very fast but capacity-limited—so oversized models push data into slower, SSD-like storage. HBF vertically stacks NAND flash (the material in USB sticks) for near-memory access in the same spirit as HBM, though slower, with 10 to 20 times larger capacity and lower cost. This design can let systems access terabytes of data much faster than conventional storage, supporting AI that is shifting from answering questions to acting as a 24/7 personal agent requiring large personalized data. Kim says HBF can cut data-center costs for firms such as Microsoft and Meta; he projects each person may eventually need around 100 TB of memory, where HBM is hard to scale economically and HBF is more plausible. He also projects that by 2038, the HBF market could exceed HBM.

Kim expects HBM (speed) and HBF (capacity) to remain a dual-memory standard for AI servers rather than replace each other. Industry roadmaps place HBF samples around 2027, and commercialization could progress faster than HBM because HBF can reuse HBM’s technical foundations. SK Hynix is advancing AIN (AI NAND) while Samsung is pushing V-NAND (vertical NAND), trying not to repeat a missed inflection point. Sentiment in Korea is euphoric: KOSPI is making repeated records, the Bank of Korea has raised GDP expectations, and consumer confidence has reached a three-month high. If HBF materializes as planned, the current cycle could extend for a long time.