Morgan Stanley: Memory stocks often look cheapest at the cycle top, but they are also the most dangerous
Original author: Li Jia
Original source: Wallstreetcn
The AI-driven memory chip supercycle is still ongoing, but the investment logic for memory stocks is changing.
Morgan Stanley's latest report points out that although tight supply and demand, rising prices, and improving profitability continue to provide strong support for the industry, as stock prices have risen sharply and earnings expectations have been revised upward in a concentrated manner, the optimistic expectations for memory stocks have already been fully priced in, and the room for further gains is being tested.
The core contradiction is that the stronger the fundamentals of the memory industry, the easier it is to underestimate the risk of a cyclical peak. Historically, memory stocks have often peaked before earnings reached their peak, because the market habitually extrapolates current high profit levels into the future, causing stocks to appear cheapest in valuation precisely when industry prosperity is at its highest.
Against this backdrop, Morgan Stanley believes that the strong prosperity of the memory industry may not necessarily translate into sustained excess returns for memory stocks. As the market's expectations for earnings growth continue to rise, the relative investment appeal of the AI compute and networking supply chain is gradually surpassing that of the memory sector.
The supercycle continues, but cyclicality has not disappeared
The prosperity of this round of the memory industry has a solid demand foundation.
The construction of AI data centers continues to drive up memory demand, with high-bandwidth memory (HBM) accounting for nearly half of the cost of AI server chips. Long-term supply agreements improve suppliers' revenue visibility by locking in procurement commitments and setting price floors. At the same time, HBM production crowds out traditional DRAM capacity, and the demand for memory capacity from a new generation of AI models continues to rise, further intensifying the supply-demand tightness.
Morgan Stanley data shows that the current price per GB of HBM has reached nearly 7 times its historical peak, reflecting the strength of this memory cycle. But the structural strengthening of industry growth does not mean that the investment returns of memory stocks will improve accordingly. High profits will eventually attract new capacity, and as supply catches up with demand, or as downstream enters a phase of inventory digestion, prices and profits may peak. The historical cycles of 2000, 2016, 2019, and 2022 all reflected this pattern.
The report believes that the inflection point of this cycle may appear later, but it is still necessary to be vigilant about changes in the supply-demand landscape after 2027.
Earnings expectations are too high, and low P/E ratios hide traps
Since this cycle began in September 2025, Asian memory stocks have risen about 9 times cumulatively, and their total market value has entered the trillion-dollar level. Morgan Stanley data shows that the consensus earnings per share (EPS) estimates for Asian memory companies have been revised upward by an average of 432%, reflecting extremely high market expectations for the earnings outlook of this cycle.
However, continued earnings growth does not mean that stock prices still have room to rise by the same magnitude. Samsung and SK Hynix currently have forward P/E ratios of about 3 to 4 times. On the surface, valuations look cheap, but this may also be a valuation trap that appears near the earnings peak of cyclical stocks. Morgan Stanley points out that to support the further upside implied by current stock prices, DRAM contract prices would still need to rise by about 70% quarter over quarter from current levels, which would be very difficult to achieve.
Historical experience is equally worth noting: in 2008 and 2012, memory stocks peaked about 4 months before earnings peaks; in 2018, this time gap widened to 8 months.Memory stocks often take the lead in reflecting market concerns about a future downturn in prosperity when earnings are strongest and valuations appear cheapest.
HBM and DRAM prosperity continues, while supply risks gradually approach
In the short term, the fundamentals of the memory industry are still supported. Morgan Stanley expects HBM prices to continue rising in 2027, and contract prices for traditional DRAM and NAND are also expected to move further higher in the fourth quarter of 2026. Strong AI server demand, the shift of traditional DRAM capacity to HBM, and growing demand for enterprise solid-state drives (eSSDs) jointly support the current price increase trend.
However, changes on the supply side cannot be ignored.Memory manufacturers are significantly increasing capital expenditures, and new capacity is expected to be gradually released starting in the second half of 2027. At the same time, the consumer-grade NAND market may face pressure from falling prices, and the prosperity performance of different memory categories may diverge.
The evolution of data center architecture may also change the distribution of returns across the industry chain. Morgan Stanley estimates that memory-related content accounts for about 29% of Nvidia's NVL72 single-rack system, falling to 19% when expanded to 8 interconnected racks, while the share of networking equipment rises from 10% to 20%.As the scale of AI clusters expands, the weight of networking hardware in system value increases, and memory manufacturers may not benefit proportionally from the growth in infrastructure investment.
