Morgan Stanley warns: AI storage frenzy approaching a turning point, memory prices may peak in Q4

According to the latest report by Morgan Stanley, the AI-driven semiconductor memory industry boom is nearing 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 Chasing Wind Trading Desk, Shawn Kim, Head of Asia and Europe Technology Research at Morgan Stanley, stated in a report released on July 21 that the momentum for upward earnings revisions among memory manufacturers is weakening significantly, with data showing the net earnings upgrade rate has dropped from its 92% peak to 77%. This early signal indicates the earnings upgrade cycle is losing traction, and the market pricing for memory industry profits is returning to rationality rather than the previous state of hype.

As a result, industry giants' valuations have dropped significantly from recent highs. SK Hynix's one-year forward EPS has declined recently, with its price-to-book ratio falling to 2.5x, while Samsung's has dropped to 1.7x. Although valuations remain above long-term averages, investors have begun to react to peaking growth rates and potential profit contraction.

Meanwhile, DRAM and NAND inventory levels rose in the second quarter, mainly driven by memory module manufacturers. Although the cycle trend is peaking, Morgan Stanley analysts believe the current memory cycle will be extended, rather than collapsing abruptly.

AI infrastructure investment is the most optimistic among Morgan Stanley’s three core judgments and is the main reason supporting the prolonged memory cycle this round.

The report focuses not only on the number of model releases but also on the training and inference intensity of major AI models, funding sources, annual recurring revenue, and capital expenditure trends of hyperscale cloud vendors. Among these, hyperscale capex in Q2 2026 is seen as a key validation point.

However, there is a significant logical gap between strong AI demand and sustained memory price increases. The start of monetization for infrastructure does not equate to an oversupply of computing power; improvements in model capability and falling prices do not automatically mean cloud vendors will continually increase capex. The real determinant of memory demand slope remains whether cloud vendors continue to allocate capital to computing power, network, and data center infrastructure.

The judgment that memory prices will peak in Q4 2026 is not based solely on extrapolating the price curve but corroborated by multiple indicators.

First is the year-on-year price growth deceleration. DRAM contract price YoY growth has retreated from its cycle high, while the market's forward price-to-book valuations for memory stocks have not expanded sharply, implying that funds are not pricing in "a new round of permanently high earnings" for the industry.

Second is the directional change in inventories. DRAM and NAND inventories both increased in Q2, primarily driven by large module manufacturers. Inventory buildup itself does not necessarily mean downstream demand is deteriorating, but with price growth already decelerating, rising inventories at the module end heighten the market's sensitivity to subsequent destocking pressure.

Third is the loss of momentum in earnings expectations. The ratio of net upward earnings revisions, although still positive, has dropped from a peak of 92% to 77%, and consensus expectations for further upgrades are becoming limited. The recent decline in SK Hynix’s forward year EPS also suggests the market has begun to price in the potential for slowing profit growth in the next stage.

When all three signals appear simultaneously, market focus often shifts from "how much more can profits rise" to "how long can profit growth be sustained"—this is the classic path where memory stocks start to feel valuation pressure while fundamentals are still healthy.

Long-term supply agreements (LTA) signed between memory manufacturers and downstream customers are a key basis for the neutral stance among Morgan Stanley's three judgments. The market has not given the industry significantly higher valuation centers simply due to the existence of these agreements.

This logic has historical precedent during the pandemic: long-term agreements do not naturally eliminate cycle risk. When prices and supply-demand relationships change, agreements may be renegotiated, or customers might passively accumulate inventory. Structural demand enhancement and traditional cycle volatility can coexist—they are not mutually exclusive.

In terms of valuation, Samsung’s price-to-book is about 1.7x, SK Hynix about 2.5x, both having retreated significantly from recent highs, but still above respective long-term averages. This pattern aligns precisely with the neutral thesis: the industry is not seen as a purely cyclical stock, but the narrative of "AI reshaping memory and a comprehensive valuation re-rating" has not yet been established either.

Divergence in supply structure is leading to two distinct paths of supply and demand logic in the memory industry. Samsung, SK Hynix, and Micron are allocating more capacity resources to high-bandwidth memory (HBM), which objectively frees up space for mainstream DDR5; but the supply-tight situation for HBM itself has not eased as a result.

Commodity DRAM is under the pressure of new supply coming online, while HBM remains limited by advanced process, advanced packaging, and bandwidth capability constraints. The two markets are heading towards different supply-demand equilibrium paths.

High memory prices and bandwidth limitations also lay the commercial foundation for a new round of memory innovation. Morgan Stanley estimates the related market’s total addressable market at about $25 billion, covering multiple technology paths around capacity, bandwidth, power consumption, and system architecture—not just a single chip category. Long-term equipment demand is closely related to EUV delivery schedules, with Morgan Stanley forecasting ASML EUV shipments to rise from 92 units in the 2027 fiscal year to 104 units in 2028, supporting expectations for continued capacity expansion in advanced logic and advanced memory manufacturing, but not meaning all memory manufacturers will benefit simultaneously.