YK Research

High-Bandwidth Memory (HBM)

Three Companies Control What NVIDIA Can't Build Without.

Last updated: 14 March 2026 · YK Research

Contents

  1. March 2026 Update: Micron at $426
  2. Market Snapshot: Three-Player Oligopoly
  3. The Real Bottleneck: Memory Controls NVIDIA
  4. Margin Structure: 9.1x Premium Over DRAM
  5. Execution Over Capital: Samsung Is Losing
  6. Micron: Highest Conviction at $426
  7. Three Horses, Different Races
  8. Valuation and Scenarios
  9. Risk Matrix
  10. Investment Framework

March 2026 Update: Micron at $426

Micron (MU) Current Price: $426.13
Original Bull Case Target: $300-400
Thesis Validation: The HBM thesis is playing out stronger than expected. Micron captured 22% market share from 5%. HBM4 ramp on schedule for 2026. 118% utilization proves structural supply shortage. Stock exceeded original bull case targets.
Performance vs Bull Case
+6.5% Above Target
Market Share Gained
17 points (5%→22%)
HBM Revenue Run Rate
$11.8B (2027E)
NVIDIA (NVDA) Price
$180.25

Market Snapshot: Three-Player Oligopoly

Market Size (2023)
$4.0B
Market Size (2030E)
$91.0B
CAGR (2023-2030)
56%
Market Utilization
118%
Number of Suppliers
3
ASP Premium vs DRAM
9.1x

Why HBM Is the Real Bottleneck

Everyone focuses on NVIDIA's GPU pricing power. Here's the reality: NVIDIA is constrained by HBM supply. Memory costs 3.5x more than the GPU die in an H200. Three companies control what NVIDIA can't ship without.

The Three Oligopolists

SK hynix: Execution Leader

  • 51% market share through superior yields
  • 80% yields vs Samsung's 50%
  • First to HBM4 sampling Q1 2025
  • $14.5B revenue target by 2025

Micron: The Leapfrog Play

  • 22% market share up from 5%
  • First HBM3E shipper in Q1 2024
  • 30% power efficiency edge
  • 59x revenue growth potential by 2027

Samsung: Spending Without Winning

  • 27% market share down from 42%
  • 50% yields: scrapping half of production
  • $6.8B capex with nothing to show for it
  • 12 months late to NVIDIA qualification

Market Structure

  • Multi-year qualification creates stickiness
  • 18-24 month customer lock-in cycles
  • Process execution beats scale in this market
  • Each generation resets competitive dynamics
🔍
“This isn't expensive DRAM. It's systems integration. The value is TSV manufacturing, thermal management of 12-layer stacks, and packaging expertise. Capital doesn't buy excellence. Execution does.”

The Real Bottleneck: Memory Controls NVIDIA

Every H200 needs 5-7 HBM3E stacks at $1,550 each. Memory cost: $8,800. GPU die cost: $2,500.

The Cost Breakdown

HBM Stacks per H200
5-7
HBM Cost per Stack
$1,550
Total Memory Cost
$8,800
GPU Die Cost
$2,500
Memory vs Die Ratio
3.5x
Memory % of BOM
30%

Utilization: A Seller's Market

118% utilization for 18+ months. Customers can't ship without you. Who has the pricing power? Not NVIDIA.

Why This Won't Change

  • 3 players globally. New entrants can't break in. Even Intel gave up.
  • $5B+ capex to reach scale. Proprietary TSV processes and advanced packaging.
  • Multi-year qualification cycles. NVIDIA doesn't swap suppliers for 5% savings.
  • Content inflation: HBM per GPU going from 80GB in 2023 to 512GB by 2030.
🔍
“When NVIDIA's H200 launch got delayed, it wasn't GPU dies. It was memory. Memory is the bottleneck in AI infrastructure. Only three companies can make it.”

Margin Structure: Not Your Father's DRAM Business

HBM margins are expanding to 35-40%. Standard DRAM sits at 25%. Different business.

The Premium That Defies Gravity

Standard DRAM Price
$3.50/GB
HBM3E Price
$32/GB
Premium Multiple
9.1x
DRAM Margins
25%
HBM Margins
35-40%
Margin Expansion
+15 pts

Product Economics by Generation

ProductCostASPMargin$/GBStatus
HBM3 8-Hi$850$1,20029%$50Current
HBM3E 8-Hi$1,020$1,55034%$65Ramping
HBM3E 12-Hi$1,560$2,40035%$672025
HBM4 12-Hi$1,908$3,20040%$672026

Why Margins Keep Expanding

  • Not a memory business. The value is systems integration, thermal design, and TSV manufacturing.
  • Each generation adds complexity. HBM4 means 16-Hi stacks, 2048-bit interface, 50% less power.
  • 18-24 month qualification prevents price shopping.
  • 118% utilization gives suppliers pricing power.
🔍
“Samsung outspends SK hynix 2:1 on capex but has 50% yields while SK hynix hits 80%. Capital doesn't buy excellence in HBM. Execution does.”

Execution Over Capital: The Samsung Paradox

The Samsung Paradox

Samsung has unlimited capital, world-class fabs, and decades of memory expertise. Yet it's hemorrhaging market share. HBM rewards process execution, not scale.

Samsung 2024 Capex
$6.8B
Samsung Yield Rate
50%
SK hynix Yield Rate
80%
Market Share Lost
15 points

Yield Rates Determine Winners

What 50% Yields Mean

  • Every other HBM stack goes to waste.
  • $1,550 ASP at 50% yields means negative margins.
  • NVIDIA can't bet H200 supply on coinflip yields.
  • 42% to 27% share in two years.

Why Execution Beats Scale in HBM

Technical Challenges

  • TSV drilling: thousands of sub-micron holes
  • 12-layer stack thermal management
  • Advanced packaging (CoWoS, Fan-out)
  • Reliability at extreme densities

Process Learning Curve

  • Optimization takes years
  • Iteration beats capital investment
  • Customer feedback loops are critical
  • First-mover advantages compound
🔍
“Micron skipped HBM3 entirely, went straight to HBM3E, and shipped first. 5% to 22% share in two years. That's not capital allocation. That's surgical execution.”

Micron: Highest Conviction at $426

$200M to $11.8B revenue in 4 years. 59x growth. NVIDIA design win lock-in.

Current Position

Current Price
$426.13
Market Cap
$46B
HBM Market Share
22%
HBM Revenue (2023)
$200M
HBM Revenue (2027E)
$11.8B
Revenue Multiple
59x

Why Micron Wins

1. Leapfrog Execution

  • Skipped HBM3. Went straight to HBM3E.
  • First to ship HBM3E. Q1 2024.
  • NVIDIA H200 qualified. 18+ months of design win stickiness.
  • 30% power advantage. Data centers care about this.

2. Geopolitical Hedge

  • Only US supplier. Reduces Korea and Taiwan concentration.
  • CHIPS Act subsidies fund capacity expansion.
  • NVIDIA's second source strategy needs them.
  • DoD and hyperscaler preference.

Revenue Transformation Path

Positioning vs Competitors

SK hynix: Dominant
51% Share
Best Yields (80%)

Market leader. Superior execution. Safe but priced as the winner.

Micron: Ascending
22% Share
Fastest Growth

Taking Samsung's share through execution and geopolitical positioning.

Samsung: Struggling
27% Share
Yield Issues (50%)

Resources are there. Execution is not. Show-me story.

The Asymmetric Opportunity

The market hasn't priced in the HBM shift. Micron still trades like a cyclical DRAM company.

  • 18-month qualification creates a moat.
  • 30% better power efficiency than competitors.
  • Only US supplier in a critical technology.
  • HBM could be 20%+ of Micron revenue by 2027.
🔍
“At $426, Micron is pricing in continued DRAM cycles. The market hasn't recognized that HBM is becoming the dominant profit driver. $11.8B revenue at 35-40% margins by 2027.”

Three Horses, Different Races

The HBM4 Race

SK hynix: The Execution Machine

Market Position
HOLD/BUY
2027E Revenue
$26.8B

Bull Case

  • 51% market share defended through superior yields
  • First to sample HBM4 (Q1 2025)
  • Proprietary MR-MUF packaging advantage
  • Locked into NVIDIA network

Bear Case

  • Consensus play. Market already knows they dominate.
  • Priced accordingly. Limited upside.
  • Vulnerable to geopolitical Korea risks

Micron: The Leapfrog Play

Market Position
BUY: Highest Conviction
2027E Revenue
$11.8B

Bull Case

  • 59x revenue growth potential ($200M to $11.8B)
  • Only non-Asian supplier (geopolitical hedge)
  • 30% power efficiency advantage
  • NVIDIA H200 design win stickiness

Bear Case

  • Smallest player.
  • Limited capacity vs leaders.
  • Must execute HBM4 transition perfectly

Samsung: The Turnaround Gamble

Market Position
WATCH
Current Yields
50%
Capex (2024)
$6.8B
Share Lost
15 points

Bull Case (IF yields improve)

  • Unlimited capital and fab capacity
  • Custom HBM for Microsoft and Meta
  • Aggressive HBM4 timeline (Q4 2025)
  • If yields hit 65%+, could recapture 35% share

Bear Case (Current trajectory)

  • 50% yields. Half of production scrapped.
  • 12 months late to NVIDIA qualification
  • 42% to 27% share in two years
  • Execution doesn't match the spend

Why No One Else Can Enter

  • China's YMTC tried with state backing. Gave up.
  • Intel invented memory chips. Doesn't compete in HBM.
  • $5B+ capex and years to reach scale.
  • Proprietary process knowledge in TSV, thermal, and packaging.
🔍
“Think of this like foundry wafers: oligopoly, lock-in, technology transitions. Not commodity DRAM. Three suppliers control what NVIDIA can't live without.”

Valuation and Scenarios

Micron Scenarios (Current: $426.13)

Bear Case (25%)
$280
-34% from current

HBM commoditizes. Samsung recovers share. Yields plateau at 65%. Memory cycle depression.

Base Case (50%)
$520
+22% by 2027

Micron holds 22% share. HBM4 succeeds. 35% margins. 18x P/E on mixed revenue.

Bull Case (25%)
$750
+76% by 2027

Micron reaches 30% share. HBM becomes 25% of revenue. 40% margins. 22x P/E.

Valuation Drivers

Key Metrics to Track

Operational Metrics

  • HBM yield rates: 70%+ needed for competitive margins
  • Market share: 22% today. Watch for 25%+
  • ASP trends: HBM4 premium vs HBM3E
  • Capacity utilization: Should remain >100% through 2025

Financial Metrics

  • HBM revenue mix: Path to 20%+ of total by 2027
  • Blended gross margins: HBM lifts overall to 32%+
  • Capex efficiency: HBM revenue per $ invested
  • Free cash flow: $8B+ potential at peak cycle
🔍
“At $426, the market prices Micron as a cyclical memory company. HBM is becoming a high-margin, sticky, oligopoly business. That justifies 20x+ multiples on the HBM portion.”

Risk Matrix

RiskSeverityProbabilityImpact on ThesisMitigant
HBM commoditizationHIGHLOWIf margins compress to DRAM levels at 25%, the entire thesis breaks. Oligopoly dissolves into commodity.3-player market with massive barriers. Each generation resets dynamics. History shows sustained premiums.
AI demand plateauHIGHMEDIUMIf AI buildout slows, HBM growth falls below 56% CAGR. Content inflation stops. Utilization normalizes.Even at 30% CAGR, the market reaches $45B by 2030. Still massive vs the $4B base.
Samsung execution recoveryMEDIUMMEDIUMIf Samsung yields improve from 50% to 70%, they recapture share. Micron's 22% is at risk.Market is large enough for multiple winners. The geopolitical hedge gives Micron a defensible niche.
New memory technologyMEDIUMLOWCXL, processing-in-memory, or other architectures could disrupt HBM by 2028-2030.HBM4/5 roadmap runs through 2030. Technology transitions take 5+ years.
Chinese competitionMEDIUMLOWIf China develops domestic HBM capability, global oligopoly breaks. YMTC returns to market.YMTC already tried and failed. Export controls limit advanced packaging equipment access.
Memory cycle downturnLOWMEDIUMA traditional DRAM cycle downturn could drag HBM demand despite structural advantages.AI infrastructure is less cyclical than consumer and enterprise. HBM is structurally different.

Investment Framework

Core Investment Thesis

🔍
“Three companies control what NVIDIA can't live without. In 2024, HBM hit 118% utilization. Demand exceeded supply by 18%. When customers are sold out for 18 months and can't ship without you, who has pricing power? Memory suppliers.”

Allocation Strategy

MICRON: Core
5-7%
Highest Conviction

Asymmetric opportunity. 59x revenue potential. Geopolitical hedge. Execution is proving out.

SK HYNIX: Anchor
2-4%
Defensive Hold

Market leader with best yields. Lower upside. Sleep-well-at-night execution.

SAMSUNG: Watch
0%
Turnaround Play

Wait for yields to hit 65%+ before entry. Show-me story.

Entry/Exit Triggers

Buy Triggers (Add Aggressively)

  • Micron HBM4 yields hit 70%+
  • Market share breaks above 25%
  • HBM revenue becomes 15%+ of total Micron revenue
  • Pullback to $350-380 range

Sell Triggers (Reduce/Exit)

  • HBM utilization drops below 90%
  • ASP declines over 15% year-over-year
  • Samsung achieves 70%+ yields
  • New entrants announce credible HBM capacity

Timeline

  • 2025: HBM4 sampling begins. Market share battle intensifies.
  • 2026: HBM4 production ramp. Next-gen GPU launches.
  • 2027: Full-scale HBM4 deployment. $91B market trajectory.
  • 2030: Market maturity. Content inflation still drives demand.

Quarterly Catalysts

  • Yield rate improvements (target: Micron 70%+)
  • Market share data from each company
  • HBM ASP trends and margin expansion
  • Capacity utilization rates industry-wide

Long-term Structural

  • AI infrastructure spending sustainability
  • Content inflation per GPU: 80GB to 512GB
  • New memory architecture threats
  • Geopolitical supply chain shifts
🔍
“This isn't a cyclical trade on memory prices. Memory is becoming the AI infrastructure constraint. Suppliers capture more value than anyone expects. The market hasn't realized it yet.”