Morgan Stanley warns: AI storage frenzy approaching inflection point, memory prices may peak in the fourth quarter.

According to the latest report from Morgan Stanley, the frenzy in the semiconductor storage industry driven by artificial intelligence (AI) is approaching a turning point, with memory contract prices expected to peak in the fourth quarter, signaling a profound shift in the current industry cycle. According to Windcatcher Trading, Shawn Kim, Asia and Europe technology research director at Morgan Stanley, stated in a report released on July 21 that the market's momentum for profit upgrades for memory manufacturers is significantly weakening, with data showing that the net profit upgrade rate has fallen from a peak of 92% to 77%. This early signal indicates that the profit upgrade cycle is losing momentum, and the market's pricing of profits in the storage industry is returning to rationality, rather than the previous frenzy. As a result, valuations of industry giants have sharply retreated from recent highs. SK Hynix's one-year forward earnings per share has recently declined, with its price-to-book ratio falling to 2.5 times, and Samsung's price-to-book ratio dropping to 1.7 times. Although valuations are still above long-term average levels, investors have started to react to the peak growth rates and potential profit contraction. Meanwhile, inventory levels for DRAM and NAND in the second quarter have seen an increase, driven mainly by memory module manufacturers. While the rate of cyclical change is peaking, Morgan Stanley analysts believe that this round of storage cycle will lengthen rather than collapse. AI capital spending remains strong, but does not directly imply continuous upward trend in storage prices AI infrastructure investment is the most positive of Morgan Stanley's three core judgments and a key basis for the prolonged period of prosperity in the current storage cycle. The focus of the report's observations is not only on the number of model releases, but also on the training and inference intensity of the top models, funding sources, annual recurring revenue, and the capital expenditures trend of hyperscale cloud providers. The hyperscale capital expenditure in the second quarter of 2026 is seen as a more crucial validation point. However, there is a significant logical gap between strong AI demand and continuous upward trend in storage prices. The monetization of infrastructure does not mean compute power is oversupplied; improvement in model capabilities and price declines do not automatically imply that cloud providers will indefinitely increase capital expenditures. The real determinant of the slope of storage demand is whether cloud providers will continue to invest in compute power, networks, and data center infrastructure. Signs of a price turning point first appear in inventory and profit expectations The judgment that storage prices will peak in the fourth quarter of 2026 is not based solely on extrapolating price curves, but on the accumulation and confirmation of signals from multiple dimensions. Firstly, the year-on-year growth rate of prices is declining. The year-on-year growth rate of DRAM contract prices has fallen from its cyclical peak, and the forward price-to-book ratios of storage stocks have not expanded significantly in sync, indicating that capital is not pricing the industry based on the assumption of "a new round of permanently high profits". Secondly, there is a directional change in inventory. Inventory for DRAM and NAND in the second quarter has increased, mainly driven by large module manufacturers. While inventory increases do not necessarily mean downstream demand deterioration, against the background of declining year-on-year price growth, an increase in module inventory will amplify the market's sensitivity to future destocking pressures. Thirdly, there is a slowdown in the momentum of profit expectations. While the net upward revision proportion of profit expectations is still in positive territory, it has fallen from a peak of 92% to 77%, with limited room for consensus expectations to continue to be revised upwards. The recent decline in SK Hynix's EPS for the next year indicates that the market has begun to digest the possibility of a slowdown in the next stage of profit growth. When these three signals appear simultaneously, the market's focus on trading often shifts from "how much more can profits rise" to "how long can profit growth be sustained" this is the typical path where the valuation of storage stocks begins to be pressured even when fundamentals remain healthy. Long-term contracts have not led to a revaluation, Samsung and SK Hynix still above long-term averages Long-term supply agreements (LTA) signed by storage manufacturers with downstream customers are one of the key reasons for Morgan Stanley's neutral stance in its three core judgments. The market has not assigned significantly higher valuations to the industry just because of the existence of these agreements. This logic has historical references from the pandemic period: long-term agreements do not naturally eliminate cyclical risks. When prices and supply-demand dynamics change, agreements may be renegotiated or lead to passive inventory accumulation by customers. Structural demand enhancement can coexist with traditional cyclical fluctuations, and the two are not mutually exclusive. In terms of valuation levels, Samsung's price-to-book ratio is around 1.7 times, while SK Hynix's is around 2.5 times, both have significantly retreated from recent highs but still remain above their respective long-term averages. This situation accurately corresponds to the neutral judgment: the industry is not seen as purely cyclical, but the narrative of "AI reshaping storage, comprehensive revaluation of valuations" has not yet been established. Tight supply of HBM remains a hard constraint, divergence between commodity DRAM and high-end storage The differentiation in supply structure is bringing two different supply-demand logics to the storage industry. Samsung, SK Hynix, and Micron are allocating more capacity resources to high-bandwidth memory (HBM), creating space for commoditized DDR5; however, the tight supply situation of HBM itself has not been relieved because of this. Commodity DRAM is facing pressure from increased supply, while high-bandwidth memory continues to be constrained by advanced process, packaging, and bandwidth capabilities. The two markets are moving towards different supply-demand equilibrium paths. The high prices of storage and limited bandwidth also form the commercial basis for a new round of storage innovation. Morgan Stanley estimates the total addressable market size of related markets to be around $25 billion, covering multiple technological paths focused on capacity, bandwidth, power consumption, and system architecture, rather than a single chip category. Long-term demand on the device side is closely related to the delivery pace of EUV, with ASML's EUV shipment expected to increase from 92 units in the 2027 fiscal year to 104 units in the 2028 fiscal year, corresponding to the continuous expansion of equipment demand for advanced logic and advanced storage manufacturing, but this does not mean that all storage manufacturers will benefit simultaneously. This article is reprinted from "Wall Street News"; GMTEight editor: Li Fo.