摩根士丹利依然看好海力士和三星，Q4存储行情将变化

TL;DR

• Morgan Stanley maintains "Overweight" ratings on SK Hynix and Samsung Electronics, with target prices of KRW 2.6 million and KRW 381,000, respectively.
• Based on the share prices used in the report, the two target prices imply potential upside of approximately 74% and 65%, respectively, which does not guarantee that share prices will necessarily rise.
• The report raises its 2027 cloud capex growth forecast from 14% to 29%, with AI compute demand remaining the primary structural support for the memory cycle.
• DRAM and NAND contract prices are still rising, but the sequential growth rate has moderated from peak levels; Morgan Stanley expects the industry to transition to the late-cycle phase in Q4 2026.
• Long-term supply agreements improve order visibility but cannot fully eliminate cyclical fluctuations in pricing, demand, and new supply additions.

Morgan Stanley's Asia tech report "Memory – A Small Wrinkle," released on August 6, continues to favor Korea's memory leaders, maintaining its target prices for SK Hynix and Samsung Electronics. Specifically, the target price for SK Hynix is KRW 2.6 million, while the target price for Samsung Electronics' common stock is KRW 381,000. Based on the baseline share prices used in the report, these correspond to potential upside of roughly 74% and 65%, respectively.

Such high target prices do not imply that Morgan Stanley believes memory prices will continue rising at the previous pace. On the contrary, the report has already observed narrowing price increases, rising channel inventory, and new capacity gradually entering the market, and it expects the memory industry to shift into the late-cycle phase in Q4 2026.

Morgan Stanley remains bullish, primarily betting on three themes: upward revisions to AI capex, improved earnings visibility from long-term supply agreements, and the recent valuation pullback in memory stocks already reflecting some peak-cycle concerns.

Earnings forecasts also reflect this caution. Morgan Stanley raised its 2026 expected EPS for SK Hynix by 13%, mainly due to the inclusion of a KRW 63.27 trillion one-time investment gain in Q2; at the same time, it lowered its 2026 operating profit forecast for the company by 7%. Samsung Electronics' 2026 expected EPS was cut by 10%, mainly reflecting weakness in consumer businesses such as smartphones.

Earnings forecast adjustments for 2027-2028 are relatively smaller for both companies. Based on the revised forecasts in the report, SK Hynix's 2027 EPS is expected to grow approximately 25% year-over-year, while Samsung Electronics is expected to grow around 49%, largely corresponding to the report's cited range of 25% to 50%.

Prices Still Rising, but the Memory Cycle Is Slowing

The most direct driver of the recent rally in memory stocks has been product price increases. AI server demand has boosted demand for HBM, server DRAM, and enterprise SSDs, while capacity expansion has struggled to keep pace, driving DRAM and NAND prices significantly higher.

Morgan Stanley cites TrendForce data showing that sequential contract price increases for overall DRAM and NAND peaked at 96% and 88%, respectively, in Q1 2026. In Q2, prices are still expected to rise 61% and 58%, respectively, but Q3 growth is expected to moderate to 16% and 13%, before further slowing to 6% and 3% in Q4.

The report's latest channel checks indicate early Q3 DRAM contract price settlements are up approximately 15% sequentially, slightly below the prior expectation of 20%; NAND contract prices are up around 20%. PC DRAM contract prices in Q3 are expected to rise 15% to 20% sequentially, a clear moderation from the 45% to 50% increase seen in Q2.

Therefore, a more accurate description is not that "memory prices have peaked," but rather that prices are still climbing while the rate of increase has already come down. Morgan Stanley expects the industry to gradually move into the late-cycle phase by Q4 2026, when the operating leverage from pricing weakens and the difficulty of delivering further earnings surprises increases.

It is worth noting that the narrowing of price increases does not necessarily stem entirely from supply improvements. The report points out that some consumer DRAM price increases have slowed because buyers are approaching the ceiling of their cost tolerance; meanwhile, AI-related customer demand remains robust, and suppliers are shifting some consumer-grade capacity toward enterprise SSDs and other products.

Sequential change in memory contract prices. The sequential growth rates of DRAM and NAND contract prices are moderating quarter by quarter from peak levels.

AI Compute Still in Short Supply, Cloud Capex Revised Higher

The first pillar supporting Morgan Stanley's bullish stance is that AI data center compute demand remains unmet in full.

The report cites earnings calls and management commentary from four major U.S. cloud providers: Alphabet and Microsoft both indicate cloud demand still exceeds internal available capacity; Amazon expects capacity to remain insufficient to meet demand in 2026, with a significant portion of 2027 capacity already pre-booked; and Meta expects industry compute supply to remain tight for the foreseeable future.

While this information does not prove that all AI investments will ultimately generate expected returns, it does suggest that major cloud providers have not yet meaningfully scaled back infrastructure buildout. As a result, Morgan Stanley's cloud capex tracking model has raised its 2027 year-over-year growth forecast from 14% a month ago to 29%.

This is critical for memory makers. AI servers require not only GPUs but also HBM, server DRAM, and enterprise SSDs. As long as cloud providers continue expanding data centers, memory demand will not be determined solely by traditional consumer electronics cycles such as PCs and smartphones.

But "structural AI demand" and "cyclical price corrections" can coexist. Morgan Stanley's core view is precisely that AI demand may extend memory companies' earnings cycle, yet it will not make memory pricing and inventory cycles disappear entirely.

Cloud capex growth forecast. The chart shows the 2027 cloud capex growth forecast revised up from 14% to 29%.

LTAs Improve Order Visibility but Cannot Lock in Future Profits

The second pillar comes from long-term supply agreements, or LTAs (Long-Term Agreements). Compared with traditional quarterly procurement, multi-year contracts allow suppliers to confirm a portion of demand in advance and plan capacity and capital expenditure accordingly.

SK Hynix's official Q2 earnings release states that the company has completed LTA negotiations with approximately 10 key customers and is continuing discussions with others. Morgan Stanley further relays from the company's earnings call that these agreements typically last around five years, though terms and pricing mechanisms vary by customer and product; some agreements include deposits, and prices are subject to adjustment with market fluctuations.

SK Hynix has not disclosed the specific proportion of capacity or revenue covered by LTAs, only stating that it intends to keep agreements at an "appropriate level" — enhancing downside protection while reserving capacity for incremental demand. Therefore, it would be inaccurate to characterize its LTAs as having locked in the majority of future revenue.

Samsung Electronics' specific terms are primarily based on Morgan Stanley's summary of its Q2 earnings call. According to the report, Samsung plans to include 60% to 70% of its capacity under long-term agreements, has already reached agreements with five global large-scale data center customers, and has five more in final-stage negotiations. The agreements adopt a rolling five-year structure, require customer prepayments, and set price floors for certain mainstream products.

Potential customers listed in the report's table — including AWS, Microsoft, Google, Meta, and Oracle — are explicitly noted as based on media reports and should not be described as officially confirmed counterparties by Samsung.

The primary significance of LTAs is to enhance demand visibility, support capacity investment, and reduce some price volatility, rather than to fully lock in future profits. Different contracts vary in the capacity covered, duration, price adjustment mechanisms, and default protections; if AI buildout slows, product specifications change, or market prices undergo major adjustments, the protection offered by these agreements still has limits.

LTA comparison across major memory makers. The chart compares long-term agreement coverage, duration, and pricing arrangements across Samsung, SK Hynix, Micron, SanDisk, and Kioxia.

High Target Prices Bet on a Prolonged Cycle; New Supply Determines Upside

The market has already begun pricing in a slowdown in memory earnings growth ahead of time. Morgan Stanley notes that the 12-month forward P/E for DRAM-related stocks typically leads the next 12 months of EPS by roughly two months, and recent valuations have already pulled back notably, reflecting investors pricing in decelerating earnings growth.

This also explains why memory companies' current earnings are strong, yet their share prices may not necessarily continue to rise in tandem. Investors' focus is shifting from existing profits in 2026 to sustainability in 2027-2028: whether AI capex can continue growing, whether LTAs can withstand a downturn cycle, and whether prices can hold after new capacity comes online.

High profits themselves also attract supply. Morgan Stanley points out that current DRAM gross margins are close to 90%, which is unusually high from a historical perspective. If high returns prompt leading manufacturers to accelerate capacity expansion or attract new entrants, existing margins could revert toward long-term averages.

Here it is important to distinguish between two metrics: the near-90% figure refers to DRAM gross margins as discussed in the report, while SK Hynix's Q2 company-level operating margin of 76% is a different metric that cannot be directly compared. SK Hynix officially disclosed Q2 2026 revenue of KRW 79.3187 trillion and operating profit of KRW 60.5426 trillion, representing an operating margin of 76%.

Chinese manufacturers' capacity expansion is also listed as a long-term risk. The report states that new supply from ChangXin Memory Technologies (CXMT) and YMTC could alleviate tightness in certain product segments; notably, CXMT's roadmap includes plans to supply HBM in the Chinese market as early as 2027. This is Morgan Stanley's representation of the company's roadmap and does not mean that Chinese HBM has already achieved volume production or can immediately replace high-end Korean products.

Additionally, the report expects major manufacturers to bring new capacity into production between 2027 and 2028. If AI demand continues growing at a rapid pace, this new supply may not immediately create oversupply; however, if cloud capex growth decelerates, supply release could accelerate price declines.

Therefore, the key assumption behind the KRW 2.6 million target price is not that memory prices will rise indefinitely, but that AI demand can extend the earnings cycle, LTAs can reduce volatility, and the recent valuation pullback has already priced in a considerable portion of cyclical risk. What truly needs to be validated next is whether these long-term agreements can survive a price downturn, and whether memory companies can find new EPS growth drivers beyond 2028.

DRAM valuation versus earnings expectations. The chart reflects that valuations typically lead changes in earnings expectations and should not be interpreted as implying a stable, mechanical causal relationship between the two.