Morgan Stanley has warned that Asian memory-chip stocks can appear inexpensive precisely when the market is closest to a turning point, despite record earnings forecasts and still-rising prices for AI-related components. The bank said Samsung Electronics and SK Hynix trade at forward price-to-earnings multiples of roughly three to four times, but argued that current valuations already assume an unusually aggressive next leg in memory pricing.
In Morgan Stanley’s model, DRAM contract prices would need to rise by about 70% quarter on quarter from current levels to support the additional upside implied by share prices. That leaves little margin for operational setbacks, weaker-than-expected demand, or a moderation in the pace of price increases after a powerful sector rally.
The warning arrives after Asian memory stocks surged about ninefold from the beginning of the current cycle in September 2025, according to Morgan Stanley. The rally pushed the sector’s aggregate market value into the trillion-dollar range while consensus earnings-per-share forecasts climbed by an average of 432%.
Low multiples reflect extraordinary earnings assumptions
A low forward P/E multiple usually suggests that a company’s shares are cheap relative to expected profits. In cyclical industries such as memory chips, though, the figure can be misleading when earnings forecasts are at or near a peak.
Morgan Stanley’s analysis rests on that distinction. Samsung and SK Hynix may look inexpensive against projected earnings, but those projections depend on a supply shortage lasting longer and producing steeper price increases than the market has historically sustained.
Memory manufacturers have benefited from the rapid expansion of AI data centers, which require large quantities of high-performance memory and storage. Producers have also shifted some manufacturing capacity away from conventional DRAM toward high-bandwidth memory, or HBM, tightening availability in other parts of the market.
That combination has improved pricing power. Yet Morgan Stanley said the market has already priced in much of the favorable scenario, making share performance increasingly sensitive to whether companies can keep surpassing elevated earnings expectations.
The bank also pointed to previous memory cycles in which equities peaked before profits did. In 2008 and 2012, memory-chip shares topped out around four months before earnings reached their high point, Morgan Stanley said. In 2018, the gap widened to about eight months. Those episodes show why equity markets can start discounting a future slowdown even while reported results remain strong.
HBM pricing reaches historic levels
HBM remains the clearest source of strength in the current cycle. Morgan Stanley said the price per gigabyte of HBM is now near seven times the previous historical peak, reflecting its importance in AI servers and the limited number of suppliers able to manufacture advanced products at scale.
HBM stacks multiple layers of memory vertically to increase bandwidth and reduce the physical space required near high-performance processors. That makes it particularly useful for AI accelerators, where rapidly moving large volumes of data is as important as raw computing capacity.
Morgan Stanley expects HBM prices to continue rising through 2027, supported by AI server construction and the diversion of DRAM production lines toward HBM. The forecast gives Samsung and SK Hynix a favorable near-term earnings backdrop, although it also places more pressure on them to execute against already demanding market expectations.
The bank projected that conventional DRAM and NAND contract prices could also rise further in the fourth quarter of 2026. It cited sustained AI-server demand, reduced DRAM capacity following the shift toward HBM, and stronger enterprise solid-state drive demand.
Enterprise SSDs are storage drives used in servers and data centers. Their demand is becoming more relevant as AI systems require not only memory near processors, but also fast storage for training data, model checkpoints and other large workloads.
Capacity additions could split the memory market
Morgan Stanley expects supply additions to emerge gradually from the second half of 2027. The timing matters because new capacity would arrive after a period in which producers have limited output growth and redirected manufacturing resources toward higher-margin AI products.
The bank said consumer NAND could face price pullbacks as fresh capacity enters the market. That creates the prospect of diverging conditions within memory: HBM could remain constrained by specialized manufacturing requirements and AI demand, while more standardized NAND products used in consumer devices could come under greater pressure.
Such a split would complicate the broad “memory supercycle” narrative that has driven much of the sector’s re-rating. Companies with greater exposure to HBM and enterprise infrastructure may be better insulated than suppliers dependent on consumer storage demand, where buyers are generally more price-sensitive and inventory adjustments can be abrupt.
AI rack designs shift spending toward networking
Morgan Stanley also examined how AI data-center architecture can alter spending across the semiconductor and hardware supply chain. In Nvidia’s NVL72 single-rack configuration, the bank estimated storage represents around 29% of total system content, while networking accounts for 10%.
When the configuration expands to eight interconnected racks, storage’s share falls to 19% and networking rises to 20%, according to Morgan Stanley. The figures illustrate that scaling AI clusters does not simply multiply demand for every component evenly. Connecting larger systems requires more networking equipment, which can shift a greater portion of spending toward switches, cables and related infrastructure.
That allocation has implications beyond chipmakers. Operators of computing-intensive services, including blockchain infrastructure providers, may face higher costs for memory, storage and networking as data-center builders compete for equipment. The effect would vary widely by workload: validating transactions and running nodes generally require far less specialized memory than training frontier AI models, while services built around data availability, indexing or large-scale analytics can be more exposed to server hardware costs.
Morgan Stanley’s caution is therefore directed less at the immediate health of memory demand than at the expectations embedded in memory-stock prices. The sector can continue reporting strong earnings and rising contract prices while shares struggle if the market starts anticipating supply growth, slower price gains or a more selective AI infrastructure buildout.
As AI reshapes markets from chips to crypto, explore emerging opportunities with Toobit’s top 5 coins every trader should watch.
Disclaimer: The content on this page is provided for general informational purposes only and does not represent the views or financial advice of Toobit. We make no guarantees regarding the accuracy or completeness of this information and shall not be held liable for any errors, omissions, or outcomes resulting from its use. Investing in digital assets involves risk; users should independently evaluate their financial situation and the risks involved. For further details, please consult our Terms of Service and Risk Disclosure.
