"Samsung and SK Hynix Still Have Opportunities"... The Era of HBF to Power 'Ultra-Large AI' Is Approaching [Tech Talk]

Stacking NAND Flash Instead of DRAM: HBF
Slower Speed, But Much Larger Capacity
A Supplementary Memory for Large-Scale AI

HBM (High Bandwidth Memory) has emerged as one of the core semiconductors in the era of artificial intelligence (AI). This is because it serves as an external storage device that holds massive AI data, working in tandem with graphics processing units (GPUs) that accelerate AI training. However, as AI is expected to become even larger in the future, there will come a time when the capacity of HBM alone will not be sufficient. Memory manufacturers are already moving quickly to prepare for the era of HBF (High Bandwidth Flash).

Following HBM, HBF... Now Stacking NAND Flash

Samsung Electronics' 9th Generation NAND Flash. Samsung Electronics

HBM is a product that vertically stacks DRAM (RAM), one of the memory devices, and connects them using through-silicon vias (TSV) to increase data transfer speeds, or bandwidth. HBF operates on a similar principle. The key difference is that, instead of stacking DRAM, it stacks NAND flash.

Both NAND flash and DRAM are memory devices, but their detailed characteristics differ. NAND flash is a 'non-volatile memory,' which means it retains data for extended periods even when the power is off. In contrast, DRAM is 'volatile memory,' losing all data once the power is cut. As a result, NAND flash offers much larger storage capacity but slower data transfer speeds, while DRAM is the opposite, providing faster data transfer but with limited storage capacity.

Until now, NAND flash has been used in internal solid-state drives (SSD) in computers or as auxiliary storage devices like USB drives. DRAM, on the other hand, has served as the main memory for central processing units (CPUs) in computers and smartphones, enabling fast data read and write operations. DRAM has since evolved into HBM, becoming a key component in AI accelerators, but NAND flash had remained largely unchanged.

AI Is Growing Rapidly, and HBM Alone Is Not Enough

The recent spotlight on HBF stems from projections that AI models will grow exponentially in size. Leading large language models (LLMs) such as ChatGPT, Gemini, and Claude are now capable not only of reading text but also generating and interpreting images and videos. Since images and videos are far larger data types than text, the models themselves must also become larger.

This means we are entering an era where both memory speed and storage capacity are crucial. GPU manufacturers like NVIDIA will need to deploy strategies that combine both HBM and HBF. If HBF serves as a basket holding the entire model, HBM acts as an intermediate stage, temporarily holding only the necessary data for rapid processing.

Some memory companies have already begun developing first-generation HBF. On August 6 (local time), SanDisk, a U.S.-based SSD and NAND flash manufacturer, signed a memorandum of understanding (MOU) with SK Hynix to standardize HBF for AI applications. The two companies plan to release prototypes in the second half of 2026 and unveil AI inference devices equipped with actual HBF samples in 2027.

NVIDIA Blackwell GPU connecting 8 HBM modules to a single chip. One HBF can provide the same capacity as 8 to 16 HBMs. Screenshot from NVIDIA website

The prototype co-developed by the two companies will offer 8 to 16 times the capacity of the current state-of-the-art commercial HBM, the HBM3e. The HBM3e provides 48GB per module (based on SK Hynix's 16-layer configuration), so the maximum data capacity reaches 768GB. Considering that NVIDIA's Blackwell GPU connects 8 HBM3e modules to a single chip, HBF could be a game changer in terms of capacity alone.

A Tighter Flash Market Than DRAM... "Opportunities for Samsung and SK Hynix"

Both Samsung Electronics and SK Hynix, Korea's leading semiconductor companies, produce NAND flash. However, it remains to be seen who will take the lead in the HBF market going forward.

In the DRAM market, Korean companies hold a dominant position with a 71.4% share as of the second quarter of 2025 (according to TrendForce). In contrast, the NAND flash market share is more evenly distributed. As of the second quarter, Samsung and SK Hynix together hold a 54% share, which is a majority but not as overwhelming as in DRAM. Competitors such as Kioxia, Micron, and SanDisk each hold more than 10% market share, and, most notably, Chinese companies like ChangXin Memory Technologies (CXMT) and Yangtze Memory Technologies (YMTC) are catching up rapidly, growing at an annual rate of 60-70%.

NAND flash factory of Yangtze Memory Technologies (YMTC) under construction in 2023. Screenshot from X

Semiconductor experts predict that within the next 10 years, the company that seizes the HBF market will dominate the AI memory business. Kim Joungho, a professor at KAIST known as the "father of HBM" for his contributions to HBM process technology, stated in his keynote speech at the International Advanced Semiconductor Substrate & Packaging Industry Exhibition held in Incheon on September 3, "Currently, the performance of Samsung and SK Hynix is determined by HBM, but in 10 years, HBF will take its place," adding, "This is because future AI will require more than 1,000 times the memory capacity compared to today."

He continued, "NAND flash is slower than DRAM, but its capacity is 10 times greater," and emphasized, "NAND flash will serve as a complementary component for AI memory. There are opportunities for both Samsung and SK Hynix, as they produce both DRAM and NAND."