AI Data Centers Reshape NAND and HDD Demand Structure, Benefiting High-Capacity SSDs in the Short Term and Supporting HDD Oligopoly Discipline in the Long Term
AI summary card
AI Data Centers Reshape NAND and HDD Demand Structure, Benefiting High-Capacity SSDs in the Short Term and Supporting HDD Oligopoly Discipline in the Long Term
Bernstein's conference notes argue that AI workloads are pushing data centers from a two-tier storage architecture toward a three-tier architecture, driving a surge in demand for high-capacity enterprise SSDs, but the current NAND price premium of more than 20x over HDD makes HDD replacement uneconomic again.
- Traditional data centers once used HDDs for about 80%-85% of bit storage, but AI data preparation, vectorization, and pre-computation tasks are making high-capacity enterprise SSDs the core source of incremental demand.
- The price gap between NAND and HDD has widened from about 4-5x to about 20-25x, far above the 2-3x TCO conversion range, restraining migration from HDD to SSD.
- NAND shortages are concentrated in the latest-node high-capacity products, with related capacity accounting for only about 30%-35% of total industry capacity, so AI demand is driving sharp ASP increases.
- The HDD industry is dominated by a small number of manufacturers, and capacity expansion is constrained by the complexity of HAMR, heads, mechanical structures, and materials science, so supply discipline is expected to persist.
- Long-term agreements are shifting from one-sided supply commitments to two-way agreements that include customer purchase commitments and financial guarantees, helping improve supplier investment visibility.
- HBF uses stacking, wide interfaces, and advanced packaging similar to HBM to apply NAND's density and cost structure to streaming data scenarios in AI inference, but it remains in the early validation stage.
Report interpretation
Overview
This report is Bernstein's webinar notes with Robert Soderbery (former Western Digital EVP and former head of SanDisk's flash business), focused on industry changes in NAND, HDD, enterprise SSDs, long-term agreements, and HBF in AI data centers. The core conclusion is that AI is pushing data center storage from the traditional two-tier structure of "compute SSD + nearline HDD" toward a three-tier structure of "compute SSD + high-capacity enterprise SSD + nearline HDD," creating an incremental market for NAND; however, after the sharp rise in NAND prices, HDD replacement has slowed significantly on economic grounds, and the industry is more likely to move toward a more refined tiered architecture.
Core views
The report argues that high-capacity enterprise SSDs are indispensable in AI data center buildouts and that demand has relatively low price elasticity; however, the current dollar/GB price gap between NAND and HDD is too large, causing ordinary data centers and some AI scenarios to reconsider HDD and tiered storage. On the HDD side, a small number of suppliers, complex technology paths, and the experience of oversupply over the past decade jointly encourage restraint in capacity expansion. On the NAND side, the real shortage is in latest-node capacity and high-capacity QLC products; prices will not simply fall back quickly, but may enter a higher new normal unless the AI capex bubble fully bursts.
Analysis framework
The report uses expert interviews and an industry supply-demand framework to analyze the investment implications for SNDK, WDC, and STX around changes in data center architecture, TCO conversion prices, supplier capacity expansion constraints, the binding force of long-term agreements, node migration capex, QLC capacity improvements, and HBF technology potential.
Methodology notes
A small number of suppliers avoid expanding capacity too quickly when demand is strong in order to sustain returns.
Constrained by HAMR, heads, mechanical structures, materials science, and complex supply chains, HDD capacity expansion is not a simple linear increase, so the industry in which WDC, Seagate, and Toshiba operate has a degree of supply discipline.
When the cost gap between NAND and HDD narrows to about 2-3x, some hyperscale cloud providers may start switching from HDD to Flash.
Different hyperscalers have different architectures. Google is more HDD-centric, with a conversion point closer to 1.5x; Meta and others place more emphasis on application performance and may accept Flash earlier. But the current price gap of more than 20x makes conversion uneconomic.
Long-term agreements are shifting from only constraining supplier supply to including customer purchase commitments, pricing mechanisms, and financial guarantees.
This improves visibility for supplier investment and capacity planning, but if the market declines severely, customers may still reduce the impact by paying penalties, extending terms, or renegotiating.
HBF turns NAND into a stacked and wide-interface structure similar to HBM, used for large-scale streaming data reads in AI inference.
NAND's dynamic read/write performance and latency are not as good as DRAM, but once it enters streaming reads, it can provide relatively fast throughput and higher density; if power consumption, performance, and reliability meet requirements, it could become a new differentiated opportunity in AI storage.
Asset mapping & comparison
Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).
- SanDisk Corp (SNDK)A key beneficiary of NAND, QLC, high-capacity enterprise SSDs, and HBF
- Strengths
- AI data centers need high-capacity latest-node NAND; QLC is key to 128TB and 256TB SSDs; long-term agreements and financial guarantees improve revenue visibility.
- Weaknesses
- The NAND business has historically been highly cyclical, latest-node capacity expansion is capital-intensive, and investor communication and disclosure were once considered somewhat confusing.
- Comparison
- Compared with HDD companies, SNDK is more sensitive to NAND prices and latest-node supply; compared with traditional consumer NAND, its AI high-capacity SSD demand is more structural.
- Risks
- Near-term earnings may already be high, a rapid decline in NAND prices would cause cyclical downside, and if NAND weakness turns into a structural problem, DCF value may be revised downward.
- Western Digital Corp (WDC)A core beneficiary of HDD supply discipline, the HAMR transition, and cloud storage demand
- Strengths
- The HDD market has remained tight for 2-3 years, and the small number of industry manufacturers have the ability to maintain supply discipline; Bernstein reiterates Outperform and a USD 590 target price.
- Weaknesses
- Highly exposed to hyperscale cloud capex and changes in procurement models; the HAMR technology transition may weigh on gross margin and EPS.
- Comparison
- Compared with NAND manufacturers, WDC's HDD business is protected by current high NAND prices; however, if NAND technology improves faster than expected, HDD share could still come under pressure.
- Risks
- Cloud capex digestion, NAND substitution risk, and HAMR execution risk.
- Seagate Technology PLC (STX)A beneficiary of HDD supply discipline and HAMR leadership
- Strengths
- The report argues that Seagate's fundamentals are improving, its 5-year EPS CAGR exceeds 60%, and its HAMR leadership is sufficient to support a 21x P/E and a USD 1,000 target price.
- Weaknesses
- Also affected by hyperscale cloud demand and HDD procurement cycles.
- Comparison
- Compared with WDC, Seagate's investment narrative places greater emphasis on share and profit improvement from HAMR leadership.
- Risks
- Cloud capex slowdown, WDC catching up faster in HAMR, and NAND technology progress leading to HDD substitution.
- Pure Storage (PSTG)An indirect beneficiary of high-capacity Flash architectures and outsourced storage engineering capabilities for hyperscale customers
- Strengths
- Direct Flash Module bypasses traditional hard drive form factors, is designed around the optimal layout for NAND, and is gaining momentum among hyperscale customers.
- Weaknesses
- It still needs to compete with the economies of scale and standardized ecosystem of traditional enterprise SSDs.
- Comparison
- Compared with standard enterprise SSDs, Pure places more emphasis on system architecture, modules, and software combinations.
- Risks
- If customers shift back toward standardized SSDs or internal storage engineering, growth momentum may be constrained.
Key data
- Traditional Data Center HDD/NAND Bit Structureapproximately 80/20 to 85/15The expert said that in traditional data centers measured by GB or EB, the vast majority of capacity is on HDDs.
- Incremental AI High-Capacity Enterprise SSD Demandnew large business may be approximately 60%-70% NANDHigh-capacity SSDs from 32TB, 64TB, 128TB to 256TB have become a key competitive focus.
- NAND/HDD Price Gapapproximately 20-25xPreviously about 4-5x; currently far above the 2-3x TCO conversion range.
- TCO Conversion Rangeapproximately 1.5-3xMost markets may migrate around 1.5x, while some initial markets may begin migrating around 3x.
- Share of Latest-Node NAND Capacityapproximately 30%-35%AI high-performance and high-capacity SSDs are concentrated in consuming the latest nodes, causing localized extreme shortages.
- NAND Node Migration Capexapproximately USD 50bn per generation; replacing HDD may require 2-3 generations, approximately USD 150bnThis does not include the capex required to serve the AI market itself.
- SanDisk Purchase Obligations and Financial GuaranteesUSD 42bn in purchase obligations, of which USD 11bn are financial guaranteesFinancial guarantees account for about one-quarter, which can cushion but cannot fully eliminate downside price risk.
- NAND Capacity Expansion Timelineequipment takes about 1 year for delivery, followed by several months of debugging; overall planning cycle is about 15 monthsPrices need to remain high for several quarters before manufacturers are more likely to commit to meaningful capex.
Impact & implications
From an investment perspective, the report supports structural opportunities in both NAND and HDD: SNDK benefits from AI high-capacity SSDs, QLC, long-term agreements, and potential HBF upside; WDC and STX benefit from HDD supply discipline, HAMR cost curves, and continued cloud storage demand. In the short term, NAND price increases will suppress some consumer and client demand, but high-end AI demand remains strong; HDD may regain some demand share in AI data centers.
Risks
- If AI capex cools broadly, NAND prices and high-capacity SSD demand may fall below expectations.
- If NAND prices fall rapidly from high levels, long-term agreements may have financial guarantees, but protection could still be weakened through renegotiation.
- Hyperscale cloud customers may use software methods such as deduplication, compression, and tiered management to reduce storage demand.
- If NAND technology delivers a larger-than-expected cost decline, it could again increase the long-term risk of SSDs replacing HDDs.
- The HAMR transition for HDD manufacturers may bring risks to technology, yield, gross margin, and EPS.
- Consumer and client SSD demand may be compressed by high prices, for example shifting from 1TB configurations down to 256GB.
What to watch
- The pace of migration to latest-node NAND capacity and equipment delivery cycles.
- Whether AI high-capacity enterprise SSD prices enter a new normal or fall back quickly.
- Subsequent disclosures on purchase obligations, financial guarantees, and third-party financial instruments in SNDK's long-term agreements.
- Whether HDD manufacturers continue to maintain capacity expansion discipline, and progress in HAMR mass production.
- Whether hyperscale cloud providers increase the HDD share again in AI data centers.
- Validation progress for HBF in power consumption, performance, reliability, and specific AI applications.
- Yield, performance, and supplier differences for QLC in 128TB and 256TB enterprise SSDs.