内存与存储：新内存范式是否回归？最佳观点

Presentation from Bernstein's Strategic Decisions Conference best idea, where we brushed off our New Memory Paradigm thesis from a decade ago and +1 212 845 7822mark.newman@bernsteinsg.com addressed where we got things right and wrong and further upside from here. Enclosed arethe presentation slides with key points from the speech.AI is driving a large and unprecedented demand surge across all forms of memory +1 917 344 8339april.li@bernsteinsg.com and storage.As workloads evolve from training and basic chatbot inference to advancedinference and agentic AI, memory and storage requirements change. HBM & DRAM wereearly beneficiaries, but NAND & HDD also benefit as AI workloads become more advanced. +1 917 344 8481phoebe.sun@bernsteinsg.com price sensitive, which is supporting much stronger pricing than in prior cycles.Is the New Memory Paradigm is back?We first wrote about the New Memory Paradigm in our 2013 Blackbook and 2014 Blackbook as a structural improvement thesis forthe memory industry. Much of that thesis has played out as slower tech-driven supplygrowth and greater industry consolidation are creating stronger upcycles and less severedowncycles, which in turn creates higher highs and higher lows. migration has continued to decline and has remained well below demand growth. That hasmade it more difficult and more time-consuming for supply to catch up in upcycles, whilealso reducing the severity of downcycles. expansion.Planar NAND had followed a pattern similar to DRAM, with GB per fab growthfalling sharply as planar NAND approached its technical limits. The transition to 3D NANDcreated a one-time step-up in GB per wafer and GB per fab. However, 3D NAND is nowalso beginning to face diminishing returns.NAND is also earlier in the cycle than DRAM, that the industry is both consolidated and growing. Since consolidation in 2012, theHDD industry had been shrinking because of aggressive NAND cannibalization. Thoseheadwinds are now largely exhausted. What remains is the business-critical, or nearline,segment, which is fundamentally a capacity market where NAND cannot compete. prepayments, and take-or-pay terms helping reduce cyclicality.Overall, we continue to like the sector and see further upside.SNDK is our preferred short-term idea, given the significant earnings upside that remains as NAND is earlier in thecycle than DRAM. Seagate is our preferred long-term idea, supported by steadier earningsgrowth from both volume and pricing, as well as HAMR-driven cost declines. BERNSTEIN TICKER TABLE INVESTMENT IMPLICATIONS Outperform. DETAILS We are seeing an unprecedented demand surge in memory and hard disk drives, with important implications formemory and HDD stocks. Paradigm thesis from 2013 and assess whether it remains in place across DRAM, NAND, and hard disk drives. We will concludewith valuation and our top picks, STX as the long-term pick and SNDK as the short-term pick, both initiated Outperform lastSeptember. At first glance, recommending stocks that are up roughly 4,000% and 700% over the past 12 months may seem aggressive.However, we continue to see meaningful upside in both names. EXHIBIT 1:Agenda More reports @wechat: Hillwood2024 resides on GPUs and CPUs, down through HBM, DRAM, NAND, HDD, and tape at the bottom. As you move up the stack,performance increases but cost rises. As you move down the stack, performance declines but cost falls. EXHIBIT 3:Memory/Storage Hierarchy for AI from training and basic chatbot inference to advanced inference and agentic AI, memory and storage requirements change aswell. High Bandwidth Memory (HBM) and Conventional DRAM were the early beneficiaries, but NAND and HDD also benefitmore as AI workloads become more advanced. categories.What is different this time is that hyperscalers building AI data centers are far less price sensitive than historicalend customers. That dynamic has helped drive pricing up 3x to 4x over roughly the past six months. EXHIBIT 5:Data Center Mix for NAND, DRAM, and HDD Blackbook as a structural improvement thesis for the memory industry. Much of that thesis has played out, although not withoutexceptions. structural undersupply. In that setup, upcycles last longer and downcycles become less severe. As supply becomes harder,slower, and more expensive to add, and as industries consolidate, behavior improves and cycles show higher highs and higherlows. EXHIBIT 6:2013/2014 Black Book on the New Memory Paradigm growth, has continued to decline and has remained well below demand growth. That has made it more difficult and more time-consuming for supply to catch up in upcycles, while also reducing the severity of downcycles. EXHIBIT 8:DRAM Structural Supply Growth vs. Demand Growth today DRAM and HDD have also become highly concentrated industries, which has supported better behavior overall.NANDis a younger industry and has not consolidated to the same extent, but it is still relatively concentrated. EXHIBIT 9:Ind