Game Changer in the Era of AI Inference, HBF Memory and Its Impact on the Semiconductor Paradigm

As of November 2025, with the AI industry moving past the experimental phase into full-scale commercialization, intriguing changes are occurring. While GPU computational performance was once considered the core competitive edge of AI systems, the ability to process data quickly and efficiently is now emerging as a more critical factor. Amidst this paradigm shift, a new memory architecture overcoming the limitations of existing High Bandwidth Memory (HBM) is gaining attention, namely High Bandwidth Flash (HBF).

HBF is not merely a new memory technology. It is evaluated as a next-generation memory that breaks the boundaries between DRAM and NAND flash, simultaneously implementing large capacity, high speed, and non-volatility optimized for AI inference. Considering that large-scale AI models like GPT-4 have 1.8 trillion parameters and require up to 3.6TB of memory for inference, the current maximum capacity of HBM3E at 192GB is woefully inadequate.

Due to these capacity constraints, the current structure requires connecting dozens of GPUs in parallel, which is inefficient. HBF is being highlighted as a key technology that can solve this problem. Han Yong-hee, a researcher at Growth Research, emphasized, “HBF secures high bandwidth through a vertical stacking method like HBM, while its NAND flash-based non-volatile characteristics allow for large-capacity storage, making it essential for implementing AI services that require constant memory.”

Personally, what makes this technology interesting is that it not only enhances performance but also has the potential to fundamentally change the user experience of AI services. For instance, in AI assistants or personalized services, non-volatile memory is essential to continuously remember and learn from user conversations and preferences. However, existing HBM has limitations in implementing such services due to its volatile nature, where data is lost when the power is turned off.

Technical Challenges and Differentiation Strategies of Korean Companies

However, there are two significant technical challenges in implementing HBF. The first is that yield drops sharply when applying the TSV (Through-Silicon Via) process to the NAND stacking structure, and the second is the complexity of designing a high-performance logic die (controller) that must control hundreds to thousands of NAND channels in parallel. This means that a completely new approach is needed rather than simply combining existing technologies.

Interestingly, major Korean memory semiconductor companies are approaching this issue in different ways. SK Hynix has introduced VFO (Via-Free Overpass) technology, which eliminates TSV altogether and connects the chip vertically on the outside. This seems to be an approach to fundamentally solve the yield problem of the existing TSV method. On the other hand, Samsung Electronics is attempting differentiation through FinFET-based high-performance logic die design, focusing more on the second challenge of complex controller design.

These differences in technical approaches are likely to be key factors in determining the competitive landscape in the future HBF market. If SK Hynix’s VFO technology succeeds, it could gain an advantage in terms of manufacturing cost and yield, while Samsung Electronics’ FinFET-based logic die could have strengths in the high-performance AI application market if it demonstrates superior performance.

Notably, this technological advancement is linked to the 400-layer NAND process, which will be fully realized from 2026. The 400-layer stacking is twice the level of the currently commercialized 200-layer NAND flash, requiring more precise etching processes to implement. In this process, Solbrain (headquartered in Daejeon, Korea), which supplies the key etching solution, is emerging as a major beneficiary company.

Impact on the Global Semiconductor Ecosystem

The emergence of HBF technology signifies more than just the launch of a new memory product. Currently, the AI semiconductor market is overwhelmingly dominated by NVIDIA’s (headquartered in California, USA) GPUs, but if memory-centric architectures like HBF become mainstream, the market structure could change significantly. Especially in AI inference tasks, memory performance is becoming a more critical bottleneck than computational performance, meaning innovations in memory technology could determine the overall competitiveness of AI systems.

This represents a new opportunity for the Korean semiconductor industry. In the system semiconductor field, American companies like Qualcomm (headquartered in California, USA) and Broadcom (headquartered in California, USA) have held overwhelming dominance, and NVIDIA has led the GPU sector. However, in the memory field, Samsung Electronics and SK Hynix occupy over 70% of the global DRAM market, suggesting they have a high chance of taking the lead in new memory architectures like HBF.

Currently, the global HBM market size is estimated at approximately $15 billion in 2025, and it is expected to grow by more than 40% annually, surpassing $50 billion by 2028. If HBF is fully commercialized, this market is likely to expand further. Particularly in a situation where AI inference services are rapidly spreading, the market potential of HBF, which can overcome the capacity limitations of existing HBM, seems significant.

However, considering the technical complexity and manufacturing difficulty, it still seems that time is needed for the commercialization of HBF. Especially, the yield issue and the complexity of controller design are challenges that are difficult to solve in the short term. However, given the rapid growth of the AI market and the increasing demand for large-capacity memory, the impact on the market once these technical challenges are resolved is expected to be enormous.

Personally, what is most fascinating is that HBF technology can change the paradigm of AI services itself, not just improve memory performance. The current limitations of AI services like ChatGPT or Claude in maintaining conversation context are ultimately due to memory constraints, but with the commercialization of HBF, much more natural and continuous AI interactions will become possible.

Additionally, the potential for HBF’s use in edge AI devices is high. To directly run large AI models on devices like smartphones or autonomous vehicles, multiple memory chips currently need to be combined, but if HBF can solve this with a single chip, it could significantly contribute to miniaturization and efficiency improvement of devices.

Ultimately, HBF is likely to become the central axis of a hybrid memory architecture integrating DRAM-HBM-NAND, going beyond mere technological advancement. As the era of AI inference fully arrives and the memory paradigm shifts, HBF is poised to be recognized as a key technology driving the next generation of growth in the AI industry, following GPUs. It will be interesting to see how Korean semiconductor companies leverage this opportunity and what changes they will bring to the global AI ecosystem.

This article was written after reading a Newsis article, with personal opinions and analysis added.

Disclaimer: This blog is not a news outlet, and the content is the author’s personal opinion. Investment decisions are the responsibility of the investor, and no liability is assumed for investment losses based on the content of this article.