Samsung Electronics is moving to secure an early lead in the HBF (High Bandwidth Flash) market, an emerging next-generation AI memory category. Following its recent unveiling of “zHBM,” a next-generation 3D memory, the South Korean chipmaker has now filed a trademark application for “zHBF,” signaling its push to establish proprietary technology and branding in next-generation AI memory built on NAND flash as well as DRAM. The AI memory competition that has centered on HBM is rapidly expanding into HBF.

According to industry sources on the 27th, Samsung Electronics recently filed a trademark application for “zHBF.” The company has not yet disclosed the specific product structure or performance of zHBF, but has moved first to secure the brand.

Unlike HBM, which stacks DRAM vertically, HBF stacks NAND flash to simultaneously deliver high capacity and high data transfer speeds. It is regarded as a new memory tier that bridges the performance and capacity gap between HBM and solid-state drives (SSDs).

As the AI market rapidly shifts from training to inference, the need for HBF is growing. AI models are getting larger and the volume of data to be processed is surging, making it increasingly difficult to secure sufficient memory capacity with expensive HBM alone. HBF leverages high-capacity NAND while supplying data faster than SSDs, helping to reduce memory bottlenecks in AI accelerators.

The first HBF standard specification released this month by SK Hynix and SanDisk also defines HBF as a new memory tier between HBM and SSDs. The specification sets maximum capacity at 512 gigabytes (GB) and bandwidth at approximately 0.4 to 3.0 terabytes per second (TB/s) depending on the tier.

3D Memory Strategy Extending from zHBM to zHBF

Samsung Electronics’ first next-generation 3D memory card, unveiled ahead of zHBF, is “zHBM.” Earlier this month, Samsung revealed the industry’s first zHBM concept model at FMS (Future of Memory and Storage) 2026 in Santa Clara, California.

zHBM departs from the conventional approach of placing HBM alongside AI accelerators, instead stacking memory vertically on top of the accelerator. The concept aims to reduce latency and power consumption in data movement by shortening the physical distance between compute units and memory.

The AI semiconductor performance race is effectively shifting from individual chip performance to how compute units and memory are arranged and connected.

The key question is what “z-axis” structure Samsung will apply to zHBF following zHBM. Samsung has not yet disclosed the structure or specifications of zHBF. Experts caution that the trademark filing alone does not confirm a vertical stacking structure identical or similar to zHBM.

Samsung may reveal the blueprint and outline of its zHBF technology at Semicon Taiwan 2026, which opens early next month in Taiwan. According to the event’s official program, Choi Jang-seok, senior vice president and head of the New Business Planning Team at Samsung Electronics’ Memory Business, will deliver a keynote address titled “Leading the AI Era with 3D Integration.” Organizers have listed “zHBM, zHBF, and 3D packaging” as key technologies to be covered in the presentation.

With Samsung putting zHBM and zHBF front and center alongside 3D packaging, a key point to watch is whether zHBF is also designed to narrow the physical distance between AI accelerators and NAND-based memory to improve data transfer efficiency. If the specific structure is revealed, Samsung’s next-generation AI memory strategy is expected to come into sharper focus.

HBF Standards Race and Memory Market Restructuring

Competitors are already accelerating efforts to dominate the HBF ecosystem. SK Hynix and SanDisk agreed to collaborate on HBF technology standardization in August last year, and in February this year formed a joint working group under the Open Compute Project (OCP), the world’s largest open data center technology consortium.

The competition over HBF is unfolding amid a fundamental shift in the memory industry’s standing driven by AI infrastructure expansion. According to market research firm Gartner, global memory revenue this year is estimated at $837.3 billion (approximately 1,157.4 trillion won), up 280.4% from last year’s $220.1 billion. Over the same period, non-memory revenue including central processing units (CPUs) is projected to grow 21.9% from $589 billion to $717.9 billion (approximately 992.3 trillion won).

Memory revenue surpassing non-memory is unprecedented in semiconductor history. Memory accounted for just 27.1% of the total semiconductor market last year, but is projected to exceed half at 53.8% this year.

The rapid expansion of AI infrastructure is behind this reversal. Because AI requires repeatedly retrieving and processing vast amounts of data, raising GPU compute capability alone has limits in boosting overall system performance. As a result, HBM — which rapidly exchanges large volumes of data between GPUs and memory — has become a core component of AI accelerators, while the importance of high-capacity DRAM for AI servers and enterprise SSDs for data storage has risen in tandem.

Ben Lee, director analyst at Gartner, said, “As AI infrastructure expands, the demand composition within the semiconductor industry is being restructured,” adding that this “demonstrates structural changes in where semiconductor value is created and in demand patterns across the industry.”

HBM Technology Race Accelerating in Parallel

Samsung Electronics’ zHBF trademark filing comes amid an intensifying HBM technology competition. Samsung and SK Hynix recently unveiled their next-generation HBM technology roadmaps at Hot Chips 2026, a global semiconductor conference held at Stanford University in the United States.

Samsung announced it will apply its 4-nanometer (nm) logic process to the base die of its HBM4/HBM4E products and raise pin speeds to up to 11.7Gbps, achieving single-stack bandwidth of 3.3TB/s. The company also presented a design that moves some memory controller functions to the base die to free up silicon area on the GPU.

SK Hynix said it will adopt a base die based on a 12nm logic process in collaboration with Taiwan’s TSMC, while unveiling “iHBM” technology to address thermal issues. The approach creates dedicated heat dissipation channels at points where heat concentrates within the HBM, reducing thermal resistance by more than 30% compared to existing designs. The company also presented a hybrid bonding technique that increases chip thickness by 24% while lowering thermal resistance by 35%.

Both companies are also seeing shifts in their HBM4 supply strategies for Nvidia. According to industry sources, Samsung and SK Hynix plan to expand the proportion of 8-high products over 12-high in HBM4 shipments to Nvidia in the second half of this year. This comes as HBM4’s input/output (I/O) count doubled to 2,048 from the previous generation, significantly increasing heat generation, compounded by yield degradation issues in 12-high stacking.

The next-generation HBM4E is also increasingly likely to feature 8-high as the mainstream configuration. HBM4E is slated for Nvidia’s next-generation AI accelerator, “Rubin Ultra.”

Meanwhile, China’s CXMT is also mounting a stronger challenge. CXMT is reportedly supplying its LPDDR6 to Xiaomi’s next-generation flagship smartphone application processor, the “XRING O3.” Chinese flagship smartphones have historically relied on Korean memory, but the latest low-power DRAM segment is now expected to see SK Hynix and CXMT compete as dual suppliers. However, CXMT’s LPDDR6 data transfer speed tops out at 12.8Gbps, approximately 11% lower than the 14.4Gbps offered by Samsung Electronics and SK Hynix.

Gartner projects next year’s memory revenue will reach $1.0755 trillion (approximately 1,486.6 trillion won), exceeding non-memory by $211.1 billion (approximately 291.8 trillion won). As memory emerges as a core resource in the AI era, the standing of Samsung Electronics and SK Hynix is expected to rise accordingly. According to Counterpoint Research, the two companies’ combined share of global DRAM revenue reached 65% in the second quarter of this year, and their combined HBM revenue share stood at 79% in the first quarter.