The contemporary AI arms race has collided with the hard physical limits of memory and a critical deficit in computational capacity. Conventional HBM-based solutions are becoming prohibitively expensive and capacity-constrained, struggling to keep pace with the escalating demands of Large Language Models (LLMs). The emergence of the HBF standard promises to fundamentally reshape the economics of neural network cluster architecture, delivering unprecedented data density without compromising read performance. SanDisk is already laying the groundwork for this paradigm shift, targeting a full-scale market rollout by 2028.
The semiconductor industry stands at a critical juncture: the development of the High Bandwidth Flash (HBF) standard, co-initiated by SK hynix and SanDisk, is paving the way for an entirely new class of storage devices. While the market has long relied on costly High Bandwidth Memory (HBM), this new standard offers a strategic alternative that bridges the gap between high-speed access and massive storage capacity. According to current roadmaps, the first physical HBF chip samples are expected next year, with full-scale industrial production slated for 2028.
The primary advantage of HBF lies in its ability to deliver read speeds comparable to HBM while expanding overall storage capacity by six to eight times. For AI infrastructure, this represents a fundamental reconfiguration of hardware architecture. Integrating HBF with GPUs allows for the same token generation performance using half the number of compute modules. In essence, four HBF-enabled GPUs can replace eight traditional HBM-based accelerators, significantly reducing both power consumption and the physical footprint of data centers.
From a technical perspective, HBF is an evolution of NAND flash memory, utilizing vertical stacks of 8 or 16 dies to achieve a total capacity of up to 512 GB. However, such density comes with a trade-off: while throughput remains competitive, HBF cannot match the ultra-low latency of HBM. This suggests a future market defined by synergy, where both memory types coexist and complement one another based on specific system requirements—ranging from near-instantaneous data access to the processing of gargantuan datasets.
This technological shift is poised to alleviate the chronic shortage of GPUs and HBM that has hampered AI development globally. By lowering the total cost of ownership (TCO) for infrastructure, developers will be able to scale their models more rapidly than was previously possible using exclusively expensive DRAM-based solutions.
The economic outlook for this rollout is bullish. Between fiscal years 2028 and 2030, SanDisk's revenue growth is projected to average between 16% and 19%, with profit margins remaining robust at approximately 80%. This forecast is underpinned by long-term agreements with eight major tech players who have already reserved production capacity. By 2028, these contracts are expected to account for two-thirds of the company's total output, underscoring the intense market demand for the transition to the HBF standard.
The Price of Acceleration: CXMT’s Rapid Push into Memory
