A new report from Morgan Stanley cautions that the rally in semiconductor memory driven by artificial intelligence (AI) is approaching a turning point, with memory contract prices expected to peak in the fourth quarter. This signals a major shift in the current industry cycle.
According to the report, the momentum for upward earnings revisions among memory manufacturers is clearly fading. The rate of net earnings upgrades has fallen from a peak of 92% to 77%. This early indicator suggests the profit upgrade cycle is losing steam, and market pricing for the memory industry's earnings is moving from a euphoric state toward a more rational assessment.
Shares of industry giants have pulled back significantly from recent highs as a result. SK Hynix's forward price-to-earnings ratio for the next twelve months has recently declined, and its price-to-book (P/B) ratio has retreated to 2.5 times. Samsung's P/B ratio has fallen to 1.7 times. While valuations still sit above long-term averages, investors are already reacting to peaking growth rates and the potential for margin compression.
Simultaneously, inventory levels for DRAM and NAND rose in the second quarter, driven mainly by memory module makers. Although the rate of cyclical change is peaking, Morgan Stanley analysts believe this memory cycle will be prolonged rather than ending in a sudden collapse.
Strong AI Capital Spending Doesn't Directly Equate to Sustained Memory Price Increases
AI infrastructure investment is the most positive of Morgan Stanley's three core judgments and is the primary basis for expecting a longer memory cycle. The report's focus is not just on the number of model launches but on the training and inference intensity of leading large language models, funding sources, annualized recurring revenue, and the capital expenditure trajectory of hyperscale cloud providers. Capital spending by hyperscalers in the second quarter of 2026 is seen as a key verification point.
However, there is a critical logical gap between strong AI demand and continuously rising memory prices. The monetization of infrastructure does not automatically mean computing power is oversupplied. Improved model capabilities and falling prices do not guarantee that cloud providers will endlessly increase capital spending. What truly determines the slope of memory demand remains whether cloud providers continue to invest in compute, networking, and data center infrastructure.
Signals of a Price Peak Appear First in Inventory and Earnings Expectations
The judgment that memory prices will peak in the fourth quarter of 2026 is not based simply on extrapolating price curves, but on the convergence of signals from multiple dimensions. First, the year-on-year growth rate of DRAM contract prices has declined from its cycle high, while forward price-to-book ratios for memory stocks have not expanded significantly in tandem. This suggests capital is not pricing the industry on the premise of a new era of permanently high profits.
Second, inventories have shifted direction. Both DRAM and NAND inventory levels increased in the second quarter, driven primarily by large module makers. An inventory build does not necessarily mean weaker downstream demand, but when combined with a declining year-on-year price growth rate, rising module-level inventory amplifies market sensitivity to potential future de-stocking pressures.
Third, momentum in earnings expectations is fading. While the rate of net earnings upgrades remains positive, it has fallen from a peak of 92% to 77%, narrowing the room for further upward consensus revisions. The recent decline in SK Hynix's next-twelve-months earnings per share suggests the market has already begun to price in the possibility of slowing earnings growth in the next phase.
When these three signals appear simultaneously, the market's focus often shifts from "how much higher can profits go?" to "how long can profit growth last?" This is the classic path by which memory stock valuations come under pressure, even while the underlying fundamentals remain healthy.
Long-Term Agreements Haven't Triggered a Valuation Re-Rating; Samsung and SK Hynix Still Above Long-Term Averages
The long-term supply agreements (LTAs) between memory manufacturers and their downstream customers are a key factor supporting Morgan Stanley's neutral stance on one of its three core judgments. The market has not assigned a significantly higher valuation multiple to the industry because of these agreements. History from the pandemic period provides a reference: long-term agreements do not inherently eliminate cyclical risk. When prices and supply-demand dynamics shift, agreements can be renegotiated or lead to passive inventory accumulation by customers. Structural demand growth and traditional cyclical fluctuations can coexist; they are not mutually exclusive.
From a valuation perspective, Samsung's P/B ratio is around 1.7 times and SK Hynix's is around 2.5 times. Both have clearly retreated from recent highs but remain above their respective long-term averages. This perfectly reflects the meaning of the neutral judgment: the industry is not seen as a pure cyclical stock, but the narrative that "AI reshapes memory, leading to a comprehensive valuation re-rating" has not yet been established.
HBM Supply Tightness Remains a Hard Constraint; Commodity DRAM and High-End Memory Diverge
The divergence in supply structure is creating two distinct supply-demand dynamics for the memory industry. As Samsung, SK Hynix, and Micron shift more production capacity toward high-bandwidth memory (HBM), it objectively leaves more room for commoditized DDR5. However, the supply tightness for HBM itself is not alleviated by this shift. Commodity DRAM faces pressure from new supply catching up, while HBM remains constrained by advanced process nodes, advanced packaging, and bandwidth capabilities. These two markets are moving toward different supply-demand equilibriums.
High memory prices and bandwidth constraints also create a commercial foundation for a new wave of memory innovation. Morgan Stanley estimates the total addressable market for related technologies is approximately $25 billion, covering multiple technological paths centered around capacity, bandwidth, power consumption, and system architecture, rather than a single chip category. Long-term equipment demand is closely linked to the delivery pace of EUV lithography machines. Morgan Stanley forecasts ASML's EUV shipments will rise from 92 units in fiscal year 2027 to 104 units in fiscal year 2028, reflecting expectations for continued equipment demand in advanced logic and advanced memory manufacturing, but this does not mean all memory manufacturers will benefit equally.
Comments