Trends | Divergence Emerges in the HBF Race as a Giant Takes a Different Technology Path

11/16 2026 396

Preface:

In August 2026, SK Hynix and SanDisk jointly released the first standard specification for HBF (High Bandwidth Flash) at the FMS Summit. Just six months after the alliance was formed, Google and Tenstorrent had already joined, and the ecosystem appeared to be taking shape. At the same time, however, another memory giant—Samsung—is pursuing a completely different path: it is also stacking NAND, but its product definition, target scenarios, and business logic hardly fall within the same frame of reference as SK Hynix's HBF. More tellingly, the customer HBF most needs to win over—NVIDIA—has yet to appear on the alliance roster. One technology direction, two product definitions, and one absent key buyer.

The divergence in the HBF race is not a question of who moves faster, but rather that there is still no consensus on what this technology should ultimately look like.

Author | Fang Wensan

Image source | Internet

HBF's positioning has been "complementary" from the start, not "replacement"

To understand the divergence, one must first see clearly the position HBF is meant to occupy in the memory hierarchy. HBF's logic is not complicated: use HBM's TSV stacking process to handle NAND flash, combining flash's capacity advantage with read bandwidth close to that of HBM. The specification published by SK Hynix defines 8-layer and 16-layer NAND stacking, with a maximum capacity of 512GB and bandwidth in three tiers, covering approximately 0.4TB/s to 3.0TB/s, connected to the processor via the UCIe interface. Its bandwidth is about half to 60% of HBM's, its capacity is 8 to 16 times that of HBM, and its power consumption is about 40% lower.

This set of parameters itself defines the boundaries. HBF does not seek to match HBM in bandwidth; it targets the gap where HBM cannot fit and SSDs are too slow—large-capacity KV Cache, long-context inference, and the long-term memory state of Agentic AI. These data need to be "fast enough," but not "the fastest." SK Hynix executive Ahn Hyun said directly at SEDEX 2026: large-capacity KV Cache that HBM cannot fully cover will be handled by HBF.

HBF is a "tiered complement" business, not a "next-generation HBM" business. If it is evaluated as a replacement for HBM, its bandwidth will never be enough; if it is designed as a capacity extension of HBM, it has a reason to exist. This distinction is precisely the starting point of all subsequent divergence.

Two routes: the data center "open layer" and the mobile "on-device layer"

The divergence between Samsung and SK Hynix is most intuitively reflected in product definition. SK Hynix is taking an alliance-based, open-standard, data-center-first route. The HBF specification is publicly disclosed through OCP, and alliance members include SanDisk, Google, and Tenstorrent, with target scenarios clearly pointing to AI inference servers and the custom NPU ecosystem. Google's participation is especially critical: Google's TPU ecosystem is expanding, and HBF can partially replace HBM as well as standard DDR memory—a practically attractive proposition for Google. SK Hynix positions HBF as "an extension of its HBM-centered strategy," with NAND stacking layers planned at 8 to 12 layers, targeting data-center-level workloads.

Samsung is taking another route. It is also doing TSV-stacked NAND, with a product called zNAND-O, but the target is not data centers; it is on-device AI on mobile devices such as smartphones, with stacking layers planned at 4 to 8 layers. Samsung has explicitly clarified the difference between the two: HBF was originally designed to fill the memory hierarchy gap beside AI accelerators on the server side, while zNAND-O is positioned for on-device devices to store large models. Regarding data center NAND demand, Samsung judges that eSSD will remain dominant in the short term, and HBF is not a priority.

The root of this divergence lies not in technical capability, but in product structure. Samsung has a massive smartphone business, and the storage demand of on-device AI is directly tied to its terminal shipments; SK Hynix's business focus is on high-performance computing and HBM, and the data center is its home ground that it must defend.

NVIDIA's "not choosing" is a more important signal than choosing sides

Beyond Samsung and SK Hynix, there is a third attitude: not following suit. Multiple sources indicate that NVIDIA does not plan to introduce HBF in the short term and will continue to use HBM as its core memory solution, while believing that enterprise SSDs (eSSD) can already relatively well address capacity and speed constraints. NVIDIA is working with Kioxia and SK Hynix to develop high-performance AI SSDs supporting PCIe 7.0, and is advancing a software platform called SCADA, attempting to shorten the distance for data to reach the GPU by bypassing the traditional CPU read path.

NVIDIA's choice is not a denial of HBF's technical value, but a pricing of "system modification cost." For HBF to enter the near-memory tier of the GPU, memory controllers, packaging structures, thermal design, and software stacks all need corresponding adjustments. If the path of eSSD plus software optimization can cover a considerable portion of capacity demand with lower system complexity, then for NVIDIA, HBF's priority will naturally be pushed back. TrendForce's analysis points in the same direction: considering the complexity of managing multi-tier storage, NVIDIA and AMD may lack the incentive to broadly support HBF.

This means that the pace of HBF's deployment on the data center side does not entirely depend on how fast SK Hynix and SanDisk push forward. If the largest demand-side player does not open the door, then no matter how good the standard is, it can only first seek its first scaled outlet in customized ecosystems such as Google TPU and Tenstorrent.

The essence of the divergence: three interests, three timetables

Looking at the three routes together, the logic of the divergence becomes clear. SK Hynix's timetable is "exchange openness for ecosystem." It has already gained a share advantage in HBM, and HBF is its attempt to replicate the same logic on the NAND side: build the standard first, pull in allies first, use OCP's open framework to lower customer access barriers, and turn players with custom chip capabilities such as Google into early anchors. This path requires large upfront investment and has a long return cycle, but once the standard is established, its positioning value is also the highest.

Samsung's timetable is "use the on-device side to nurture the technology." Smartphones are its ready-made shipment channel, on-device AI has real demand for storage capacity, and the bandwidth requirements of on-device scenarios are far lower than those of data centers. First get the yield and cost of TSV-stacked NAND running smoothly in a market with greater fault tolerance and a shorter validation cycle, then decide whether to extend to data centers—this is a more conservative order with lower cash flow risk.

NVIDIA's timetable is "use the existing hierarchy to solve the problem." Its stance represents the general mindset of the largest downstream buyer: whether there is still a gap between HBM and eSSD depends on whether that gap is worth restructuring the system for. HBF's capacity advantage is real, but if the cost of system complexity exceeds the capacity benefit, delay is the rational choice.

Conclusion

SK Hynix is betting that an open standard can define a brand-new memory tier, Samsung is using its most familiar on-device market to find the first commercialization path for this technology, and NVIDIA is waiting for the account between system cost and capacity benefit to be calculated more clearly. Divergence has already occurred, while consensus is still on the way.

Internet citations:

Jiwei.com: "The booming HBM, the giant keeping a backup move! Major vendors place separate bets on the storage 'Plan B'"

IT home: "SK Hynix and SanDisk will release the first HBF standard specification at FMS 2026, showcasing AI-oriented storage solutions"

Global Semiconductor Observer: "SK Hynix and SanDisk release the first HBF standard specification"

Solemnly declare: the copyright of this article belongs to the original author. The reprinted article is only for the purpose of spreading more information. If the author's information is marked incorrectly, please contact us immediately to modify or delete it. Thank you.