WhiteLine Daily said HBF is not a NAND replacement for HBM. The idea is to add a larger-capacity memory tier between HBM and SSD, placing storage closer to compute. While a common standard is now in place, products are still ahead of the sampling stage, so the report said the near-term focus should remain on SanDisk earnings, NAND pricing, and data center SSDs.
HBF gets its first public standard
SK hynix and SanDisk released the first HBF technical specification through the Open Compute Project, or OCP. Under that specification, HBF comes in 8-layer and 16-layer NAND stacks, with capacity of up to 512GB per package. Bandwidth is divided into three tiers at roughly 0.4 TB/s to 3.0 TB/s, and the memory connects to CPUs, GPUs, and other accelerators through UCIe. Google and Tenstorrent have also joined the HBF consortium.
The significance, according to the report, is not simply the arrival of another memory product. HBF now has public standards for capacity, bandwidth, and interface design for the first time. That gives different vendors a common framework for product development and moves HBF from a single-company concept closer to an industry standard.
WhiteLine Daily said HBF is not positioned as a direct substitute for HBM. Data that is read and written frequently during model execution, and that has the strictest latency requirements, would still stay in HBM or DRAM. Model weights with larger capacity needs and lower read frequency could be placed in HBF closer to the xPU. Compared with pulling that data from remote SSD storage, that arrangement could be faster. Compared with fitting everything into HBM, it could also be cheaper.
MoE may be an early use case
The report pointed to mixture-of-experts, or MoE, models as a possible first deployment path. These models have a large number of parameters, but only a subset of experts is activated for each inference task. Many expert weights remain on standby for long periods. They consume capacity, yet they do not need the same lowest-latency treatment as KV cache in every step of computation.
Right now, those weights are often spread across multiple accelerator cards, which creates heavy inter-card communication when they are called. If HBF can sit next to the xPU through UCIe, part of that traffic could shift from remote communication to local reads. The report said that could reduce pressure on HBM capacity and may lower deployment costs for very large MoE models.
SanDisk previously ran internal simulations using the Llama 3.1 405B model. Under the assumption that HBM capacity was not constrained, the overall performance gap between HBF and HBM was about 2.2%. WhiteLine Daily said that suggests HBF has potential at least in the model-weight reading scenario. It also noted that the result came from vendor simulation, not customer deployment data.
New NAND narrative, but revenue is still some distance away
The report said HBF gives NAND a new AI application angle, with SanDisk the most directly exposed name. Even so, the company’s previously published timeline shows first HBF samples are due in the second half of 2026, while the first inference device samples carrying HBF are expected in early 2027. On that schedule, WhiteLine Daily said it is still too early to talk about revenue contribution.
For that reason, the report said the more important near-term item is SanDisk’s earnings report due in the early hours of Aug. 6 Beijing time. The main points to watch are NAND pricing, data center operations, margins, and long-term customer agreements. In the previous quarter, SanDisk’s data center revenue reached $1.467 billion, up 233% quarter over quarter. The next report needs to show whether demand and pricing can continue to support that growth.
SK hynix also showed a 375-layer 4D NAND under development at FMS 2026. Performance per unit of power was said to be about 2.5x higher than the previous generation, and the company plans to begin mass production of related high-performance, high-capacity enterprise SSDs in early 2027. Compared with HBF, which is still before the sample stage, that SSD product line is closer to revenue realization.
What the report says to watch next
WhiteLine Daily’s short-term conclusion is that HBF adds a new valuation clue for NAND, but it is not yet an earnings benefit that has already materialized. Near term, the report said investors should watch SanDisk earnings, NAND prices, and data center SSD developments. Over a longer horizon, the key questions are whether HBF can be sampled on schedule, enter customer devices, and turn technical standards into actual orders.