U.S. Memory Deep Dive: Data-Center Storage Prices to Rise at Least 25% in 3Q; LTAs Cap the Peak, but the Micron and Sandisk Sell-Off Creates an Entry Point
目录
Too Long; Didn’t Read
I. The Variable the Market Mispriced: The Second Derivative, Not the Direction of Supply and Demand
II. Why Negative Consumer Signals May Be False Signals
III. LTAs Exchange Price Caps for Cycle Duration
IV. Why Configuration Cuts Are Both Negative and Evidence of Shortages
V. Micron and Sandisk: Duration Drives Valuation, but the Models Are Already Aggressive
VI. The Most Important Counterpoint: Why Morgan Stanley Still Prefers Nvidia and Broadcom
VII. Five Indicators to Watch: Mispricing or a Cycle Peak?
Conclusion: At Least 25% Data-Center Price Growth Is Only the Starting Point; Duration Is the Core Valuation Driver
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
Share prices have pulled back, while DDR5 spot prices continue to rise. The key debate is whether long-term agreements and configuration cuts will end the cycle or turn a one-off spike into a longer-lasting high-profit plateau.
Too Long; Didn’t Read
This sell-off is primarily a trade on the slope of price increases, not a reversal in price direction. Morgan Stanley’s procurement survey indicates that like-for-like data-center storage prices will still rise at least 25% QoQ in 3Q 2026, above both its own and third-party forecasts. Meanwhile, Micron’s valuation has fallen sharply since late June even as DDR5 16Gb spot prices continued to rise. This divergence between share prices and fundamentals is the report’s most important finding.
The deceleration in the second derivative of pricing is real, but it does not mean the cycle is over. According to industry data cited in the report, DRAM prices rose 70% QoQ in 1Q and more than 40% in 2Q. Morgan Stanley states that memory revenue exceeded US$200 billion this quarter, versus US$46 billion a year earlier. Its procurement survey also indicates that like-for-like data-center prices will rise at least 25% in 3Q. These datasets are not fully comparable and should not be mechanically combined into a single price curve, but both indicate that uncontrolled price increases are unsustainable while the direction of pricing remains upward.
Long-term supply agreements and configuration cuts will suppress the peak but may extend the cycle. LTAs provide customers with supply and price protection while giving suppliers longer order visibility. Morgan Stanley believes Nvidia has reduced rack-level LPDDR5 configurations, while cloud providers are also optimizing memory architectures, lowering memory content per system and the intensity of procurement competition. However, token volumes are growing roughly 10x annually and AI spending is rising by more than 50%, far ahead of the 3%–5% growth in traditional end markets. Customers are adapting to shortages, not losing their need for memory.
Micron offers more direct exposure to DRAM, HBM, and cycle duration, while Sandisk depends more heavily on earnings revaluation in NAND and enterprise SSDs. Based on July 17, 2026 prices, Morgan Stanley assigns Micron a US$1,200 price target, implying approximately 41% static upside, and Sandisk a US$1,750 target, implying approximately 29% upside. Both models are ultimately underwriting the persistence of exceptional profitability for two to three years, not merely an upside surprise next quarter.
This is not a report arguing that memory is already superior to every AI compute stock. Morgan Stanley still views Nvidia and Broadcom as offering better risk-reward within its coverage, but believes memory has rapidly closed the gap following the pullback. The most reasonable interpretation is that the sell-off has created a potential entry window that still requires further data confirmation. Price targets should not be treated as guaranteed returns.
I. The Variable the Market Mispriced: The Second Derivative, Not the Direction of Supply and Demand
The most valuable aspect of this report is that it acknowledges all three of the market’s concerns are valid: the pace of price increases will slow, capital expenditure will rise, and customers will reduce configurations. The disagreement is not over the facts, but over how those facts should be reflected in valuation.
In a traditional memory cycle, a peak in the rate of price increases often signals that peak earnings are approaching. The reason is straightforward: PCs, smartphones, and traditional servers typically grow only 3%–5%, so price increases quickly pressure end demand. Once supplier profitability improves, capacity expansion turns shortages into oversupply. The market therefore tends to sell before EPS reaches its peak.
What makes this cycle different is that data centers are almost solely determining marginal demand. Morgan Stanley notes that some recent signals weighing on share prices have come from consumer electronics, PCs, and smartphones. These markets do affect spot pricing and inventories, but they may not represent like-for-like procurement pricing in data centers. Following discussions with procurement teams, Morgan Stanley concluded that data-center shortages have not eased, could tighten further in 2027, and may intensify again in 2028.
A like-for-like data-center procurement price increase of at least 25% in 3Q therefore represents both deceleration and strength. Compared only with an increase of more than 40% in 2Q, it is clearly negative. Compared with a normal cycle, already-compressed market valuations, and still-rising DDR5 spot prices, it provides fundamental support. We previously discussed the divergence between the correction in memory stocks and continued strengthening in DRAM and NAND demand. The new evidence in this report is that data-center buyers remain willing to pay higher prices to secure delivery.
II. Why Negative Consumer Signals May Be False Signals
Understanding this cycle requires dividing the memory market into two layers. The first is consumer demand: smartphones, PCs, commodity NAND spot markets, channel orders, and inventory signals are intertwined and have the greatest influence on near-term sentiment. The second is data centers: HBM, server DRAM, rack-level low-power DDR5, and enterprise SSDs feature more concentrated procurement, longer qualification cycles, and higher costs of shortages.
