目录
I. Legacy-Product Shortages: Profiting From the Timing Gap Between Supply Exits and Demand Declines
Why Major Vendors Are Abandoning Products That Are Still Rising in Price
Why Legacy Customers Have Not Left at the Same Pace
Shortages Can Persist, but Price Acceleration Cannot Continue Indefinitely
II. How AI Creates DDR4 Demand: Start With Capacity per Drive, Then Calculate Supplier Revenue
1GB of Cache and an 8Gb Die Are the Same Order of Magnitude
Brokerage Estimates Imply a Large Market, but It Remains Potential Demand
NOR Provides Another Growth Driver, but the Two Demand Streams Must Not Be Double-Counted
III. How Price Increases Translate into Profit: Nanya Offers a Clear Case Study
Revenue Growth Began with Higher Revenue per Unit
Gross-Margin Expansion Does Not Require an Equivalent Decline in Costs
Winbond’s Growth Reflects Both Pricing and Business-Mix Changes
Stress-Testing Prices Is More Useful Than Focusing Only on EPS Growth
IV. Both Benefit from Higher Legacy-Memory Prices, but Nanya and Winbond Have Different Earnings Mixes
Nanya: Greater Pricing Sensitivity Requires Stronger Capital Discipline
Winbond: NOR Can Improve Business Quality but Cannot Yet Replace DRAM as the Core Driver
Profit Quality Must Be Compared Using a Consistent Cash-Flow Definition
V. When Does Capacity Expansion Change Supply-Demand—and When Does It First Consume Cash?
Winbond’s 25,000-Wafer Target Does Not Directly Imply 67% Output Growth
Nanya’s New Fab Must Deliver Both Growth and Technology Upgrades
Assess Chinese Supply by Product Overlap Before Aggregate Scale
YMTC Changes NAND Supply First, Then DRAM Cache Demand Indirectly
VI. The Duration of Peak Earnings Determines Whether a Low P/E Is Meaningful
The Two Brokers’ Price Targets Rely on Different Earnings and Return Assumptions
Estimate Mid-Cycle Earnings Before Selecting a Multiple
Sustainable Growth Must Survive the Cyclical Downturn
How the Next Data Points Should Change the Investment View
本内容基于公开资料和研报数据整理，不构成任何投资建议，不代表任何个人观点，仅供学习参考，请理性阅读
Nanya and Winbond’s profit surge began with price increases following supply cuts in legacy products. AI is now adding demand for enterprise SSD cache and server NOR, potentially extending the shortage window. Whether today’s high profits endure will depend on customer migration, new capacity, and the cash left after capital expenditure.
I. Legacy-Product Shortages: Profiting From the Timing Gap Between Supply Exits and Demand Declines
DDR4’s profitability depends on whether customers reduce purchases faster or slower than major vendors withdraw supply. Mature technology and slowing end-market growth do not necessarily imply ample supply. If capacity exits faster than demand declines, remaining suppliers can still raise prices and materially increase profits even without shipment growth.
Nanya Technology offers a clear example. Morgan Stanley’s July 12, 2026 report noted that second-quarter bit shipments were flat quarter on quarter, while average selling prices rose by more than 60%, driving a 68% sequential increase in revenue. Gross margin reached 79.5%, about 12 percentage points higher than in the previous quarter. Shipped memory capacity barely increased, but revenue per unit of capacity rose sharply—the principal driver of the earnings inflection.
This example also frames the analysis. Nanya Technology and Winbond Electronics may indeed be winning more AI-related orders, but the entire earnings surge cannot be attributed to AI demand growth. Scarcity premiums created by supply reductions are already sufficient to transform a memory manufacturer’s profitability over a short period.
Why Major Vendors Are Abandoning Products That Are Still Rising in Price
HBM and advanced server memory have raised the opportunity cost of producing conventional DRAM. Wafer capacity, equipment budgets, R&D; personnel, and customer-support resources are finite, forcing companies to compare the returns that different products can generate over the next several years. A legacy product’s current price increase does not automatically make it the most attractive destination for new investment.
Morgan Stanley’s August 26 report on China’s memory competition linked HBM’s consumption of advanced DRAM resources to tightening supply of conventional products. This mechanism should not be reduced to “one more HBM wafer immediately means one fewer DDR4 wafer.” The products involve different processes, production-line conversions, yields, and packaging arrangements; actual displacement typically occurs gradually through investment decisions and product-mix shifts.
The more consequential development for Nanya and Winbond is major vendors’ diminishing willingness to maintain the same level of incremental capital investment in DDR4. As long as advanced products continue to offer stronger long-term returns, major vendors may extend supply of selected legacy products without restoring full expansion plans. Conventional DRAM supply elasticity could therefore remain lower than in the previous cycle.
“Exit,” however, does not mean that every vendor stops deliveries on the same day. Long-term customer procurement arrangements, the residual value of existing production lines, and prevailing prices all affect the actual timetable. Announcing end-of-life, accepting last-time orders, making final shipments, and materially reducing effective output occur at different points. The duration of a shortage should be assessed by the volume of qualified products that can actually be delivered.
Why Legacy Customers Have Not Left at the Same Pace
Migrating away from DDR4 often requires corresponding changes to controllers, motherboard designs, firmware, or certification. In long-lived industrial equipment and consumer products, memory is only one component of the broader system. Even if a more advanced product offers a lower price per bit, redesign and recertification still carry additional costs.
The customer decision can be reduced to a comparison: the incremental cost of continuing to buy higher-priced DDR4 versus the development expense, validation time, and supply risk of changing platforms. If the former is lower in the near term, customers may accept the price increase. The longer elevated prices persist, the more economically attractive migration becomes.
Price increases therefore benefit suppliers while also accelerating substitution. They initially lift margins on existing orders, then encourage customers to reduce capacity, adjust product mixes, or accelerate platform upgrades. The more suppliers depend on high-priced purchases from the same legacy customers, the more closely they must monitor whether those customers are reducing overall procurement volumes.
JPMorgan’s July 15 report on Winbond projected that third-quarter contract prices for DDR3- and DDR4-related products would rise by approximately 50% quarter on quarter, above its previous assumption of 20%—30%. The report also indicated that consumer customers were still accepting higher prices. This demonstrates strong pricing pass-through at the time, but does not by itself establish that customers can absorb comparable increases for several years.
Shortages Can Persist, but Price Acceleration Cannot Continue Indefinitely
The supply-demand gap, absolute price level, and rate of price increases should be analyzed separately. Even while the industry remains undersupplied, prices may stay elevated as sequential increases narrow each quarter. A subsequent slowdown in manufacturers’ revenue growth would not necessarily mean the shortage has ended. Conversely, continued price increases may partly reflect inventory accumulation and cannot alone confirm end demand.
The July report on Nanya favored a continued upcycle in specialty DRAM, while the July Winbond report argued that enterprise SSDs could sustain tightness over the next two years. The August report on China’s memory competition identified 2028 as a key window for assessing the impact of new supply on industry returns. These are conditional timelines, not guarantees that elevated profitability will persist until a specific date.
The most important indicators for this window are like-for-like contract prices, manufacturers’ realized selling prices, bit shipments, and customer inventories. Simultaneous improvement across all four would more convincingly indicate growth supported by both end demand and supply conditions. If selling prices rise while shipments continue to decline, current profits become increasingly dependent on scarcity premiums, raising subsequent substitution risk.
II. How AI Creates DDR4 Demand: Start With Capacity per Drive, Then Calculate Supplier Revenue
Enterprise SSDs provide a clear pathway from AI demand to conventional DRAM, but translating “cache is required” into “how much Nanya and Winbond earn” requires assumptions about drive type, capacity, supplier qualification, and market share. Each step requires corresponding evidence.
1GB of Cache and an 8Gb Die Are the Same Order of Magnitude
JPMorgan’s July 15 report stated that a 1TB enterprise SSD typically requires approximately 1GB of DRAM cache, commonly using DDR4 8Gb. Lowercase “b” denotes bits and uppercase “B” denotes bytes; 8 bits equal 1 byte, so 8Gb corresponds to 1GB. Misstating 8Gb as 8GB would overstate total demand eightfold.
Following the report’s example, a 7.68TB SSD configured with approximately 1GB per TB would require about 7.68GB of DRAM. This is the output of a proportional calculation, not evidence that an actual product necessarily contains exactly 7.68GB of memory. Real products use feasible component configurations, and capacity, channels, and die density may require rounding.
DRAM in these SSDs supports controller operations, including mapping management and caching. Higher SSD NAND capacity may require additional corresponding resources, but the precise amount also depends on controller architecture, mapping methodology, performance targets, and firmware. The demand relationship is sufficient for this analysis; each caching mechanism need not be treated as a separate incremental market.
In particular, a “typical enterprise SSD configuration” cannot be extrapolated directly to all SSDs. Whether different drive types use discrete DRAM, which DRAM generation they adopt, and how much they require per TB must be verified separately. Without disclosed configurations for actual products, industry estimates must remain scenario-based.
Brokerage Estimates Imply a Large Market, but It Remains Potential Demand
JPMorgan projects enterprise SSD shipment capacity of more than 515EB in 2026, 840EB in 2027, and more than 1,190EB in 2028. Under its assumption of approximately 1GB of DRAM per TB, 840EB in 2027 represents 840 million TB of SSD capacity and approximately 840 million GB of potential DRAM demand.
The calculation uses decimal capacity units, with 1EB equal to 1 million TB. The 840 million GB figure refers to DRAM capacity, while 840EB refers to SSD NAND capacity; the two cannot be compared directly in the same column. The estimate nevertheless indicates that cache demand could be large enough to warrant inclusion in specialty DRAM supply-demand models.
However, these enterprise SSD capacity figures cover the entire enterprise SSD market and do not imply that AI creates all shipments. Traditional databases, virtualization, enterprise applications, and replacement of existing equipment also generate enterprise SSD demand. Estimating the AI-driven increment requires comparison with a procurement baseline absent the current AI expansion, or direct shipment evidence from AI-related customers and products.
