数据中心的HDD与NAND：前西部数据闪迪业务执行副总裁访谈纪要及要点

Memory/Storage: HDD vs. NAND in the Datacenter with ex-WDCEVP, SanDisk Flash Business - Transcript and Takeaways We recently hosted a webinar with Robert Soderbery, former EVP at Western Digital whoran the Sandisk flash business, overseeing cloud, consumer, and AI storage markets. Wediscussed the dynamic of NAND vs. HDD in data centers as well as LTAs and HBF. This noteis an edited transcript with key takeaways. A replay is available here. Mark C. Newman+1 212 845 7822mark.newman@bernsteinsg.com April Li+1 917 344 8339april.li@bernsteinsg.com Data center storage has shifted from a simple two-tier model - HDDs for capacityand small SSDs for compute - to a three-tier architecture driven by AI.Historically,HDDs represented about 80-85% of bits, but AI workloads created strong demand forhigh-capacity enterprise SSDs to support data prep and pre-compute tasks that HDDscannot handle, pushing new AI storage deployments temporarily toward NAND. Phoebe Sun+1 917 344 8481phoebe.sun@bernsteinsg.com Sharp NAND price increases have widened the ASP gap vs. HDD, slowed HDD-to-SSD substitution, and forced AI operators toward more cost-efficient tiereddesigns.SSD displacement of HDD is now economically unattractive for both hyperscalersand NAND suppliers.The hyperscaler TCO crossover requires a ~2-3x NAND-to-HDD ASP gap; the current gap exceeds 20x, making substitution uneconomic. For NANDsuppliers, displacing HDD would require massive incremental capex across multiple nodetransitions at low returns and highly unlikely. Two additional trends could reshape the NAND landscape.First, the industry is movingtoward true two-way long-term agreements- replacing historically one-sided LTAswith contracts that include customer purchase commitments and financial guarantees,improving planning visibility and supporting investment (though severe downturns wouldlikely prompt renegotiation over strict enforcement).Second, any shift of compute fromcloud to edge strongly favors NAND over HDD given NAND's advantages in space,power efficiency, and performance. NAND pricing is spiking because AI demand is concentrated on the latest-node,high-capacity NAND, which represents only ~30-35% of total industry capacity, so theconcentrated demand is creating a very sharp increase in ASP. HDD pricing reflects a moreconventional supply-demand imbalance: demand exceeds supply, but less extreme. HDD vendors operate in a tight oligopoly and are intentionally avoiding capacityexpansion. HDD technology remains complex enough - across HAMR, heads, mechanics,and materials science - that adding capacity has real friction, supporting LT supplydiscipline. HBF uses similar-to-HBM stacking, wide interfaces, and advanced packaging to sitalongside GPUs, but exploits a key AI inference characteristic: data is often streamedrather than frequently rewritten.Although NAND has slow dynamic read-write andhigher latency, it can deliver very fast sustained throughput once streamingbegins, making it well suited for AI inference. By combining HBM-style bandwidth withNAND’s much higher density and different cost structure, HBF could be compelling for AIworkloads, though it is still early stage. BERNSTEIN TICKER TABLE DETAILS We recently hosted a webinar with Robert Soderbery, former EVP at Western Digital who ran the Sandisk flash business,overseeing cloud, consumer, and AI storage markets. We discussed the dynamic of NAND vs. HDD in data centers as well as LTAsand HBF. This note is an edited transcript with key takeaways. A replay is available here. This transcript has been edited for clarity and consistency. Speaker Key: RS:Robert Soderbery MN:Mark Newman MN: Good morning, everyone. I'm Mark Newman, Bernstein's US IT Hardware Analyst, and great pleasure to welcome RobertSoderbery today, as our special expert in the industry. Rob, I think it'll be really great, since your background, in flash and workingat Western Digital, it'll be good to know, how you see the current mix of NAND versus HDD in data centers. How has thatchanged over time, and how do you expect that to evolve over time? RS: Yeah, happy to help, Mark. I think to understand it, it's best to walk back a little bit, because the situation's gotten morecomplex with the arrival of the AI data center. But let's start with the conventional data center. And recognize that thehyperscalers don't differentiate in their purchasing, typically, between these two, so it can be actually quite hard to disentanglethe whole thing. So, starting with the conventional data center, the world was pretty simple You used, Enterprise SSDs forcompute, these are small drives - 2, 4 or 8TB drives, and you used, Enterprise Nearline HDDs for storage. And so you had really a two-tier architecture. If you're going to measure in terms of gigabytes or exabytes, the vast quantity ofthat was in HDD, although, of course, those HDDs are much lower cost, so from a spend perspective, it's more balanced. Thatwas a world that was 80/20, 8