Memory Market to Reach $1.7 Trillion by 2028, CPU Demand Surge and Supply Constraints Drive Profit Growth

Overview

J.P. Morgan’s report updates the long-term outlook for the global memory market. Key conclusions: (1) AI demand expands from GPUs to CPUs, with the CPU-GPU ratio narrowing, making AI servers and CPU DRAM demand new growth drivers; (2) Based on updated demand forecasts and supply chain validation, the report raises 26-28E memory TAM expectations by 37-53%, with 28E market size reaching $1.7 trillion; (3) High-end memory (HBM) faces long-term supply gaps, with ASP and capacity allocation revised upward, sustaining high industry margins at 76%+; (4) NAND growth is driven by enterprise SSD demand, with eSSD market doubling to 1,100EB in two years, value TAM nearing HBM scale; (5) Supply-side investment accelerates to $45 billion over three years (vs. $30 billion previously), but bottlenecks remain in EUV and capacity construction. Overall, the memory industry is in a "higher for longer" upcycle, with valuation frameworks transitioning to profit-driven models.

Core views

CPU demand emerges as a new growth engine. Following GPUs driving AI computing, CPUs (especially AI-specific CPUs like Vera) see explosive demand in task orchestration, state management, and API execution. NVIDIA projects $20 billion in Vera CPU sales by 2026, AMD raises server CPU TAM growth to 35%/year; Intel hints at CPU-GPU ratios balancing out from 1:8. The report estimates 26-28E AI CPU DRAM demand at 461M/2.3bn GB, while raising overall server DRAM demand by 5-22%. AI server DRAM (including AI headnodes and CPUs) rises from 14% to 30%+ of total market demand by 27-28E. Tight supply-demand dynamics, HBM shortages persist. The report raises 26-28E HBM TAM by 17-21%, driven by ASIC demand acceleration (ASIC bit mix rises from 33% in 26E to 39% in 27E) and conservative assumptions for HBM4E (long lifecycle). Despite some HBM capacity coming online in 2027, HBM3E/HBM4 demand growth (especially from ASICs like Google TPU and Amazon Trainium) remains strong, widening HBM S/D gaps. The report expects HBM3E ASP to hold steady in 26E, rise 19% Y/Y in 27E, with HBM4 and HBM4E premiums at 29% and 61%. DRAM wafer allocation to HBM rises from 24% in 26E to 31% in 28E, tightening conventional DRAM supply. NAND driven by enterprise SSDs, doubling in two years. High-end demand for fast, high-density SSDs grows, with single-server SSD capacity up 40% Y/Y. 26E eSSD market exceeds 500EB (43% of NAND bit demand), projected to reach 1,100EB in two years (CAGR +52%). With ASP premiums, eSSD value TAM could hit $30 billion, surpassing current HBM TAM. High-performance SLC adoption impacts NAND capacity allocation and cost curves. High industry margins, multi-year supply shortage consensus forms. Reflecting CoWoS upgrades, CPU memory growth, and consumer downturns, the report raises 26-28E memory TAM by 37-53%, with S/D gaps persisting into 2028. While tight supply could theoretically push prices up 220-250% Y/Y, long-term agreements (LTAs) with hyperscalers temper actual ASP increases, resulting in stable pricing. 27E memory TAM is estimated at $1.3 trillion (based on 3.6x P/S), offering 47% upside; using profit-based Market Cap/OP, upside reaches 89%. Capacity investment accelerates but bottlenecks remain. DRAM: Three-year capex totals $36.4 billion, CY28 WSPM reaches 2.8mn (+880k), with 60% allocated to HBM; Capex/Sales averages 13%. EUV access and infrastructure are key constraints. NAND: Three-year investment of $8.6 billion, CY28 WSPM up 880k to 1.44mn, focused on process migration and post-2H28 capacity. Compared to DRAM, NAND investment is lower but shows signs of 2029 greenfield projects. Chinese competition continues but high-end constraints remain. DRAM volume share rises from 6% in 2025 to 8-11% in 26-28E, NAND from 12% to 12-16%; capacity share reaches 18-19%. However, product mix limits value share (DRAM/NAND value share only 10%/12% by 28E). Key focus: (1) Wafer allocation between conventional DRAM and HBM (higher HBM share tightens conventional DRAM S/D); (2) Equipment export control updates (e.g., MATCH Act).

Analysis framework

The report employs a multi-dimensional supply-demand model framework. On the demand side, it builds bottom-up forecasts by chip and application (cloud provider projects, niche ASICs), covering DRAM headnodes, standalone CPUs, and enterprise SSDs, based on GPU/ASIC units and memory configurations (per-chip capacity determined by channels and DIMM setups). On the supply side, it estimates WSPM growth and capacity allocation trends using fab plans, process roadmaps, and investment cycles. S/D gaps quantify ASP and margin trends, incorporating CSP hardware investment and LTA contract share to assess sustainability. Valuation methods include traditional P/S and profit-based Market Cap/OP to reflect memory firms’ transition from cycle troughs to high-margin phases.

Methodology notes

• Industry/Industrial Analysis FrameworkSupply-demand framework

  Supply-demand imbalances directly drive pricing and capacity allocation decisions. Prolonged DRAM/NAND shortages create positive feedback loops of ASP support and investment incentives.

  The report uses DRAM S/D gaps (bit demand vs. capacity shipments) to forecast ASP ranges and allocation shifts. Supply discipline (e.g., higher LTA share) and demand certainty strengthen pricing power, enabling stable ASPs even during shortages.

• Industry/Industrial Analysis FrameworkVolume-Price Breakdown

  Memory market growth is driven by bit demand and ASP changes. Shifts in GPU-CPU ratios and new products (HBM, eSSD) lift ASPs while AI workloads boost bit demand.

  The report details contributions from AI vs. traditional servers, HBM vs. conventional DRAM, and eSSD vs. QLC SSD. AI server DRAM rising from 14% to 30%+ lifts overall ASPs; HBM’s 85% CAGR, though small, significantly impacts profits. This mix is key to assessing 28E TAM and margins.

• Company Fundamentals & Financial FrameworkFree cash flow analysis

  Memory’s capital intensity makes Capex/Sales a cycle indicator. High margins encourage upfront investment to capture growth.

  The report tracks three-year DRAM/NAND Capex ($36.4bn and $8.6bn) and Capex/Sales (13% and lower). With shortages and 80%+ margins, firms invest aggressively, constrained by EUV and fab lead times. Capex timing affects future S/D balance.

• Valuation methodsPS valuation

  PS multiples better reflect long-term market share during cycle transitions, as PE volatility spikes.

  The report uses 3.6x P/S (below historical median) for 27E TAM. AI demand and margin stability justify re-rating; Market Cap/OP offers higher upside (89% vs. 47%) as profits outpace revenue.

• Valuation methodsROIC–WACC spread

  Memory firms’ ROIC exceeding WACC creates value, especially during high-margin shortages.

  Implicitly, high investments amid shortages yield strong returns (77% margins) well above capital costs, justifying capex and stock re-rating.

• Cycle & Sentiment FrameworkInflection Point Analysis

  Memory cycles shift from oversupply to tight supply, with inventory, S/D gaps, and ASPs as key indicators.

  The report positions 26-28E as a "higher for longer" upcycle, driven by GPU-to-CPU demand shifts, AI server adoption, and delayed supply (EUV bottlenecks). This cycle’s strength and duration exceed past ones.

• Event Gaming & Behavioral FinanceExpectation Gaps/Management

  Detailed models reveal underestimated shortages. LTAs manage expectations, avoiding volatility.

  While nominal S/D gaps persist in 2028, LTAs with hyperscalers smooth ASP hikes, diverging from market expectations of sharp spikes and reducing bubble risks.

• Industry/Industrial Analysis FrameworkCapacity Cycle/Equipment Cycle

  DRAM/NAND investment lags (2-3 years) prolong shortages as decisions trail demand by 1-2 years.

  EUV and fab bottlenecks mean $45bn investments (vs. $30bn) won’t fully deploy until 2028, sustaining S/D gaps. Leaders (e.g., SK Hynix, Samsung in HBM4/4E) widen process advantages.

Asset mapping & comparison

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

• SK hynix(000660.KS)

  HBM technology leader with ASP premiums. 1cnm mono-die process leads vs. 1bnm peers; Samsung’s 12Hi HBM4E samples trail. HBM4 holds high-teens% share in NVDA demand.

  Strengths

  HBM4/4E tech edge (1cnm, SF 4nm base die), back-end yield improvements, NVDA wallet share upside. Stable ASIC share.

  Weaknesses

  Smaller conventional DRAM share; HBM reliance on NVDA; new fab delays risk.

  Comparison

  vs Micron: SKH leads in HBM but MU has more DRAM capacity; vs Samsung: SKH’s 1cnm HBM4 edges Samsung’s 1bnm.

  Risks

  NVDA Rubin GPU demand shortfall pressures HBM4/4E; Chinese HBM breakthroughs threaten long-term share.

• Samsung Electronics(005930.KS)

  Integrated memory leader in DRAM, NAND, HBM. 28E benefits from market expansion and capex.

  Strengths

  Largest scale, NAND share leader. HBM4E 12Hi samples shipped, covering 8Hi-16Hi. Full product synergy.

  Weaknesses

  HBM4 process (1bnm) lags SKH, ASP pressure; NAND capex priority lower vs. DRAM.

  Comparison

  vs SKH: larger but HBM weaker; vs MU: stronger NAND but comparable HBM progress.

  Risks

  SKH HBM share gains pressure margins; Chinese competition weighs on NAND ASP.

• Micron Technology(MU)

  US leader in DRAM, HBM, NAND. ID1 fab and Singapore back-end expansions support capacity.

  Strengths

  Aggressive HBM capacity (ID1, Singapore), ASIC share rising by 27E. Stable conventional DRAM costs.

  Weaknesses

  HBM4/4E slightly behind; new capacity tight until 27E.

  Comparison

  vs SKH/Samsung: HBM lags but capex larger; middle-ground competitiveness.

  Risks

  New capacity delays/cost overruns; US export limits China sales.

• KIOXIA Holdings(285A.T)

  Japanese NAND leader, direct eSSD beneficiary. Two-year eSSD CAGR +52%, premium SSD TAM over $30bn.

  Strengths

  NAND process lead (3D layers). eSSD growth with strong client stickiness.

  Weaknesses

  No DRAM, lacks full solutions; smaller vs. Korean giants.

  Comparison

  Competitive in NAND eSSD but lacks broad memory coverage.

  Risks

  eSSD slowdown; process lead erosion costs competitiveness.

• Winbond(2344.TW)

  Taiwanese DRAM niche player (consumer/industrial). Limited AI upside despite broader market growth.

  Strengths

  Low cost, niche client loyalty.

  Weaknesses

  Small scale, tech lag; no HBM progress.

  Comparison

  vs majors: downmarket focus, minimal high-end presence.

  Risks

  AI/server DRAM growth squeezes traditional market share.

• Tokyo Electron(8035.T)

  Semiconductor equipment leader (SPE), benefits from $45bn memory capex. Key EUV/etching supplier.

  Strengths

  EUV/etching bottlenecks ease with equipment orders.

  Weaknesses

  Cyclical, reliant on fab funding/construction timelines.

  Comparison

  Indirect beneficiary vs. direct memory players.

  Risks

  Capex delays/insourcing pressure equipment demand.

• Silicon Motion(SIMO)

  SSD controller vendor, leveraged to eSSD (CAGR +52%). High-end SSDs need advanced controllers.

  Strengths

  SSD controller share lead, eSSD growth lifts demand.

  Weaknesses

  Narrow product focus; distant from end clients.

  Comparison

  Mid-chain eSSD beneficiary, less direct than memory firms.

  Risks

  eSSD slowdown/competition pressures growth.

Key data

• 28E Memory TAM$1.7 trillionvs. prior $1.1 trillion, up 53%; driven by CPU, HBM, eSSD demand revisions
• CPU-GPU Ratio Evolution2023 5.4:1 → 2025 3.2:1 → 2028E 2.4:1ASIC racks higher (7.5:1→2.6:1); NVDA racks now 2:1; CPU DRAM is 19-24% of total demand (27-28E)
• HBM TAM Revision17-21% (26-28E)ASIC demand acceleration and HBM4E lifecycle assumptions are key drivers
• HBM ASP Growth26E mid-teens%, 27E +32% Y/YHBM3E up 19% in 27E, HBM4 +15%, HBM4E premium at 61% over HBM3E
• HBM DRAM Wafer Allocation24% (26E) → 31% (28E)HBM bit demand CAGR +85%, squeezing conventional DRAM supply
• DRAM CY28 WSPM2.8mn (+880k from CY25)Three-year Capex $36.4bn, 60% to HBM; Capex/Sales avg. 13%
• eSSD Market Size26E 500EB+ → 28E 1,100EB+ (CAGR +52%)eSSD value TAM over $30 billion, nearing HBM; SLC adoption rising
• Memory Industry Margins26-28E avg. 76-77%Above historical levels; LTAs and discipline sustain profits
• Three-Year Total Investment$45 billion (vs. $30bn prior)DRAM $36.4bn, NAND $8.6bn; reflects shortage expectations and high margins
• China DRAM/NAND Capacity Share26-28E 18%/19% (vs. 6%/12% in 2025)But value share lags (10%/12%) due to product mix constraints
• AI CPU DRAM Demand26E 461M GB (1% of total) → 27E 2.3bn GB (3% of total)Vera CPU projected at 600K/3M units in 26-27E, with 768GB-1.5TB per unit
• 27E TAM Upside (P/S)47% (3.6x P/S)Based on $1.3 trillion TAM estimate
• 28E OP-Based Upside89% (Market Cap/OP)Profit growth outpaces revenue, reflecting cycle leverage

Impact & implications

The report outlines multi-dimensional impacts from this valuation update. Demand: AI CPU demand extends the "higher for longer" cycle beyond GPUs, broadening memory applications. Future Physical AI/World AI could further lift demand. Supply: $45bn investments face EUV/fab bottlenecks, prolonging S/D gaps. HBM capacity favors leaders (SKH, SEC, MU), widening gaps. Pricing: LTAs stabilize ASPs, avoiding volatility. Profits: 76-77% margins attract investment and fund R&D (HBM4E, new processes). Stocks: P/S or profit-based multiples suggest re-rating potential (47-89%). Regional: China’s DRAM/NAND volume grows but value share lags due to product mix; watch export controls and local breakthroughs.

Risks

• AI demand growth shortfalls or bubbles pressure TAM/ASP
• Equipment export controls (e.g., MATCH Act) tighten, China struggles to fill gaps, altering competitive dynamics
• EUV bottlenecks delay capacity, prolonging shortages or triggering overshoots
• New architectures (3D stacking, chiplets) disrupt DRAM/NAND roadmaps
• CSP capex slowdowns on ROI concerns hit cloud memory demand
• HBM concentration with NVDA exposes to single-client volatility
• LTAs stabilize ASPs but reduce flexibility to demand shocks

What to watch

• NVDA Rubin GPU/AMD MI ASIC 2026 shipments and memory configs, driving HBM demand
• CSP in-house CPUs (Graviton/Cobalt/Axion) 2027 volume and DDR/LPDDR5 needs
• DRAM big three (SKH/Samsung/MU) 26-27 capex deployment, especially HBM/EUV
• DRAM wafer allocation to HBM (24% in 26E → 31% in 28E), reflecting priorities
• High-end NAND (SLC/TLC) mix and eSSD ASP trends, sustaining NAND margins
• LTA contract share growth, indicating pricing discipline
• Chinese memory (YMTC DRAM, CXMT NAND) process/cost progress, reshaping markets
• New AI-driven HBM capacity needs (e.g., HBM5) rekindling S/D imbalances