The sharp sell-off in memory stocks may be over, while AI demand and capital returns open a tactical re-entry window

Overview

The report views the recent correction in memory stocks as a normal bump in a maturing cycle rather than the end of the AI memory demand thesis. The industry is expected to enter the late stage of the cycle in the fourth quarter of 2026, with price increases and the breadth of earnings estimate upgrades beginning to slow, but AI infrastructure buildout is still creating tight supply in DRAM, HBM, and enterprise storage. As it becomes more difficult to rely purely on price increases to drive earnings upside surprises, the investment narrative will gradually shift toward earnings stability from long-term agreements, free cash flow, capital discipline, and shareholder returns.

Core views

Over the long term, AI capex, the adoption of agentic AI, and the importance of memory to model performance constitute structural demand, with Samsung Electronics and SK hynix expected to deliver earnings growth of about 25% to 50% in 2027. In the near term, momentum-chasing trades may continue after the rapid deleveraging in July 2026, but third-quarter DRAM price performance is slightly below prior expectations, consumer electronics demand is weakening, and channel inventories are rising. The current valuation of about 3 times next-twelve-month earnings reflects strong expectations of an earnings decline; if AI demand extends the duration of earnings, multiple expansion and capital returns may become more important sources of returns than earnings estimate upgrades.

Analysis framework

The research combines second-quarter 2026 results, management guidance, supply-chain checks, and price and inventory data to evaluate three pillars: AI capex, long-term agreement re-rating, and the memory cycle. It also updates DRAM, NAND, and HBM price and shipment assumptions, and uses a residual income model to assess target prices for SK hynix and Samsung Electronics.

Methodology notes

• Industry cycle and structural demandThree-pillar framework of AI spending, long-term agreement re-rating, and the memory cycle

  Simultaneously examine structural AI demand, contracted revenue, and traditional cycle variables

  AI capex determines long-term demand strength, long-term agreements affect earnings visibility and capital returns, while prices, inventories, and supply determine the short-term cycle position. The three dimensions can move in different directions at the same time, so a strong AI cycle can coexist with a traditional memory downturn.

• Cycle assessmentSecond derivative of pricing framework

  Identify the late stage of the cycle by the peak in year-over-year price increases

  The report believes prices can still rise, but a slowdown in the pace of increases means the second derivative of cycle momentum has peaked. Operating leverage will find it harder to continue outpacing price growth, and the breadth of earnings estimate upgrades will also slow accordingly.

• Equity valuationResidual income valuation model

  Estimate target prices based on book value and excess returns

  This model is used to assess target prices for SK hynix and Samsung Electronics; the target prices correspond to about 2.6 times and 2.1 times expected 2027 price-to-book, respectively.

• Valuation paradigmComparison of normalized earnings and structural re-rating

  Compare traditional cycle-average valuation with an AI-driven longer earnings duration

  Traditional memory stocks are usually valued on normalized earnings, but if AI demand represents a structural change rather than an ordinary cycle, the historical mean-reversion framework may underestimate earnings duration, free cash flow, and the value of capital returns.

Asset mapping & comparison

Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).

• SK hynix (000660.KS)

  A core beneficiary of AI memory and DRAM supply bottlenecks, with the target price maintained at KRW 2,600,000.

  Strengths

  Leadership in HBM, higher DRAM exposure, and improved demand visibility from long-term agreements, with the target price implying 74% upside.

  Weaknesses

  The 2026 EPS upgrade mainly comes from a one-time asset disposal gain, while the late stage of the cycle will reduce the likelihood of sustained earnings upside surprises.

  Comparison

  Compared with memory module makers with higher consumer exposure, it benefits more directly from AI capex and DRAM tightness.

  Risks

  Uncertainty over HBM specifications and pricing negotiations, new supply in 2028, margin mean reversion, and a slowdown in AI capex.

• Samsung Electronics (005930.KS)

  The Top Pick highlighted in the report, with the target price maintained at KRW 381,000.

  Strengths

  Broad memory capacity and customer base, with plans to include 60% to 70% of capacity in long-term agreements and reduce volatility through price floors and prepayments.

  Weaknesses

  Weak consumer electronics and smartphone businesses, and memory cost pressure leading to a 10% cut in the 2026 EPS forecast.

  Comparison

  Its business is more diversified than SK hynix, and its long-term agreement coverage plan is clearer, but the drag from consumer businesses is also more pronounced.

  Risks

  Further deterioration in consumer demand, weaker-than-expected HBM competitiveness, memory supply expansion, and a pullback from high margins.

• DRAM and HBM

  The report’s most preferred memory segments and key bottlenecks in AI infrastructure buildout.

  Strengths

  Demand from hyperscale cloud providers continues to exceed supply, customers are actively signing long-term agreements, and HBM still has strong pricing demands.

  Weaknesses

  Third-quarter DRAM price increases are below previous expectations, and pricing momentum is expected to weaken from the fourth quarter.

  Comparison

  Compared with memory modules and consumer products, they more directly capture data center capex.

  Risks

  Final changes to HBM specifications, thermal management constraints, failed pricing negotiations, and new capacity coming online from late 2027 to 2028.

• NAND and legacy memory

  Overall positive, with greater preference for legacy products such as NAND SLC and DDR4 due to supply exits and shortages.

  Strengths

  Strong enterprise SSD demand, and severe SLC shortages in automotive, industrial, communications, and edge AI applications.

  Weaknesses

  Consumer NAND demand is softening, while MLC and QLC face price ceilings and customer resistance.

  Comparison

  Legacy-node SLC pricing performs better than MLC, and enterprise products outperform consumer products.

  Risks

  Capacity expansion by Chinese manufacturers such as YMTC, rising channel inventories, HDD substitution, and continued weakening in consumer demand.

• Memory module makers and distributors

  Relatively cautious; the report explicitly prefers upstream DRAM and legacy memory over module makers.

  Strengths

  If consumer demand recovers, they could benefit from inventory value and shipment improvement.

  Weaknesses

  About 90% of end-market exposure comes from consumer markets, inventories have risen to an average of 12.5 weeks, and procurement behavior is becoming conservative.

  Comparison

  Compared with upstream original memory manufacturers, they have lower direct exposure to AI demand and are more vulnerable to channel destocking.

  Risks

  Weak consumer electronics demand, continued inventory accumulation, procurement normalization, and high-priced memory suppressing end-device sales.

Key data

• 2027 cloud capex growthUp 29% year over yearThe previous forecast was 14%, raised by 15 percentage points after second-quarter 2026 results from U.S. hyperscale cloud providers.
• 2027 earnings growth of major Korean memory companiesAbout 25% to 50%Refers to expected earnings growth for Samsung Electronics and SK hynix, one of the report’s medium- to long-term bullish arguments.
• Third-quarter 2026 DRAM contract pricesUp about 15% quarter over quarterEarly transactions were slightly below the previous expectation of 20%, and incremental pricing momentum is expected to slow into the fourth quarter.
• Third-quarter 2026 NAND pricesUp about 20% quarter over quarterProduct divergence is significant, with more severe shortages in legacy-node SLC.
• PC DRAM contract pricesUp 15% to 20% quarter over quarter in the third quarter of 2026Significantly slower than the 45% to 50% increase in the second quarter, while negotiations between buyers and sellers remain tense.
• Cloud service provider DRAM long-term agreement coverage targetAt least 90% of demandIndustry checks indicate that cloud service providers remain willing to sign long-term agreements, and orders from new cloud service providers are also under negotiation.
• Samsung Electronics long-term agreement planPlans to cover 60% to 70% of capacityIt has reached agreements with five major global data center customers, with another five customers in final negotiations, and has set prepayments and minimum prices for mainstream products.
• SK hynix long-term agreement progressAbout 10 customers have completed negotiationsAgreements are usually about five years and include stronger two-way commitments, customized pricing, and deposit arrangements.
• Channel inventoryMemory module makers and distributors average 12.5 weeksAs of the end of the second quarter of 2026, below the previous cycle peak of 15 weeks, but procurement in non-AI consumer markets is becoming conservative.
• Cloud service provider inventoryCurrently 2.5 to 3 months, target 4 to 5 monthsCustomers still hope to build higher safety inventory, but procurement cadence is expected to normalize from the third quarter of 2026.
• SK hynix earnings forecast revisionEarnings per share for 2026 to 2028 adjusted by 13%, -2%, and 1%, respectivelyThe 2026 upgrade mainly comes from a one-time asset disposal gain of KRW 63.27 trillion; the target price is maintained at KRW 2,600,000, implying 74% upside.
• Samsung Electronics earnings forecast revisionEarnings per share for 2026 to 2028 adjusted by -10%, -2%, and 0%, respectivelyThe downgrade mainly reflects weaker Chinese consumer demand and weaker smartphone shipments and margins; the target price is maintained at KRW 381,000, implying 65% upside.

Impact & implications

The investment implication is shifting from chasing prices and earnings upgrades to assessing earnings durability and cash returns. Low valuations after the near-term correction support a rebound, but the late stage of the cycle is unfavorable for sustained earnings upside surprises. Over the medium term, investors should prioritize DRAM, HBM, and legacy memory segments with severe shortages that directly benefit from AI capex and supply bottlenecks, while reducing reliance on module makers with higher consumer exposure. If long-term agreements can stabilize prices and demand in a downcycle, and companies further disclose buybacks or other capital returns, industry valuations may break away from the traditional cycle-average framework.

Risks

• Capacity expansion by Chinese manufacturers such as CXMT and YMTC and entry into more non-Chinese customer supply chains could weaken long-term supply discipline and returns.
• New global wafer fab capacity coming online from late 2027 to 2028 could ease current shortages and push DRAM and NAND prices downward.
• DRAM industry gross margin approaching 90% is significantly above historical levels and could mean-revert due to capacity expansion, new entrants, or customers shifting to lower-cost solutions.
• AI data center buildout may eventually slow, and hyperscale cloud providers may reduce capex around 2028 due to financing constraints or normalization of the upgrade cycle.
• Persistent weakness in consumer electronics, PCs, and smartphone demand could further depress price increases and shipments.
• Rising interest rates, such as the U.S. 10-year Treasury yield rising to 4.7%, could weigh on valuations of high-growth memory stocks and increase market volatility.
• Long-term agreements have not yet gone through a full downcycle, and the effectiveness of their price floors, purchase commitments, and capital returns still needs to be proven.
• HBM’s final stacking specifications, thermal management, supply, and pricing remain under negotiation, and design changes could affect revenue and margin expectations.

What to watch

• Whether DRAM and NAND price increases continue to slow as expected in the fourth quarter of 2026.
• Further disclosures from Samsung Electronics, SK hynix, and other suppliers on buybacks, dividends, and capital return plans.
• Whether long-term agreements can stabilize revenue, margins, and free cash flow during a price downcycle.
• Whether 2027 capex from U.S. hyperscale cloud providers continues to exceed market consensus expectations.
• Whether HBM ultimately adopts 8-Hi or 12-Hi configurations, and the corresponding pricing and margin negotiation results.
• Whether inventories at module makers, distributors, and cloud service providers continue to rise, and whether procurement cadence normalizes.
• Financing, construction, and mass-production progress for new capacity from CXMT, YMTC, and Solidigm.
• Whether the divergence between AI enterprise SSD demand and consumer NAND demand widens.
• Whether earnings growth at Samsung Electronics and SK hynix from 2027 to 2028 can offset the decline in traditional memory prices.