The Memory Supercycle: DRAM, HBM, NAND & AI Infrastructure

Intermarket Research Report: DRAM · HBM · NAND · Logic · Foundry · Etch · Deposition · Hyperscaler CAPEX April 2026

DRAM | HBM | NAND | LOGIC | FOUNDRY | ETCH | DEPOSITION | HYPERSCALERS | KEY WINNERS | RISKS

Table of Contents

Executive Summary ............................................................. 3

DRAM: An Unprecedented Price Supercycle ....................................... 4

HBM: The Heart of the AI Supercycle ........................................... 5

NAND: Structural Scarcity Through 2027–2028 ................................... 6

Logic Semiconductors: The AI Demand Engine .................................... 7

Foundry: TSMC's Structural Dominance & the AI Fab Race ........................ 8

Etch Equipment: Structural Oligopoly Strengthens .............................. 9

Deposition Equipment: ALD & CVD Power the AI Architecture ..................... 10

Hyperscaler CAPEX: $630 Billion Reshaping the Supply Chain .................... 11

WFE Spending: No Signs of Deceleration ........................................ 12

Key Players Positioned to Flourish: 2026–2030 ................................. 13

Multi-Year Trajectory: What Happens Next ...................................... 14

Key Risks & Headwinds to Monitor .............................................. 15

Segment Summary: Market Sizes & 2026 Outlook .................................. 16

Glossary of Terms ............................................................. 17

Executive Summary

Executive Summary

The global memory and semiconductor market is undergoing its most significant structural transformation in decades. Driven by the insatiable compute demands of generative AI, large language models, and hyperscaler infrastructure buildouts, every segment of the semiconductor supply chain — DRAM, HBM, NAND, foundry, etch, and deposition — is simultaneously experiencing constrained supply against accelerating demand. The global semiconductor market reached $791.7 billion in 2025 according to the Semiconductor Industry Association, and is on track to breach $1 trillion in 2026 — a milestone that would have seemed implausible just three years ago. High Bandwidth Memory, valued at $34–35 billion in 2025, is projected to reach $54.6 billion in 2026 and $100 billion by 2028, emerging as the single most critical bottleneck in the AI compute stack. The four largest hyperscalers — Amazon, Google, Meta, and Microsoft — are collectively targeting $600–630 billion in 2026 CAPEX, roughly triple the figure just two years prior. Approximately 75% of this spending is directly tied to AI infrastructure: GPUs, HBM-dense servers, data center construction, networking, and power systems. When combined with Oracle and the Stargate Project, total committed AI infrastructure spending exceeds $690 billion for 2026. Wafer fabrication equipment spending is responding in kind, with the WFE market reaching an estimated $115.7 billion in 2025 and projected to grow 24% to $139 billion in 2026 per Barclays. TSMC, the world's dominant foundry, has its 2-nanometer process and all advanced capacity fully booked through 2028, while investing $165 billion in US-based manufacturing across six fabs in Arizona. For investors, operators, and supply chain participants, this report maps the key dynamics, market positions, and multi-year growth trajectories of every critical node in the semiconductor ecosystem. Each section includes verified market data, corrected figures where original sourcing contained errors, and forward-looking analysis through 2030.

Memory · Segment Deep Dive

DRAM: An Unprecedented Price Supercycle

The DRAM market is in the throes of a structural supply shortage with no precedent in modern semiconductor history. TrendForce revised Q1 2026 conventional DRAM contract prices upward to +90–95% quarter-over-quarter, a record-breaking revision driven by a broadening supply-demand gap across every application category — PCs, servers, smartphones, and AI accelerators. Even tier-1 PC OEMs with secured supply allocations are reporting rapidly depleting inventories. In Q3 2025, global DRAM revenue hit $41.4 billion, up 30.9% QoQ, and Q4 2025 saw Samsung's DRAM revenue surge 43% QoQ to $19.3 billion.

DRAMpocalypse — TrendForce Q1 2026 Pricing Alert PC DRAM is expected to increase by more than 100% QoQ in Q1 2026, while server DRAM and LPDDR5X surge ~90%. Goldman Sachs reports that both HBM and conventional DRAM demand continues to significantly outstrip supply. DRAM supplier inventories are described as "nearly depleted." Price increases are expected to persist throughout 2026.

DRAM Market Share — Q4 2025 (TrendForce, Feb 2026)

Company Market Note

SHARE

DRAM Key Metrics

QoQ) QoQ)

Contract Price

Metric Q3 2025 Q4 2025 Q1 2026 Forecast

(2026)

(2026)

2027–2028 Outlook: DDR4 Becomes a Specialty Product

The DDR4 memory standard, which has served as the workhorse of the computing industry since its introduction in 2014, is entering its end-of-life phase as all three major DRAM manufacturers accelerate their migration to DDR5 and HBM production. By 2027, DDR4 supply will be heavily concentrated at Nanya Technology and Winbond Electronics — two Taiwanese firms that have historically focused on mature-node memory — as Samsung, SK Hynix, and Micron complete their transition away from DDR4 wafer production entirely.

The economics driving this shift are straightforward: HBM commands average selling prices 5–10x higher than conventional DRAM, and DDR5 carries a meaningful premium over DDR4. With every major manufacturer operating at or near full capacity utilization, there is zero financial incentive to allocate scarce wafer capacity to DDR4 production when the same wafers can generate dramatically higher revenue as HBM or DDR5. This structural shift is permanent — manufacturers have no plans to reverse their migration trajectory.

DDR4 pricing is expected to follow a pattern similar to DDR3 in its late phase: production volumes will continue to shrink, but prices will become increasingly "sticky" and resistant to downward pressure. As supply concentrates among fewer producers, buyers will face longer lead times, reduced negotiating leverage, and potential allocation restrictions. Industrial, automotive, and embedded systems that rely on DDR4 for long product lifecycles will be particularly affected.

The geographic concentration of remaining DDR4 production adds another layer of supply chain risk. Nanya Technology and Winbond Electronics are both headquartered in Taiwan, meaning that virtually all DDR4 supply post-2027 will originate from a single geography. For defense, aerospace, and critical infrastructure applications that mandate long-term component availability, the DDR4 end-of-life transition requires immediate procurement strategy adjustments — including last-time-buy planning and qualification of DDR5 alternatives where technically feasible.

From an investment perspective, the DDR4 phase-out reinforces the broader memory market thesis: every unit of wafer capacity freed from DDR4 production flows into higher-margin DDR5 and HBM output. This mix shift is a key driver behind the revenue and margin expansion forecasts for Samsung, SK Hynix, and Micron through 2028. Nanya and Winbond, meanwhile, stand to benefit from increasing pricing power as the sole remaining DDR4 suppliers — a niche but potentially lucrative position for firms with lower capital intensity.

DDR4 Supply Transition Timeline

Timeframe Ddr4 Supply Dynamics Recommended Action

DDR4 output; Micron shifting contracts remaining lines

production; Nanya/Winbond become remaining suppliers primary sources

weeks; spot market premiums non-critical applications emerge

product with limited availability premium pricing for legacy needs

Companies with products requiring DDR4 beyond 2027 should take immediate action to secure multi-year supply contracts with guaranteed allocation. The window for favorable contract terms is narrowing rapidly as the remaining DDR4 capacity is committed. Organizations should also evaluate accelerated DDR5 migration timelines for any applications where the transition is technically feasible, as DDR5 will offer both superior performance and significantly better long-term supply security.

Memory · Ai-Critical

HBM: The Heart of the AI Supercycle

High Bandwidth Memory is arguably the single most strategically important component in the AI supply chain. Every NVIDIA GPU generation from H100 onward requires HBM, and capacity per accelerator is growing exponentially. The HBM market was valued at approximately $34–35 billion in 2025 (per Micron and Yole Group), and BofA Securities estimates the 2026 HBM market will reach $54.6 billion — a 58% year-over-year increase. Micron forecasts an HBM TAM CAGR of ~40% through 2028, with the market reaching ~$100 billion by 2028.

Correction: Original report cited $3B for 2025 HBM market. Industry sources (Micron, Yole Group, Counterpoint) place it at $34–35B.

NVIDIA GPU HBM Memory Progression

Gpu Hbm Spec Launch

H100 80 Gb Hbm3 2023

Rubin R100 bandwidth updated to 22 TB/s per GTC 2026 (originally 13–15 TB/s). Each Vera Rubin NVL72 rack requires ~1,152 TB of adjacent NAND.

HBM Revenue Market Share — Q3 2025 (Counterpoint Research)

Company Revenue Key Notes

SHARE

Correction: Original cited 53/35/11. Counterpoint revenue shares: 57/22/21.

SK Hynix overtook Samsung in annual operating profit in 2025 for the first time — 47.2 trillion won vs Samsung's 43.6 trillion won (corrected from 47.7T). Goldman Sachs forecasts HBM demand for ASIC-based AI chips will skyrocket 82%, representing roughly one-third of the total HBM market.

Storage · Flash Memory

NAND: Structural Scarcity Through 2027–2028

The NAND flash market is experiencing "structural scarcity" — a shortage that cannot be resolved within 2026 and is unlikely to normalize until new fab investments come online in 2027–2028. The crisis is driven by surging AI infrastructure demand, simultaneous capacity cuts by Samsung and SK Hynix, Samsung's termination of MLC NAND production, and Kioxia's admission that its entire 2026 production volume is already sold out.

NAND Crisis — 200%+ Price Increase in H1 2026 Samsung is hiking NAND prices ~100% in Q2 2026, following a similar increase in Q1 — implying a cumulative 200%+ rise in H1 2026. TrendForce initially forecast Q1 2026 NAND prices at +33–38% QoQ; revised to +55–60%, then again to +85–90% QoQ. Further revisions remain possible.

NAND Market Share — Q3 2025 (TrendForce)

MLC

focus

Correction: Samsung was 32.3% (not 29.1%). Kioxia was 15.3% (not 16.5%).

Samsung ended MLC NAND production (final shipments June 2026); global MLC capacity forecast to drop 41.7%. Samsung cutting wafer starts 4.5% (4.9M to 4.68M), SK Hynix cutting 10% (1.9M to 1.7M). No greenfield NAND expansions planned. Market estimated at $77.81B in 2026 (Coherent Market Insights), though estimates range $58.7B–$78.2B across firms.

Chips · Ai Accelerators

Logic Semiconductors: The AI Demand Engine

The logic semiconductor market — CPUs, GPUs, AI accelerators, FPGAs, ASICs, and SoCs — is the engine pulling demand through the entire memory and equipment chain. The global logic market was $147.88 billion in 2025, forecast to reach $231.38 billion by 2034 at 5.1% CAGR. AI chips represent less than 0.2% of wafer starts yet generate ~20% of total semiconductor revenue.

TSMC 2nm — A Game-Changing Node TSMC's N2 process entered HVM in Q4 2025 at Fab 22 near Kaohsiung with strong yields. N2 delivers superior performance-per-watt vs N3, positioning it as the preferred node for NVIDIA Rubin, AMD MI400, and high-end smartphones. TSMC has raised prices across all nodes at 5/4nm and below for 2026, with further increases expected in 2027.

Key Logic Semiconductor Players

Position

Designer Rubin

Designer

Networking XPUs

Overflow

Apple Silicon

Manufacturing · Pure-Play & Idm

Foundry: TSMC's Structural Dominance & the AI Fab Race

IDC projects the Foundry 2.0 market to exceed $360 billion in 2026, up 17% from 2025. TrendForce forecasts total foundry revenue of $218.8 billion in 2026 (+24.8% YoY), with TSMC at ~32% YoY growth. TSMC's 2nm process and all capacity at 5/4nm and below is fully booked through 2028.

TSMC 2025 revenue: $122.42B (+35.9% YoY). Guides ~30% growth in 2026. Advanced nodes (7nm and below) = 77% of revenue; N3 = 28% of Q4 2025 wafer revenue. AI accelerator revenue CAGR: mid-to-high 50% (2024–2029). US investment expanded to $165B across 6 fabs (corrected from 5) plus 2 advanced packaging facilities in Arizona.

Foundry Competitive Landscape — 2026

Foundry Share Leading Ai Capacity Outlook

2026 Edge Status

2028

dev

es

Equipment · Pattern Etching

Etch Equipment: Structural Oligopoly Strengthens

The global etch equipment market is valued at $27.33 billion in 2026, forecast to reach $39.43 billion by 2031 at 7.61% CAGR. Applied Materials, Lam Research, and Tokyo Electron collectively control ~75% of global etch revenue, benefiting from both advanced logic (GAA transistors) and memory (3D NAND layer escalation, HBM via formation) transitions.

Etch Equipment Market Share — 2025

Company Share Key Strengths

ASP

wins

deposition

globally

Equipment · Thin-Film Deposition

Deposition Equipment: ALD & CVD Power the AI Architecture

ALD equipment market: $9.55B in 2025, $10.34B in 2026, forecast $20.79B by 2034 at 9.1% CAGR. CVD market: $29.10B in 2026, reaching $54.83B by 2033 at 9.5% CAGR. As nodes shrink to 2nm and 3D NAND exceeds 300 layers, ALD is superseding conventional CVD for precision applications.

Deposition Technology Landscape — 2026

2026

TSV Lam

dielectrics Kokusai

barriers

Molybdenum Deposition — The Next Inflection Lam Research's ALTUS Halo (launched Feb 2025) is purpose-built for molybdenum ALD deposition at sub-3nm nodes — a first-mover advantage expected to become a significant equipment category by 2027 as TSMC N2 and A16 process generations ramp.

Demand Driver · Infrastructure

Hyperscaler CAPEX: $630 Billion Reshaping the Supply Chain

The four largest hyperscalers collectively plan to spend $600–630 billion in 2026 CAPEX, ~3x the figure two years prior. ~75% (~$450B) is directly tied to AI infrastructure. This buildout is the primary demand force behind every shortage described in this report.

Hyperscaler CAPEX — 2026 Estimates

Company 2026 2025 Yoy Key Notes

Capex Actual

commitment; custom Trainium ASICs

Google HBM; Gemini demand

5GW Hyperion (LA)

Azure constraints; Maia chips Correction: Amazon 2025 CAPEX was $131.8B (not $125B); YoY growth +52% (not +60%).

Stargate Project (OpenAI, SoftBank, Oracle): $500B AI infrastructure commitment over four years through 2029. Combined "Big Five" CAPEX pushes to $690B+. Deloitte forecasts AI data center CAPEX at $400–450B in 2026, rising to ~$1T by 2028.

Correction: Original cited Deloitte at "$1T in 2026, $2T in 2028." Actual: $400–450B / ~$1T.

Power availability has emerged as a binding constraint that may ultimately limit the pace of AI infrastructure buildout more than chip supply. Microsoft's $80 billion Azure backlog is unfulfilled primarily due to power constraints, not chip shortages. Meta's planned 5GW Hyperion facility in Louisiana would consume more electricity than many small countries. The intersection of semiconductor demand and energy infrastructure represents a new investment thesis that spans both technology and utilities sectors.

The concentration of spending among just four companies creates both opportunity and risk for the semiconductor supply chain. The unprecedented demand visibility — with hyperscalers placing orders 18–24 months in advance — gives equipment makers and memory producers the confidence to invest in capacity expansion. However, any pullback or delay by even one of these

four players could cascade through the supply chain, creating a bullwhip effect in equipment orders and memory pricing. The 2022–2023 memory downturn, triggered in part by a sudden pullback in enterprise server orders, demonstrated how quickly sentiment can shift even in structurally tight markets.

Supply Chain Cascade: How Hyperscaler CAPEX Propagates

Action Immediate Impact Beneficiaries

NAND Micron

Nvidia Tsmc

fabs categories

Equipment · Market Outlook

WFE Spending: No Signs of Deceleration

WFE market was ~$115.7B in 2025 (SEMI, corrected from $109–110B). KLA expects mid-$130B range in 2026. Barclays projects $139B in 2026 (+24%), $159B in 2027. Morgan Stanley: $128B for 2026. By 2034, WFE projected at $248.6B at 9.33% CAGR.

WFE Equipment Companies — FY2025 Actuals (Corrected)

Rev. Share K

Materials broadest portfolio

Halo); high spares ratio

capex doubled 2025

nopoly sub-2nm; €38.8B backlog

91.1% semi revenue Growth

International NAND critical equipment Correction: AMAT was $28.37B (not ~$22B). KLA was $12.16B (not ~$11B). Lam Research returned ~150%+ in CY2025 (not 107%).

Investment Thesis · Beneficiaries

Key Players Positioned to Flourish: 2026–2030

These companies benefit from oligopolistic structures, multi-year demand visibility from hyperscaler CAPEX, and technology moats that competitors cannot easily replicate.

Tsmc (Nyse: Tsm)

Controls 72–77% of advanced logic. N2 in HVM. Capacity booked through 2028. $165B US investment. 62%+ gross margins. Key Catalyst: Rubin GPU ramp + HBM4 volume 2026–2027.

SK Hynix (KRX: 000660) 57% HBM revenue share. First to mass-produce HBM3E. UBS forecasts ~70% HBM4 share for Rubin. Overtook Samsung in annual profit (47.2T won). $13B P&T7; packaging facility. Key Catalyst: HBM4 mass production + NVIDIA Vera Rubin allocation.

Lam Research (NASDAQ: LRCX) 28% etch + dominant ALD. 42% GAA etch wins. ALTUS Halo pioneering Mo ALD for sub-3nm. 34% recurring spares/software revenue. ~150%+ return in 2025. Key Catalyst: GAA ramp at TSMC N2/A16 + HBM via etch volume.

Micron Technology (NASDAQ: MU) DC revenue surged 137% YoY to $20.75B in FY25 (56% of sales). HBM 2026 capacity sold out. HBM4 samples for Rubin. $200B US expansion (Idaho + New York). Key Catalyst: HBM4 NVIDIA qualification + Foundry 5nm AI wins.

Applied Materials (NASDAQ: AMAT) Broadest WFE portfolio. 18% share. $28.37B FY2025 revenue. Integrated Materials Solutions: 42% of etch revenues. WFE forecast +24% in 2026. Key Catalyst: WFE acceleration + 2nm/A16 process tool ramp.

Asml (Nasdaq: Asml)

Near-monopoly on EUV lithography. High-NA EUV shipping for sub-2nm. €38.8B backlog. Key Catalyst: High-NA EUV adoption at TSMC A16 / Samsung 2nm.

Forward Outlook · 2026–2030

Multi-Year Trajectory: What Happens Next

2026: Peak Scarcity

• DRAM +90–95% QoQ in Q1 (record); NAND +85–90% QoQ revised
• HBM3E dominates AI GPU deployments; HBM4 in qualification
• Global semiconductor market breaches $1 trillion; Hyperscaler CAPEX ~$630B
• WFE spending grows 24% to ~$139B

2027: Transition Year

• HBM4 ramps for NVIDIA Rubin, AMD MI400; new NAND fabs begin easing shortages
• TSMC Arizona Fab 21 Phase 2 (3nm) enters production; DDR4 becomes specialty product
• WFE grows to ~$159B; Advanced packaging exceeds $20B

2028: Scale & Normalization

• HBM TAM reaches ~$100B; NAND shortage resolves with new capacity
• TSMC Arizona Fab 4 (sub-2nm) begins construction; 3D NAND hits 400+ layers
• Global semi market approaches $1.2T+; Advanced packaging at $25B+

2029–2030: Structural Maturity

• HBM5 development; 3D DRAM architectures may debut
• China domestic supply chain matures; 10+ custom AI ASICs competing
• TSMC targets ~25% CAGR through 2029; Semi equipment approaches $200B+/year
• HBM projected at ~50% of DRAM revenue by 2030

Analysis · Downside Scenarios

Key Risks & Headwinds to Monitor

Geopolitical / Export Controls US export restrictions on advanced equipment to China could disrupt supply chains. Deloitte forecasts $30B+ in technologies affected by trade barriers in 2026. TSMC Taiwan concentration remains a systemic risk for the global AI supply chain.

AI CAPEX Pullback Risk If AI monetization disappoints, CAPEX could be deferred. Microsoft Azure has $80B unfulfilled backlog due to power constraints — demand signal but also dampener if power buildout lags.

Currency & Macro Risk Strong USD vs KRW/TWD amplifies costs. Rising energy costs for fabs and data centers. Interest rate environment affects equipment financing.

HBM Yield & TSV Challenges TSV yield challenges with 16-high HBM stacks. Samsung quality issues with prior HBM generations required remediation before NVIDIA qualification.

China Domestic Semiconductor Rise China 22–40nm capacity forecast at 42% of global output by 2028. NAURA/AMEC achieved ~40% domestic etch substitution. Technology breakthrough could reshape Western exposure.

Reference · Quick Guide

Segment Summary: Market Sizes & 2026 Outlook

(5Y)

Shortage Micron

B* Micron

Scarcity Hynix

Booked Foundry

Equipment

Equipment

Equipment

Spending

*HBM estimates vary by firm — from $3.81B (Business Research Co.) to $54.6B (BofA). The $34.6B reflects Micron/Yole 2025 actuals. $100B is Micron's 2028 TAM forecast.

Reference · Definitions

Glossary of Terms

ALD — Atomic Layer Deposition — deposits one atomic layer at a time for extreme precision. ASIC — Application-Specific Integrated Circuit — chip for a specific purpose (e.g., Google TPU). CAGR — Compound Annual Growth Rate — annualized average rate over a specified period. CAPEX — Capital Expenditure — funds for acquiring/maintaining physical assets. CVD — Chemical Vapor Deposition — depositing thin films from gaseous precursors. DDR4/DDR5 — Double Data Rate 4th/5th generation synchronous dynamic RAM standards. DRAM — Dynamic Random-Access Memory — volatile memory for PCs, servers, AI accelerators. EUV — Extreme Ultraviolet Lithography — 13.5nm wavelength light for sub-7nm patterning. FinFET — Fin Field-Effect Transistor — 3D transistor architecture at 22nm–5nm nodes. GAA — Gate-All-Around — transistor architecture succeeding FinFET at 2nm and below. HBM — High Bandwidth Memory — vertically stacked DRAM via TSVs for AI/GPU workloads. HBM3/3E/4 — Successive HBM generations with increasing capacity and bandwidth. High-NA EUV — Next-gen lithography from ASML for sub-2nm patterning. HVM — High-Volume Manufacturing — commercial-scale chip production. IDM — Integrated Device Manufacturer — designs and fabricates own chips (Samsung, Intel). LPDDR5X — Low-Power DDR5X — energy-efficient DRAM for mobile and edge AI. MLC — Multi-Level Cell — NAND storing 2 bits/cell; being phased out for TLC/QLC. NAND — Non-volatile flash memory (NOT-AND logic gate) used in SSDs and storage. NVL72 — NVIDIA's 72-GPU server rack for AI training and inference. OSAT — Outsourced Semiconductor Assembly and Test — third-party packaging/testing. QLC — Quad-Level Cell — NAND storing 4 bits/cell; lower cost but lower endurance. QoQ — Quarter-over-Quarter — metric comparison vs. immediately preceding quarter. TLC — Triple-Level Cell — NAND storing 3 bits/cell; mainstream for SSDs. TSV — Through-Silicon Via — vertical connection through silicon, critical for HBM stacking. WFE — Wafer Fabrication Equipment — tools/machinery for manufacturing semiconductor wafers. WSTS — World Semiconductor Trade Statistics — tracks global semiconductor sales. YoY — Year-over-Year — comparison vs. same period one year prior.

Disclaimer: This report is produced for informational and educational purposes only. It does not constitute investment advice, a solicitation to buy or sell securities, or financial recommendations. All data sourced from publicly available research, SEC filings, and analyst estimates as of April 2026. Readers should conduct their own due diligence and consult a licensed financial professional.

Intermarket Universe Research Report Sources: TrendForce · Goldman Sachs · BofA Securities · Barclays · Morgan Stanley · Counterpoint Research · IDC · SEMI · SIA · WSTS · Deloitte Global · Micron IR · TSMC IR · SK Hynix Newsroom · CNBC · Reuters

Intermarket Universe · April 2026

This research is for informational purposes only and does not constitute investment advice. Intermarket Universe does not hold positions in any securities mentioned unless disclosed. All estimates are the author's own analysis derived from public information.