The bottleneck for AI compute has shifted from compute itself to memory bandwidth and capacity. HBM (High Bandwidth Memory) , as a core component of AI chips, has a 2026 market size of $80B+, but there are only 3 suppliers globally — SK Hynix, Samsung, Micron. This article provides an in-depth analysis of this "memory three kingdoms" battle.
Why Is HBM Critical in the AI Era?
The Memory Wall
AI model sizes have grown from BERT (340M parameters) in 2018 to Llama 3 (405B) in 2024 and Gemini (estimated 1T+) in 2026, with compute growing 1,000× but memory bandwidth only growing 10-20×.
This is the famous von Neumann bottleneck:
| Metric | 2018 (V100) | 2024 (H100) | 2026 (Rubin R200) | Growth |
|---|
| Compute (FP16/BF16) | 125 TFLOPS | 989 TFLOPS | 25 PFLOPS | 200× |
| Memory Capacity | 32 GB | 80 GB | 288 GB | 9× |
| Memory Bandwidth | 900 GB/s | 3.35 TB/s | 22 TB/s | 24× |
Compute growth far outpaces memory bandwidth growth, causing GPUs to frequently "wait for data." HBM is the core solution to alleviate this bottleneck.
HBM vs GDDR vs SRAM
| Memory Type | Bandwidth (per pin) | Capacity Density | Power | Use Case |
|---|
| HBM4 | 6.4 Gbps/pin | High (12-Hi) | Medium | AI training / inference GPU |
| HBM3e | 4.8 Gbps/pin | High (8-Hi/12-Hi) | Medium | AI training / inference GPU |
| HBM3 | 4.0 Gbps/pin | Medium | Medium | AI training |
| HBM2e | 3.2 Gbps/pin | Medium | Medium | AI inference |
| GDDR6X | 1.6 Gbps/pin | Medium | Med-High | Consumer GPUs |
| GDDR7 | 2.5 Gbps/pin | Medium | Medium | Consumer / Workstation |
| LPDDR5X | 0.85 Gbps/pin | High | Low | Edge AI / Mobile |
| SRAM (on-chip) | 10+ Gbps/pin | Very Low | Very High | LPU / Cache |
HBM is the optimal balance of bandwidth and capacity. SRAM is fastest but capacity is too small (100× cost per GB vs HBM) , GDDR has large capacity but insufficient bandwidth.
Three-Vendor Landscape
1. SK Hynix: Current HBM Dominator (70% Share)
SK Hynix is the absolute leader in the HBM market:
| Item | SK Hynix Status |
|---|
| Market Share | ~70% (2025 total HBM market) |
| HBM4 Progress | First mass production Q1 2026, NVIDIA exclusive supply |
| HBM3e Supply | NVIDIA main supplier (80%), some AMD |
| Core Technology | Advanced MR-MUF (Mass Reflow Molded Underfill) |
| Capacity | 2026 plan: 25,000 HBM wafers/month |
| Key Customers | NVIDIA (90%), AMD, some Google |
| 2025 Revenue (HBM) | ~$30B (YoY +80%) |
| 2025 Net Margin | ~35% (far above traditional DRAM business) |
SK Hynix Key Advantages
- Earliest to mass-produce HBM3 (2018) : Technological first-mover
- Deep NVIDIA partnership: HBM3, HBM3e all launched with NVIDIA first
- Advanced MR-MUF process: Leading packaging yield
- HBM4 head start: First mass production Q1 2026
Key Events
| Time | Event |
|---|
| 2014 | Co-developed HBM with AMD |
| 2018 | First HBM2 mass production (NVIDIA V100) |
| 2020 | HBM2e mass production |
| 2022 | HBM3 mass production (NVIDIA H100) |
| 2024 | HBM3e 12-Hi mass production (NVIDIA B200) |
| 2026 Q1 | HBM4 first mass production (NVIDIA Rubin R200) |
2. Samsung: HBM4 Chaser (20% Share)
Samsung is the HBM market challenger, but hampered by yield and NVIDIA certification:
| Item | Samsung Status |
|---|
| Market Share | ~20% (2025 total HBM market) |
| HBM4 Progress | Q2 2026 mass production (1 quarter behind SK Hynix) |
| HBM3e Supply | Awaiting NVIDIA certification, primarily for AMD / Google |
| Core Technology | TC-NCF (Thermal Compression Non-Conductive Film) |
| Capacity | 2026 plan: 10,000 HBM wafers/month |
| Key Customers | AMD (some MI300X), Google TPU, Cerebras |
| 2025 Revenue (HBM) | ~$8B (YoY +150%, but only 1/4 of SK Hynix) |
| 2025 Net Margin | ~5% (low yield → low margins) |
Samsung Key Issues
- HBM3 NVIDIA certification not passed: Multiple sample submissions failed 2023-2024
- Low yield: HBM3e yield ~50% (SK Hynix ~70%)
- Technology path divergence: Samsung bets on TC-NCF (vs SK Hynix's MR-MUF)
- 2024 significant losses: HBM business massive investment but slow returns
Samsung 2025-2026 Breakthroughs
| Time | Event |
|---|
| 2024-12 | HBM3e 8-Hi passed AMD certification |
| 2025-03 | HBM3e 12-Hi passed Google TPU certification |
| 2025-Q4 | HBM4 trial production, Q2 2026 mass production |
| 2025-Q4 | NVIDIA HBM3e 12-Hi certification passed (partial orders) |
Late 2025, Samsung finally obtained partial NVIDIA HBM3e 12-Hi orders, a turning point for Samsung's HBM business.
3. Micron: HBM Dark Horse (10% Share)
Micron is the HBM market dark horse, becoming NVIDIA's second supplier with HBM3E 12-Hi 9.2 Gbps speed:
| Item | Micron Status |
|---|
| Market Share | ~10% (2025 total HBM market) |
| HBM4 Progress | Q3 2026 mass production (2 quarters behind SK Hynix) |
| HBM3e Supply | NVIDIA second supplier (~30% share) |
| Core Technology | 1znm DRAM + Advanced Packaging |
| Capacity | 2026 plan: 5,000 HBM wafers/month |
| Key Customers | NVIDIA (partial), AMD, Intel |
| 2025 Revenue (HBM) | ~$4B (YoY +200%) |
| 2025 Net Margin | ~10% |
Micron Key Advantages
- HBM3E 12-Hi 9.2 Gbps: Industry's highest speed (ahead of SK Hynix's 9 Gbps)
- US-based manufacturing: Idaho/New York fabs, CHIPS Act compliant
- NVIDIA second supplier: Avoiding single-source risk
- 2025 breakthrough: Revenue +200% YoY, fastest growth among the three
Micron Key Events
| Time | Event |
|---|
| 2024 | First HBM3E 8-Hi mass production |
| 2025-Q1 | HBM3E 12-Hi mass production (industry first) |
| 2025-Q2 | NVIDIA H100/B200 certification passed |
| 2025-Q3 | Partial B200 orders (~30% share) |
| 2026-Q3 | HBM4 expected mass production |
Micron is the fastest growing among the three, +200% YoY in 2025. But capacity of only 5,000 wafers/month limits its market share growth.
HBM Technology Roadmap
HBM4 Key Specifications
| Item | HBM3e | HBM4 | Improvement |
|---|
| Per stack capacity | 24 GB (12-Hi) | 36-48 GB (16-Hi) | 1.5-2× |
| Per pin speed | 9.2 Gbps | 12-16 Gbps | 1.3-1.7× |
| Per stack bandwidth | 1.2 TB/s | 1.5-2 TB/s | 1.3-1.7× |
| Power | 7W/stack | 8W/stack | Slight increase |
| Manufacturing process | 1z/1β DRAM | 1γ/1δ DRAM | Shrunk |
| Packaging | CoWoS-S | CoWoS-L | Larger interposer |
| Mass production | 2024-2025 | 2026 Q1 (SK) / Q2 (Sam) / Q3 (Mic) | — |
HBM4E / HBM5 Roadmap
| Year | Model | Capacity | Speed | Process |
|---|
| 2026 | HBM4 | 36-48 GB | 12-16 Gbps | 1γ/1δ |
| 2027 | HBM4E | 48-64 GB | 16-20 Gbps | 1δ |
| 2028 | HBM5 | 64-80 GB | 20-24 Gbps | 1ε |
| 2029 | HBM5E | 80-96 GB | 24-28 Gbps | 1ε+ |
Capacity and Pricing
HBM Capacity (2026 Plan)
| Vendor | 2025 Actual | 2026 Plan | 2027 Plan | Market Share |
|---|
| SK Hynix | 18,000 wafers/month | 25,000 | 35,000 | 60-70% |
| Samsung | 6,000 wafers/month | 10,000 | 18,000 | 15-20% |
| Micron | 3,500 wafers/month | 5,000 | 12,000 | 10-15% |
| Total | 27,500 | 40,000 | 65,000 | 100% |
2026 HBM capacity tight: Demand ~50,000 wafers/month, supply only 40,000 wafers/month, 20% shortfall.
HBM Pricing (per GB)
| Model | 2024 | 2025 | 2026 | 2027 (E) |
|---|
| HBM4 | N/A | N/A | $20-25/GB | $15-18/GB |
| HBM3e 12-Hi | $18-22/GB | $15-18/GB | $12-15/GB | $10-12/GB |
| HBM3e 8-Hi | $14-18/GB | $11-14/GB | $9-11/GB | $8-10/GB |
| HBM3 8-Hi | $10-12/GB | $8-10/GB | $6-8/GB | $5-7/GB |
HBM accounts for 30-50% of AI chip cost. NVIDIA B200's HBM cost is approximately $5,000-8,000 (192GB HBM3e × ~$30/GB).
HBM and AI Chip Supply Correlation
NVIDIA Rubin R200 HBM Supply Chain
| HBM Source | Share | Notes |
|---|
| SK Hynix HBM4 | 70% | First batch, exclusive supply Q1-Q2 2026 |
| Micron HBM4 | 20% | From Q3 2026, CHIPS Act benefits |
| Samsung HBM4 | 10% | From Q4 2026 (certification delay) |
NVIDIA still heavily dependent on SK Hynix. This is a single point of failure risk for NVIDIA's supply chain.
AMD MI400 HBM Supply Chain
| HBM Source | Share | Notes |
|---|
| SK Hynix HBM3e | 50% | Main supplier |
| Samsung HBM3e | 30% | Expanded after 2025 breakthrough |
| Micron HBM3e | 20% | Backup supply |
Google TPU HBM Supply Chain
| HBM Source | Share | Notes |
|---|
| Samsung HBM3e | 60% | Early partnership |
| SK Hynix HBM3e | 30% | Partial orders |
| Micron HBM3e | 10% | New entrant |
Huawei Ascend 920 HBM Supply Chain
| HBM Source | Share | Notes |
|---|
| CXMT (ChangXin Memory Technologies) | 70% | Domestic HBM, 4 Tbps |
| Samsung HBM3 | 20% | Restricted by US export controls |
| SK Hynix | 10% | Restricted by US export controls |
Huawei affected by US export controls, forced to accelerate domestic CXMT HBM substitution.
Investment Analysis
| Vendor | 2025 Revenue (HBM) | YoY | 2026 (E) Revenue | Net Margin |
|---|
| SK Hynix | ~$30B | +80% | ~$50B | ~35% |
| Samsung | ~$8B | +150% | ~$15B | ~10% |
| Micron | ~$4B | +200% | ~$10B | ~15% |
| Vendor | 2025 Increase | 2026 Increase (YTD) |
|---|
| SK Hynix | +120% | +35% |
| Samsung | +15% | +10% |
| Micron | +90% | +25% |
| NVIDIA | +180% | +40% |
SK Hynix is the biggest beneficiary of AI memory, +120% in 2025, far ahead of Samsung (only +15%).
Risks and Challenges
1. HBM Capacity Tightness
- 2026 20% shortfall (demand 50K wafers vs supply 40K)
- NVIDIA Rubin R200 shipments may be delayed due to HBM supply constraints
- Customer prepayment deposits have become the norm
2. US Export Controls
- HBM exports to China strictly limited by US Commerce Department
- Samsung, SK Hynix China-based fabs restricted
- Huawei accelerating domestic CXMT HBM substitution
3. Technology Path Divergence
- SK Hynix: MR-MUF path, leading
- Samsung: TC-NCF path, behind but catching up
- Micron: Somewhere in between
4. Competitive Technologies
- Samsung HBM-PIM: HBM with built-in processing units (processing-in-memory)
- TSMC SoIC: 3D stacked SRAM + Logic
- Micron HBM-CX: Compute Express Link integration
Future Outlook
Short-term (2026-2027)
- HBM remains tight: Demand > supply
- Prices remain high: HBM4 $20-25/GB
- Three coexist: SK Hynix 70% + Samsung 20% + Micron 10%
Mid-term (2027-2029)
- HBM4E / HBM5: Capacity 64-96 GB, speed 20-28 Gbps
- Domestic HBM rises: CXMT mass produces 8-Hi
- New entrants: YMTC may enter HBM market
Long-term (2029+)
- HBM6 / 3D HBM: Stack more layers, TSV alternatives
- PIM-HBM: HBM with built-in processing units
- Alternative technologies: On-chip SRAM capacity breakthroughs (e.g., Cerebras WSE)
Detailed Product Pages
Summary
The HBM three-way battle will continue for 3-5 years:
- SK Hynix unshakeable in the short term — HBM4 debut + deep NVIDIA lock-in
- Samsung breakthrough late 2025 — HBM3e 12-Hi finally NVIDIA-certified
- Micron fastest growing — HBM3E 9.2 Gbps industry fastest
- Capacity constraints persist — 20% shortfall in 2026
- Domestic HBM rises — CXMT 4 Tbps breakthrough
HBM is not a supporting actor — it's the "electricity, water, and gas" of the AI era.