The Correction in Memory Stocks May Be Over, but the Supply Crisis Will Continue

Overview

The report updates market size, supply-demand, capacity, and pricing forecasts for global DRAM, NAND, and HBM. Based on assumptions around CoWoS, next-generation PC and server configurations, AI CPU ramp-up, and content optimization, JPMorgan raises its 2026E-2028E memory TAM forecasts and judges that new wafer capacity will be difficult to bring online in time to close the demand gap. Although memory stocks have pulled back quarter-to-date in 3Q26 due to earnings expectations misses, the pace of CSP capital expenditure, and optimization of per-unit memory capacity, the report believes this high-cycle upturn can continue until end-2028.

Core views

Core conclusions include: first, AI memory demand is expanding from GPUs to CPUs, long-context inference, KV cache, and agent workloads, and investors may be underestimating the actual impact of demand diffusion; second, SOCAMM and HBM content optimization are response mechanisms in a shortage environment, not the start of a downcycle; third, HBM wafer efficiency losses, fab construction cycles, and physical space constraints make it difficult for new capital expenditure to quickly translate into effective supply; fourth, earnings visibility from long-term agreements and special dividends and buybacks over the next two to three years are expected to drive a rerating of memory stocks.

Analysis framework

The report combines a global memory TAM model, DRAM and NAND supply-demand models, capacity forecasts at the vendor and fab levels, capital expenditure conversion efficiency, HBM product iteration, CSP capital expenditure and cloud revenue trends, long-term agreement terms, as well as market capitalization, operating profit, and valuation metrics for cross-analysis, and uses scenario analysis to estimate the additional capacity and capital expenditure required to achieve supply-demand balance in 2028.

Methodology notes

• Industry Supply-DemandMemory Supply-Demand Balance Model

  Measures the degree of shortage based on DRAM and NAND demand, supply, consumption, and the difference in year-on-year bit growth.

  The model shows that although the growth-rate gap has narrowed to single digits due to content optimization, the absolute supply-demand gap remains large and may continue to widen into 2028E.

• Market SizeServiceable Market Size Analysis

  Forecasts global memory TAM based on price, shipment volume, and different memory product mixes.

  After incorporating CoWoS, new PC and server configurations, and AI CPU demand, 2026E-2028E memory TAM is raised by 4%-8% versus the previous model.

• Capacity AnalysisBottom-Up Capacity Model

  Forecasts future supply by vendor, fab, production start time, and monthly capacity.

  The model considers that a new fab requires about 2 to 2.5 years from groundbreaking to mass production, HBM wafer efficiency losses, and technology migration, and concludes that publicly announced expansion projects are still insufficient to meet demand.

• Capital EfficiencyBit Output per Unit of Incremental Capital Expenditure Analysis

  Compares the bit output that each unit of new capital expenditure can generate for DRAM and NAND.

  DRAM bit output per unit of capital expenditure is flattening, with new investment more focused on alleviating bottlenecks; NAND output efficiency is improving due to layer-count migration, so its potential oversupply risk is relatively higher.

• Valuation and CatalystsEarnings Visibility and Shareholder Return Analysis

  Assesses rerating potential by combining long-term agreements, margins, capital allocation, special dividends, and buybacks.

  Long-term agreements help smooth pricing and earnings curves, while clear shareholder return plans may prompt investors to place greater value on the sustainability of memory earnings.

Asset mapping & comparison

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

• Global Memory Stocks

  Sector-wide Overweight

  Strengths

  AI demand diffusion, constrained supply expansion, prices staying high, and long-term agreements improving earnings visibility.

  Weaknesses

  After a long period of outperformance, valuations are more sensitive to earnings delivery and the pace of capital expenditure.

  Comparison

  The report believes the roughly 25% correction in 3Q26 is more of an expectations reset than the end of the memory upcycle.

  Risks

  AI demand slowdown, CSP capital expenditure cuts, content optimization exceeding expectations, or earlier-than-expected supply release.

• SEC

  Near-Term Preferred Name in Korea Memory Sector

  Strengths

  A clearer path for new capacity, and shareholder return plans may bring special dividends and buybacks.

  Weaknesses

  Large-scale capacity expansion requires high capital investment, and there is uncertainty around execution and production ramp timing.

  Comparison

  The report prefers SEC over SKH in the near term.

  Risks

  New capacity ramping below expectations, rising capital expenditure, and memory prices weaker than forecast.

• SKH

  Maintain Positive View

  Strengths

  HBM capacity layout, M15X and Yongin cluster expansion, and strong AI memory exposure.

  Weaknesses

  HBM wafer efficiency losses are relatively high, and capacity expansion realization requires a longer time.

  Comparison

  Ranks below SEC in the near term, but remains part of the global overweight portfolio.

  Risks

  Delays in HBM product iteration, customer specification adjustments, and execution risk for new capacity.

• Micron

  Favored Name in the U.S. Memory Sector

  Strengths

  Benefits from global DRAM and HBM shortages and rising prices.

  Weaknesses

  Compared with SEC and SKH, new capacity additions through the first half of 2027 are relatively limited.

  Comparison

  Benefits from the industry upcycle together with Asian peers, but near-term incremental supply is smaller.

  Risks

  Delays in new fab production, technology migration falling short of expectations, and changes in industry demand.

• Kioxia

  Favored Name in Japan NAND Sector

  Strengths

  Benefits from demand growth in NAND driven by enterprise SSDs, high-bandwidth flash, and AI inference.

  Weaknesses

  NAND bit output per unit of capital expenditure is improving quickly, and supply elasticity is relatively higher.

  Comparison

  NAND's potential oversupply risk is higher than DRAM's, but the base case remains continued shortage.

  Risks

  Technology migration leading to higher-than-expected supply, slow maturation of the software ecosystem, and end-demand volatility.

• NYT/Winbond

  Favored Names in Taiwan Memory Sector

  Strengths

  Can benefit from traditional DRAM demand, AI inference memory diffusion, and rising industry prices.

  Weaknesses

  Scale and advanced product mix are limited compared with global leaders.

  Comparison

  Included in the global memory overweight universe, but the benefit path is more tilted toward traditional and niche memory.

  Risks

  Competition from Chinese manufacturers, increased mature-node supply, and product mix improvement falling short of expectations.

Key data

• Memory TAM Forecast Revision2026E-2028E raised by 4%-8%Compared with the May 2026 model, both price and shipment volume contribute.
• Next-Year Bid Supply Fulfillment Rate70%-80%It was about 90% during the pandemic peak, indicating supply remains tight.
• Next-Year HBM Price ForecastUp 42% year-on-yearAlso considers an extended 8-Hi cycle, a slower 12-Hi ramp, and delayed 16-Hi adoption.
• AI CPU Unit Market Size GrowthExpected CAGR of 155%More AI CPU shipments are expected to offset the impact of per-unit memory capacity optimization.
• Major CSP Cloud Revenue2Q26 combined US$106bn, up 43% year-on-yearCovers AWS, MSFT, and GCP, with growth continuing to accelerate from 35% in the previous quarter.
• CSP Capital Expenditure Expectation RevisionRaised 6% for this year and 17% for next yearRepresents changes in market consensus expectations after three major CSPs and Meta reported results.
• Memory as a Share of CSP Capital Expenditure31% in 2026E and 49% in 2027EBelow 10% before the AI era; a level close to 50% may cause investors to worry about sustainability.
• Global DRAM Monthly Capacity1.90 million wafers at end-2025 and 2.85 million wafers at end-2028Even with expansion, it is expected to remain below demand; HBM's share is expected to reach about 32% by end-2028.
• Additional Capacity Needed for 2028E Supply-Demand BalanceDRAM about 300,000 wafers per month, NAND about 45,000 wafers per monthThe report believes the probability of securing this capacity within the next year is low.
• Supply Required for 2028E BalanceDRAM about 5.5EB, NAND about 76EBThe corresponding capital expenditure requirements are about US$58bn and US$7bn, respectively.
• DRAM Average Selling Price ForecastUp 235% in 2026E, up 29% in 2027E, and up 7% in 2028EThe rate of increase slows, but absolute prices are still expected to continue rising.
• Long-Term Agreement TermsPrepayments account for 20%-25%, covered volume accounts for 50%-70%AI and server memory typically command a 30%-40% price premium over consumer electronics applications.
• Recent Correction in Memory StocksDown about 25% quarter-to-date in 3Q26Mainly due to short-term earnings expectations misses, CSP spending slower than expected, and memory content optimization.
• Potential Shareholder Return YieldSEC and SKH may reach 16%-20% over the next two yearsSpecial dividends and buybacks may become catalysts for valuation expansion.

Impact & implications

Structural shortages mean the cycle of memory price increases, margin improvement, and earnings upgrades may be more durable than in the past, and higher capital expenditure should be viewed more as relieving bottlenecks rather than a direct signal of uncontrolled supply. For equities, the recent correction has improved the risk-reward profile, but valuation re-expansion still requires continued CSP capital expenditure growth, no further large reductions in memory specifications, and implementation of shareholder return plans. NAND benefits from enterprise SSDs, high-bandwidth flash, and near-GPU storage demand, but its bit output per unit of capital expenditure is improving more quickly, making its potential oversupply risk higher than DRAM.

Risks

• CSP AI capital expenditure or cloud revenue growth may fall below expectations, weakening memory demand.
• SOCAMM, HBM, and server memory specifications may be further reduced, causing the drag from content optimization on demand to exceed the base case.
• New fabs, technology migration, or capacity expansion by Chinese manufacturers may drive supply growth above expectations, especially in NAND.
• HBM 12-Hi ramp, 16-Hi adoption, or related packaging ecosystem progress may fall short of expectations.
• Memory prices may peak earlier from a high base, causing the scale of earnings upgrades to narrow.
• Memory's share of CSP capital expenditure approaching or exceeding 50% may trigger customer budget constraints and investor concerns about sustainability.
• Long-term agreement pricing transparency may be insufficient, and actual price protection and volume commitments may be weaker than expected.
• Shareholder return plans may be delayed, or capital expenditure may squeeze free cash flow, limiting valuation rerating.

What to watch

• AI capital expenditure, cloud revenue, and server deployment updates from major CSPs.
• Capacity and layer-count specification changes for SOCAMM, Rubin, Rubin Ultra, and HBM4E.
• Quarterly DRAM and NAND supply-demand gaps, spot prices, and average selling price trends.
• New fab production starts and monthly capacity ramps by SKH, SEC, Micron, and Chinese memory manufacturers.
• Product cycles, yields, and customer qualifications for HBM 8-Hi, 12-Hi, and 16-Hi.
• Prepayment ratios, covered volumes, price adjustment mechanisms, and renewals of long-term agreements.
• Special dividends, buybacks, and medium- to long-term capital allocation policies of SEC and SKH.
• Commercialization progress of high-bandwidth flash, CXL memory pooling, enterprise SSDs, and near-GPU storage ecosystems.