When AI moves from &#34;generation&#34; to &#34;autonomous action,&#34; the computing power bottleneck may shift from GPUs to memory.
According to South Korean local media such as Aju Business Daily, Joungho Kim, a professor at the Korea Advanced Institute of Science and Technology (KAIST) who is hailed by the industry as the &#34;Father of HBM,&#34; recently issued a prediction: The current GPU-centric AI architecture dominated by NVIDIA will eventually be replaced by a new architecture with memory at its core.
Behind this judgment is the fundamental transformation of AI application forms. Moving from generative AI toward Agentic AI, systems need to simultaneously process massive amounts of documents, videos, and multimodal data—a trend Kim refers to as the rise of &#34;context engineering.&#34; He pointed out that to ensure speed and accuracy, memory bandwidth and capacity must increase by up to 1,000 times.
The demand side is even more startling: According to Money Today Broadcasting quoting Kim's earlier remarks, if the input scale expands by 100 to 1,000 times, memory demand could jump exponentially, with the total volume potentially expanding by as much as 1 million times.
<h2>HBM to Reach Ceiling, HBF to Take Over</h2>
Kim clearly stated that existing HBM technology—achieving ultra-high-speed transmission via vertically stacked DRAM, currently dominating the AI accelerator memory market—will be unsustainable in the Agentic AI era.
His proposed next-generation solution is HBF (High Bandwidth Flash): replacing DRAM with stacked NAND to build a &#34;giant bookshelf-style&#34; long-term memory with capacity far exceeding current limits.
By analogy, HBM is more like sticky notes on a desk—fast but limited in capacity; HBF is like an entire wall of books, capable of storing a completely different magnitude of information.
At the architectural level, SK Hynix proposed the &#34;H3&#34; architecture in a paper published by the IEEE—according to The Korea Economic Daily in February, this architecture deploys HBM and HBF side-by-side with the GPU, rather than the current design where only HBM is adjacent to the processor. This means the GPU's role will decline from &#34;protagonist&#34; to &#34;supporting character,&#34; with computing units embedded into a memory-centric system.
The timeline is gradually becoming clear.
According to Kim's prediction, HBF engineering samples are expected to appear around 2027, and Google, NVIDIA, or AMD may adopt the technology as early as 2028.
This pace is highly similar to HBM's path from the laboratory to large-scale commercialization, signaling that the industry window of opportunity has opened.
<h2>SK Hynix and Samsung, Another Direct Face-off</h2>
Kim also pointed out that the competition in the HBF field will replicate the script of the HBM era—SK Hynix and Samsung Electronics will once again be the protagonists.
Currently, SK Hynix formed an HBF standardization alliance with SanDisk in February this year, aiming to seize dominance in the ecosystem. Samsung, on the other hand, is continuing to advance next-generation HBM products like HBM4E while simultaneously investing in NAND architecture R&amp;D consistent with the HBF concept, according to Aju Business Daily.
The two giants have different strategic paths, but their goals point to the same track. Whoever can first complete the closed loop from standard-setting to mass production and delivery will largely determine the landscape of the next round of the AI memory market.

SOXL

SSNGY

FTXL

NVDA

SOXQ

SOXX

The "Father of HBM" predicts that AI architecture will shift from GPU-centric to memory-centric—memory demand in the Agentic AI era may swell by 1 million times, existing HBM technology will hit a ceiling, and a new generation of HBF technology based on stacked NAND is expected to see engineering samples around 2027 and be adopted by Google or NVIDIA by 2028. SK Hynix has partnered with SanDisk to lead in standard-setting, while Samsung is simultaneously positioning its layout

- As AI transitions from "generation" to "autonomous action," the bottleneck in computational power may shift from GPUs to memory.  
- Joungho Kim, known as the "father of HBM," predicts a new memory-centric architecture will replace the current Nvidia-led GPU model, driven by the demands of Agentic AI which necessitates vast computation.  
- The competition in the HBF (High Bandwidth Flash) sector will mirror the HBM era, with SK Hynix and Samsung positioning themselves to dominate the evolving AI memory market.

Wallstreetcn

Direxion Semicon Bull 3X

Samsung Electronics 

First Trust Nasdaq Food & Semicon

NVIDIA

VanEck Semiconductor ETF

SPDR S&P Semicon

Invesco PHLX Semiconductor ETF

iShares Semiconductor ETF

"Father of HBM" Predicts AI Architecture Disruption: Memory to Replace GPU as Core