Bernstein — Global Memory and NAND... Bernstein — Global Memory and NAND: HBF … · @2030FY_X · TwKit

@2030FY_X

Bernstein — Global Memory and NAND... Bernstein — Global Memory and NAND: HBF Deep Dive — Can HBF Become the Ultimate Solution to Break the Memory Bottleneck? • The memory wall and AI architecture migration: AI is evolving from training to advanced inference and agentic workloads, making memory capacity, bandwidth, and data reuse increasingly important. Expanding model parameters and context lengths will exceed HBM capacity, forcing data to spill into the broader storage hierarchy; at the same time, cache economics (the cost of reusing KV cache is only about 10% of recomputation) incentivizes retaining rather than recomputing data. This drives infrastructure from compute-centric to memory-centric, with demand expanding from GPU/HBM alone to conventional DRAM, NAND, and HDD. • The core value of HBF: High Bandwidth Flash (HBF), by combining HBM-class read bandwidth with 8-16x the capacity of HBM and significantly lower cost per GB, is expected to materially improve the memory efficiency and economics of AI inference. In a tiered architecture, HBM retains latency-sensitive active datasets, while HBF carries massive model parameters (such as inactive MoE experts) and reusable KV cache. Its value proposition is not simply cheap memory, but achieving dual optimization of capex and compute efficiency by reducing the number of GPUs/HBM required to host large models and avoiding costly recomputation. • High-performance NAND as a new memory tier: Various technologies are emerging to fill the gap between HBM and conventional NAND, including HBF, ultra-high-performance IOPS SSDs (such as Storage Next), low-latency NAND (such as Z-NAND), pooled flash KV cache platforms (such as NVIDIA CMX), and CXL-attached DRAM. Most of these solutions are based on NAND (especially mid- to low-density SLC), and their adoption will consume significant NAND wafer capacity. Whichever architecture prevails, it will tighten the NAND supply-demand balance and support industry pricing. • HBF economics: Due to the use of high-performance SLC flash and additional silicon area (TSV, high-bandwidth I/O, etc.), we estimate HBF's wafer capacity trade ratio at as high as 6.5x (base case), well above HBM's roughly 3x; the cost ratio is about 10x that of eSSD. But if priced per module close to HBM (with 8-16x HBM capacity), HBF gross margin could reach 80-90%, in line with or even higher than NAND flash, with a significantly accretive effect on NAND ASP, revenue, and profit. • Implications for NAND and GPU/HBM demand: HBF could replicate HBM's reshaping of the DRAM industry — creating incremental NAND demand while reducing GB output per wafer due to the high wafer consumption ratio, structurally improving supply and demand. Although HBF may reduce GPU intensity and HBM demand per workload, under the Jevons Paradox, lower total AI costs will stimulate broader applications and instead drive growth in aggregate demand for HBM and GPUs. • Technology implementation and limitations: HBF converts NAND into high-bandwidth memory through 3D stacking, CBA (CMOS bonded array) architecture, TSV vertical interconnects, UCIe interface, and SLC mode. But it remains fundamentally constrained by NAND's physical characteristics: microsecond-level latency, page-level reads that discount effective bandwidth, slow writes, and limited endurance, so it cannot replace HBM but serves as a complementary tiered capacity layer. • Commercialization challenges: The biggest obstacle is the lack of endorsement from key players such as NVIDIA. Google, Meta, and SK hynix have formed the HBF alliance, but NVIDIA, Kioxia, and others are pushing Storage Next and CMX solutions. In addition, HBF needs to consume scarce CoWoS packaging resources and requires redesigns of processors and software stacks to optimize data placement. NVIDIA's GPU roadmap through 2028 does not include HBF, meaning that even if adopted, it may be years away. • Sandisk's leading position: Sandisk aims to launch its first HBF sample in 2027, with the second generation targeting edge AI. Its NAND IP, vertically integrated manufacturing, and systems capabilities enhance execution credibility. HBF is not yet included in the company's financial guidance, constituting potential upside optionality. Investment Thesis and Ratings The report covers multiple storage and hardware companies, all assigned Outperform ratings, with price targets showing significant upside: • Sandisk (SNDK): price target USD 3,000 (current price USD 1,255.58), rated Outperform. • Seagate (STX): price target USD 1,350 (current price USD 644.21), rated Outperform. • Western Digital (WDC): price target USD 600 (current price USD 356.13), rated Outperform. • Samsung Electronics (005930 KS): price target KRW 440,000 (current price KRW 222,000), rated Outperform. • SK hynix (000660 KS): price target KRW 2,700,000 (current price KRW 1,341,000), rated Outperform. • Micron (MU): price target USD 1,300 (current price USD 903.70), rated Outperform. • KIOXIA (285A JP): price target JPY 49,000 (current price JPY 12,605), rated Underperform. Key downside risks: HBF commercialization progress below expectations; continued lack of support from ecosystem giants such as NVIDIA; crowding out of CoWoS packaging capacity; alternative technologies (such as CXL and CMX) taking market share; short-term supply volatility caused by NAND wafer capacity conversion; slower growth in AI inference demand. Valuation and Industry Implications The introduction of HBF elevates NAND from general-purpose storage to the AI memory hierarchy. Its high replacement ratio (6.5x) means that a small amount of demand can absorb a large amount of wafer capacity, and tightening industry supply-demand will support NAND pricing and margins. Over the long term, the explosion of AI inference and the proliferation of MoE models will amplify the value of tiered memory architectures, and vendors with first-mover technological advantages and ecosystem alliances will benefit first.

曝光 208 · 评论 0 · 点赞 0 · 书签 0