CPU Demand Drives Storage TAM Expansion to $1.7 Trillion; Supply Gap Supports High-Margin Cycle

Overview

This report focuses on the profound impact of AI computing architecture evolution on the global storage market. The traditional GPU-driven AI ecosystem is expanding to the CPU layer, which undertakes key functions such as orchestration, state management, and API execution. This shift drives server-grade memory demand from a GPU-centric model in 2023 to a diversified layout by 2028, with a significantly increased proportion of CPU demand. Simultaneously, bottlenecks on the supply side (especially capacity constraints in HBM and NAND) create a mismatch with demand expansion, creating a long-cycle, high-margin prosperity environment. The report adjusts the Total Addressable Market (TAM) expectations for storage from 2026 to 2028, expanding the three-year growth amplitude to 37-53%, ultimately reaching a scale of $1.7 trillion.

Core views

CPU demand becomes the new growth engine. NVDA's 2026 Vera CPU sales are expected to reach $20 billion; AMD has raised its server CPU TAM growth expectation to a 35%+ annual growth rate (to $12 billion by 2030); Intel explicitly targets a CPU-to-GPU ratio shift from 1:8 to 1:4 or even parity. Based on supply chain surveys, J.P. Morgan estimates 600K Vera CPU units in 2026 and 3M in 2027, with corresponding AI CPU memory demand rising from 461Mn GB to 2.3Bn GB, increasing the share of total DRAM demand from a historical 14% to 30%+ in 2027-28. Server DRAM density and architectural changes drive up memory content per unit. The GPU-to-CPU ratio compresses from 5.4:1 in 2023 to 3.2:1 in 2025, and is expected to reach 2.4:1 by 2028, with even higher CPU density on ASIC servers. Average memory for dual-socket servers increases from 1.5TB in 2024 to higher levels by 2028. Meanwhile, although cloud providers' self-developed CPUs (AWS Graviton 5, MSFT Cobalt 100, GOOG Axion) have a memory configuration of 1 DPC per server (vs. conventional 1-2 DPC), the increase in CPU sockets per server fully compensates for this. HBM remains tight, with ASP and supply gaps reinforcing each other. HBM TAM is上调 (raised) by 17-21% (2026-28E), driven by accelerating ASIC demand (bit mix rising from 33% in 2026 to 39% in 2027), pushing up the blended ASP of HBM3E/HBM4/HBM4E. HBM3E ASP is expected to remain flat in 2026 and rise 19% in 2027; blended HBM ASP is projected to rise mid-teens% in 2026 and record a historic 32% YoY growth in 2027. The premium of HBM4E over HBM3E is expected to rise to 61%. DRAM capacity allocation tilts towards HBM, rising from 24% in 2026 to 31% in 2028, further squeezing conventional DRAM supply. Enterprise SSD becomes a new growth point for NAND. The high-speed SSD (HBF, SLC, etc.) market is expanding rapidly, with eSSD capacity exceeding 500EB in 2026 (accounting for 43% of NAND bit demand), expected to surpass 1,100EB by 2028 with a CAGR of 52%. ASP premiums are significant; within two years, the eSSD value TAM will exceed $30 billion, surpassing the scale of HBM. Although the rising SLC mix is unfavorable for NAND bit/wafer output, the drive in total market demand offsets this negative factor. Supply side accelerates investment to address expected multi-year gaps. DRAM cumulative capital expenditure over three years is adjusted up to $36.4 billion, with year-end 2028 WSPM increasing by 880K to 2.8mn; 60% of the incremental capacity is allocated to HBM, with an overall capex-to-sales ratio of 13%. NAND three-year spending is adjusted up to $8.6 billion, with 2028 WSPM increasing by 165K to 1.44mn, but most of this is for process migration and greenfield capacity, the latter of which will only take effect in 2H28.

Analysis framework

J.P. Morgan adopted a bottom-up supply-demand framework and multi-scenario analysis. First, based on cloud provider earnings guidance and supply chain research, GPU/ASIC shipment expectations (NVDA, AMD, GOOG TPU, AMZN Trainium, etc.) were梳理 (sorted out) company by company to derive corresponding HBM bit demand. Second, through benchmarking server CPU specifications (channel count, DIMM configuration, socket density), these were quantitatively converted into DRAM capacity demand. Third, combining the S-curve of enterprise SSD penetration and the value weighting of high-speed vs. low-speed SSDs, the expansion of NAND TAM was assessed. On this basis, the report constructed a three-scenario cycle model (Bull Case, Base Case, Bear Case), with significant differences in assumptions regarding LLM parameter expansion, model transitions, cloud vendor hardware spending, and new technology adoption across scenarios. Finally, historical cycle benchmarking (2016-18 upcycle, 2018-19 downcycle, 2023-24 current upcycle) was utilized to identify the sustainability of supply-demand imbalances and the evolution of pricing power.

Methodology notes

• Industry/Sector Analysis FrameworkSupply-demand framework

  Storage market cycles are driven by the mismatch between the supply-side capacity construction cycle and the demand-side AI computing architecture upgrade.

  This report emphasizes that supply bottlenecks are key to determining pricing power. Manufacturing capacity for DRAM/NAND/HBM is constrained long-term by EUV tool procurement and fab expansion cycles, while AI inference computing demand expands rapidly within 12-24 months, creating a typical situation of supply falling short of demand. This mismatch is particularly prominent in 2026-28, granting suppliers significant ASP pricing space.

• Industry/Sector Analysis FrameworkVolume-Price Split

  Storage TAM expansion is driven jointly by volume (bit demand doubling) and price (ASP rising 20-30%).

  J.P. Morgan explicitly separates the dimensions of volume (bit demand) and price (unit ASP) in the TAM model. Between 2026 and 2028, bit demand growth exceeds 45%, while the historical ASP decline of -10~-20% reverses to +10~+30%. The resulting TAM growth (8-fold over three years) from the product of these two factors far exceeds that driven by a single dimension.

• Industry/Sector Analysis FrameworkUpstream-Midstream-Downstream Transmission

  Cloud vendor AI capex growth → Simultaneous surge in GPU/CPU/Memory demand → Intensified competition for storage capacity allocation → Enhanced pricing power in long-term contracts.

  The report tracks the mechanism of demand transmission from the CSP layer (Google, Meta, Amazon, Microsoft, etc.) to storage vendors. The proportion of storage in CSP hardware spending jumps from an initial mid-teens% to 52-73%. This elevated status makes storage capacity a scarce resource, in turn strengthening storage vendors' bargaining power in Long-Term Agreements (LTA).

• Cycle and Prosperity FrameworkProsperity Inflection Point Analysis

  The market has entered a new 7-8 year upcycle, driven by the evolution of AI computing from single-chip to heterogeneous multi-chip architectures.

  Benchmarking against history, the 2016-18 upcycle (7 quarters) and 2018-19 downcycle (6 quarters), the uniqueness of the current cycle (launched in Q3 2023) lies in the demand inflection point being driven by AI architectural innovation rather than traditional process iteration. This implies stronger cycle sustainability and greater price resilience.

• Company Fundamentals and Financial FrameworkWorking capital cycle

  Under the Long-Term Agreement (LTA) model, storage vendors experience extended accounts receivable cycles but improved cash flow stability, expanding operating leverage.

  The report specifically points out that compared to the traditional spot market, while LTA agreements may have slightly lower unit prices, they eliminate inventory risk and price volatility pressure through order certainty. In an environment with gross margins of 80%+, LTA becomes a key indicator of earnings quality.

• Valuation MethodPS valuation

  Storage stocks traditionally trade at a PE/PB discount, but when OPM (Operating Profit Margin) rises from 30% to 75%, valuation should be based on Operating Profit (OP) rather than sales.

  J.P. Morgan notes that storage stocks currently trade at 3.6x P/S but are undervalued, as OPM is expected to stabilize at 72-76% by 2028. If priced based on OP, the implied $1.7 trillion TAM at 3.6x P/S could translate to 89% upside in 2027. This reflects the market's underestimation of margin sustainability.

• Valuation MethodValuation Anchor

  CSP hardware capex scale and storage penetration rate constitute the ultimate anchor for TAM.

  The research report anchors the permanent high-value status of storage in AI infrastructure through multiple external validation points, including NVDA's 2030 spending guidance of $3-4 trillion, AMD's 35%+ annual growth expectation, and Intel's CPU commoditization. This is fundamentally different from the cyclical nature of traditional PC/mobile phone demand.

Asset mapping & comparison

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

• SK hynix (000660.KS)

  Leader in HBM technology; significant room to increase HBM4/4E share when Samsung faces capacity pressure; conventional DRAM products also benefit from rising AI server demand.

  Strengths

  Leading HBM process (1cnm mono-die); HBM4 qualification progress ahead of competitors; backend yield gradually improving; traditional DRAM business holds a solid position in the high-end server market.

  Weaknesses

  Conventional DRAM product market share is lower than Micron's; cost competitiveness is slightly weaker; HBM 4E sample shipment schedule (planned for 2H27) is relatively lagging.

  Comparison

  Compared to Samsung, which is in a catching-up position for HBM4 qualification, SK hynix holds a competitive advantage due to superior process technology and yields; compared to Micron's aggressive expansion strategy, SK hynix maintains steady growth.

  Risks

  Insufficient HBM capacity to meet explosive demand from ASIC vendors could lead to loss of market share; geopolitical factors (South Korean export controls) may affect supply.

• Micron Technology (MU)

  Aggressive DRAM/NAND capacity expansion plans (ID1 fab, Singapore packaging plant); strengthened supply chain position highly tied to cloud vendors.

  Strengths

  As a US company, benefits from convenient export licenses, relatively alleviating competitive pressure from the Chinese market; NAND capacity expansion is more aggressive than SK/Samsung, potentially securing more ASIC vendor orders; leading industry expertise and financial strength in storage.

  Weaknesses

  DRAM cost competitiveness is relatively slightly weaker; HBM technology process lags behind (1bnm base die); new capacity has a long maturation cycle and high investment pressure.

  Comparison

  Compared to Samsung, Micron has a first-mover advantage in HBM4 qualification; Micron mainly relies on capacity scale and customer stickiness; compared to SK hynix's steady strategy, Micron's aggressive expansion faces higher uncertainty.

  Risks

  Delays in new capacity launch or slower-than-expected ramp-up will drag down financial performance; escalation of US-China trade controls may limit emerging market share.

• Samsung Electronics (005930.KS)

  Benefits from the dual-main business structure of DRAM/NAND in its memory business; HBM4 sample shipments verify technical progress, but faces the greatest competitive pressure from China.

  Strengths

  Strongest comprehensive competitiveness; complete product line covering DRAM/NAND/HBM; HBM4/4E qualification certification launched, with room to catch up despite slightly lagging progress; best-in-class brand recognition and customer stickiness.

  Weaknesses

  Rapid volume share increase by Chinese competitors puts pressure on Samsung's traditional markets; HBM cost (relative to single-layer single-chip processes) and pricing competitiveness remain to be verified.

  Comparison

  Compared to SK hynix and Micron, Samsung has the advantage of product line completeness, but no single item has absolute dominance; compared to Chinese vendors, Samsung leads in technology but faces increasing pressure on cost/price competitiveness.

  Risks

  Accelerated domestic substitution supported by the Chinese government may erode mid-to-low-end DRAM/NAND market share; if HBM capacity allocation and yields fall short of expectations, it may lose orders from major customers like NVDA.

• Kioxia Holdings (285A.T)

  Specialized NAND vendor, benefiting from eSSD market expansion and growth in high-speed SLC applications; relatively neutral geographical location in Japan.

  Strengths

  Deep accumulation of NAND professional technology; cost competitiveness is mid-to-high among Japanese and Korean vendors; high flexibility in switching between SLC/TLC, adapting to various application needs.

  Weaknesses

  Capacity scale is relatively smaller than Samsung/SK/Micron, resulting in weaker bargaining power; capital expenditure capability and new technology investment are relatively constrained.

  Comparison

  Compared to the deep competitiveness of the three giants, Kioxia is positioned in specific niche markets (eSSD, high-speed products); compared to the cost advantages of emerging Chinese NAND vendors, Kioxia competes relying on technical specialties.

  Risks

  Fierce competition in the NAND industry; failure to rapidly commercialize high-end product lines (SLC/HBF) may result in lost market opportunities; difficulties in capital financing may constrain capacity expansion.

• Winbond (2344.TW)

  Specialty Memory vendor, benefiting from diversification of storage applications but with relatively limited scale.

  Strengths

  Layout in specialized niche markets (industrial/automotive/IoT, etc.), fast technology iteration, and high customer stickiness; Taiwan's geographical location is relatively advantageous for international trade.

  Weaknesses

  Smallest overall capacity scale, unable to participate in large-capacity AI computing memory competition; relatively limited capital expenditure and R&D investment.

  Comparison

  Compared to general memory giants, Winbond has the smallest volume but the highest degree of specialization; compared to Chinese vendors, Winbond is more competitive in high-end application fields.

  Risks

  If expectations for specialty memory demand in mainstream AI applications decline, it may face performance risks; if competition with large factories invades the mid-range market, cost pressure will surge.

Key data

• Global Storage Total Addressable Market (TAM)$89.6 billion in 2026 → $133.7 billion in 2027 → $168.1 billion in 2028Upward revision of 37-53% compared to the March 2026 model; three-year growth amplitude expanded from 3x to 8x.
• CPU-to-GPU Ratio (Servers)5.4:1 in 2023 → 3.2:1 in 2025 → 2.4:1 in 2028Narrowing trend drives increased DRAM capacity per unit; AI server DRAM bit demand revised up 20-22% for FY27-28E.
• AI CPU DRAM Demand Share19% in 2027 → 24% in 2028AI headnode + AI standalone CPU creates absolute unit demand: 461Mn GB in 2026 → 2.3Bn GB in 2027.
• HBM TAM Upward Revision17-21% (during 2026-28E period)Mainly due to accelerating ASIC demand, upward revision of Rubin GPU units, and ASIC bit mix rising from 33% to 39%.
• HBM ASP GrowthMid-teens% in 2026 → 32% YoY in 2027 (historical high)HBM3E rises 19% in 2027, HBM4 rises 15%; HBM4E commands a 61% premium over HBM3E.
• HBM DRAM Capacity Allocation24% in 2026 → 31% in 202860% of incremental WSPM allocated to HBM, squeezing conventional DRAM supply.
• Enterprise SSD (eSSD) Market Size500EB in 2026 → 1,100EB in 2028; Value TAM exceeds $30 billionCAGR 52%; Rising share of SLC high-speed solutions reduces bit/wafer output, but total volume demand offsets this.
• DRAM Capacity and Capital Expenditure2.8mn WSPM in 2028 (increase of 880K); Three-year cumulative capex $36.4 billionCapex-to-sales ratio 13% (higher than historical average of 30%); EUV procurement and infrastructure are key bottlenecks.
• NAND Capacity and Capital Expenditure1.44mn WSPM in 2028 (increase of 165K); Three-year capex $8.6 billionMost spending is for process migration; greenfield capacity will only take effect in 2H28; priority is lower than DRAM.
• Storage Share in CSP Hardware CapexFrom mid-teens% → Expected 52% in 2026 → 73% in 2027Reflects the status change of storage from an accessory to a strategic asset.
• Storage Stock OPM Expectation76% in 2026 → 75% in 2027 → 72% in 2028Higher than AI semi (63%), foundry (56%), and CSP cloud (39%); relatively lower OPM makes pricing based on OP rather than sales more reasonable.
• NVDA Vera CPU Sales (Standalone)$20 billion in 2026; Unit expectation 600K in 2026 → 3M in 2027Based on 768GB-1.5TB memory configuration, unit demand is 461Mn GB in 2026.
• Chinese Vendor Market Share (Volume)DRAM 6% (2025) → 8-11% (2026-28E); NAND 12% (2025) → 12-16% (2026-28E)Volume share increases significantly, but value share is limited by product mix (DRAM 10%, NAND 12%).
• Implied Upside for Storage Stocks Based on P/S89% (2027 baseline, based on OP perspective)3.6x P/S × $133.7 billion TAM ÷ Current Market Cap; reflects the discount relative to traditional EPS pricing.

Impact & implications

This research has profound implications for the storage ecosystem. First, on the demand side: The heterogenization of AI computing architecture (multi-layer combination of CPU+GPU+HBM+SSD) means storage demand is upgrading from a cyclical consumer good to strategic infrastructure, with cloud providers' permanent capital commitment extending the demand cycle. Second, on the supply side: Fabs face choices regarding investment priorities between HBM vs. conventional DRAM and NAND, while EUV tool shortages and infrastructure construction cycles lasting 24-30 months make manufacturing capacity expansion difficult. Third, on pricing: Supply bottlenecks grant storage vendors bargaining power in LTA agreements, reversing the historical ASP growth space from -10~-20% to +10~+30%, significantly improving earnings quality. Fourth, on the valuation framework: Storage stocks are transitioning from traditional cyclical commodity valuation (PE/PB discount) to long-term high-margin asset valuation, requiring the market to re-examine their value share and sustainability within AI infrastructure. Fifth, on the competitive landscape: While domestic vendors' volume share is increasing, their value share is limited by product structure (mainly mature processes and low-end products), with international giants still dominating advanced processes (HBM, ASIC).

Risks

• If cloud vendor AI infrastructure investment slows due to ROI concerns, it will directly suppress storage demand growth, and TAM expectations could be revised down by 15-25%.
• Escalation of Chinese government export control intensity (MATCH Act or subsequent policies) could cut off advanced capacity supply chains; while decentralization trends strengthen, overall capacity utilization may decline.
• Competition for advanced process capacity between TSMC and Samsung may push up HBM/ASIC base chip costs, compressing the pricing space for finished storage ASPs.
• If commercialization of new memory technologies (such as RRAM, MRAM) progresses faster than expected, they may partially replace traditional DRAM/NAND application scenarios.
• Geopolitical tensions leading to further supply chain fragmentation will increase cross-border procurement costs and inventory costs, putting pressure on storage vendor profit margins.
• If the signing speed of Long-Term Agreements (LTA) fails to meet expectations, a resurgence in spot price volatility could trigger a price war, causing profit margins to fall from 80%+ back to the industry historical average of 40-50%.

What to watch

• Actual shipment progress of NVDA Vera CPUs (target 600K in 2026, 3M in 2027) and confirmation of memory specifications, which directly affect the validation of AI CPU DRAM demand.
• Actual hardware capex spending and AI business ROI of major cloud vendors (Google/Meta/Amazon/Microsoft) in 2026-27 to judge demand sustainability.
• Storage vendor HBM capacity investment and product launch schedules for 2H2027, which will determine the outcome of the high-end market share contest.
• Progress in capacity ramp-up and evolution of cost competitiveness by Chinese vendors (such as CXMT, Tsinghua Unigroup) in DRAM/NAND to assess the threat level of domestic substitution.
• Developments in US/Japan export control policies towards China, especially regarding key supply chain links such as EUV/DUV tools and advanced packaging equipment.
• Spot prices and contract price spreads for DRAM/NAND, reflecting the degree of supply tightness and changes in pricing power.
• Qualification certification progress and unit cost trends for HBM3E vs. HBM4 to judge the multi-generation coexistence cycle and ASP support strength.