sk hynix hbm memory technology and supply chain
on this page
hbm market dynamics 2025
market leader: sk hynix controls 46-49% of hbm market1
supply crisis: 2025 production nearly sold out, premium pricing 1
technology edge: hbm3e 12-layer stacks, 1.18 tb/s bandwidth 2
nvidia dependency: primary customer for h200/gh200 gpus3
expansion: m15x fab construction, production q3 20264
competition: samsung struggling with yields, micron ramping 5
hbm constrains ai chip production. memory bandwidth determines performance limits for large language models and training workloads.
hbm technology fundamentals
architecture
hbm stacks dram dies vertically:6
tl;dr
generation | bandwidth | stack height | capacity | interface width |
---|---|---|---|---|
hbm2 | 256 gb/s | 4-8 layers | 4-8 gb | 1024-bit |
hbm2e | 461 gb/s | 8 layers | 16 gb | 1024-bit |
hbm3 | 819 gb/s | 8-12 layers | 24 gb | 1024-bit |
hbm3e | 1.18 tb/s | 8-12 layers | 36 gb | 1024-bit |
hbm4 (2026) | 1.6+ tb/s | 12-16 layers | 48 gb | 2048-bit |
through-silicon via (tsv)
hbm tsv specifications:
- tsv density: 4,000-8,000 vias per die
- via diameter: 5-10 μm
- pitch: 40-50 μm
- aspect ratio: 10:1 to 20:1
- electrical path: 50-100 μm through silicon
tsvs provide vertical interconnection.
thermal management
sk hynix’s proprietary mr-muf (mass reflow molded underfill):2
- reduces chip stacking pressure by 6%
- increases productivity 4x through reduced process time
- improves heat dissipation by 45% vs traditional underfill
- advanced mr-muf adds 10% additional thermal improvement
required for 700w+ accelerator thermal management.
sk hynix market position
competitive position
sk hynix position
early production: first to mass-produce hbm3e in 20242
nvidia partnership: exclusive supplier for h200 initial production
yield advantage: 20-30% better yields than samsung on hbm3e
capacity: largest dedicated hbm production capacity
roadmap: path to hbm4 and beyond
production capacity and expansion
tl;dr
facility | location | capacity | status | investment |
---|---|---|---|---|
m14 | icheon, korea | 30k wafers/month | operational | existing |
m15 | cheongju, korea | 20k wafers/month | operational | existing |
m15x | cheongju, korea | 50k wafers/month | q3 2026 | $15 billion |
m16 | icheon, korea | planned | 2027+ | $20 billion |
m15x: hbm-specific fab with euv capability.4
financial performance
hbm revenue and projections:
- 2024: $8-10 billion hbm revenue
- 2025: $12-15 billion projected (sold out)
- 2026: $18-22 billion with m15x ramp
- gross margin: 50-60% on hbm vs 20-30% on commodity dram
- price premium: 5-8x per gb vs standard dram
sk hynix ceo kwak noh-jung: “sold out of hbm production for 2024 and almost sold out for 2025”1
manufacturing ecosystem
equipment suppliers
hbm production equipment:7
equipment partners
towa corporation: resin compression molding (22 units to sk hynix fy2023)
disoc: wafer grinding to 30-40μm thickness
advantest: specialized hbm test systems
hanwha semitech: thermal compression bonding (tcb) machines
applied materials: etching and deposition for tsv formation
materials
hbm manufacturing materials:7
- ncf (non-conductive film): die attachment and insulation
- tim (thermal interface material): heat dissipation between layers
- mr-muf materials: sk hynix proprietary underfill
- photoresists: for tsv and redistribution layer patterning
- seed layers: copper/titanium for tsv metallization
japanese suppliers dominate specialized materials.
test and validation
hbm test requirements:
- known good die (kgd): pre-stacking test
- mid-bond testing: assembly validation
- final test: bandwidth and thermal
- burn-in: temperature aging
- system-level validation: gpu configuration test
testing: 15-20% of total cost.
competitive landscape
samsung memory
samsung hbm status:5
- yield issues: 10-20% lower yields than sk hynix on hbm3e
- thermal problems: inferior heat dissipation in nvidia qualification
- market share: declining from 40% to ~35%
- customer concentration: lost nvidia h200 to sk hynix exclusive
samsung investing in yield improvements.
micron technology
micron hbm position:8
- hbm3e production: ramping in late 2024/early 2025
- capacity: targeting 20% market share by 2026
- differentiation: focus on power efficiency
- customers: amd, potential intel discrete gpus
- us advantage: domestic supply for government contracts
chinese development
china hbm status:9
- cxmt (changxin memory): developing hbm2 equivalent
- technology gap: 3-4 years behind leaders
- equipment constraints: lack of advanced equipment access
- target market: domestic ai chips, import substitution
limited competitive impact through 2028.
supply chain dynamics
vertical integration
memory manufacturer expansion:
- sk hynix: investing in packaging equipment and materials
- samsung: developing in-house test capabilities
- micron: partnering with osats for turnkey solutions
increasing supply chain integration.
geographic concentration risks
tl;dr
risk factor | impact | mitigation |
---|---|---|
korea concentration | 80% of hbm production | micron us expansion |
equipment dependency | japan controls key tools | limited alternatives |
materials bottleneck | specialized chemistry | strategic inventory |
power disruption | fabs require stable power | backup generation |
geopolitical | korea-japan tensions | diplomatic engagement |
pricing and allocation
hbm pricing dynamics 2025:
- spot prices: 50-100% premium over 2023 levels
- contract pricing: annual agreements with 20-30% increases
- allocation priority: nvidia > amd > custom chips > others
- prepayment requirements: major customers paying 12+ months ahead
- minimum orders: 10,000+ units for direct supply
supply constraints through 2026.
technological roadmap
hbm4 (2026-2027)
hbm4 specifications:10
- bandwidth: 1.6-2.0 tb/s per stack
- capacity: 48-64 gb per stack
- interface: 2048-bit wide bus
- power efficiency: 30% improvement per gb/s
- new features: on-die ecc, ai-specific operations
sk hynix targeting q4 2026 for initial production.
packaging evolution
hbm packaging roadmap:
- hybrid bonding: direct cu-cu connections replacing micro-bumps
- silicon bridges: connecting multiple hbm stacks
- optical interconnects: photonic links for extreme bandwidth
- advanced cooling: integrated microfluidic channels
- chiplet integration: hbm as standardized chiplet
future memory technologies
- pim (processing in memory): computation within memory stack
- cxl-attached memory: disaggregated memory pools
- persistent memory: storage-class memory integration
- 3d xpoint successors: new non-volatile technologies
- quantum memory: cryogenic memory for quantum computers
ai workload optimization
workload requirements
tl;dr
workload | bandwidth need | capacity need | optimal config |
---|---|---|---|
llm training | critical | very high | 8-12 hbm3e stacks |
llm inference | high | moderate | 4-8 hbm3 stacks |
computer vision | moderate | low | 2-4 hbm2e stacks |
recommendation | very high | high | 8-12 hbm3e stacks |
scientific compute | variable | very high | application specific |
software optimization
hbm utilization techniques:
- data layout: optimizing for 1024-bit access patterns
- prefetching: hiding latency through prediction
- compression: reducing bandwidth requirements
- caching: intelligent use of on-chip sram
- scheduling: maximizing concurrent memory operations
software optimization: 2-3x effective bandwidth improvement.
market outlook
market projections
hbm market size:11
- 2024: $12 billion market
- 2025: $18 billion (50% growth)
- 2026: $28 billion (55% growth)
- 2028: $50 billion projection
- 2030: $80-100 billion potential
ai accelerators: 85%+ of demand.
semiconductor implications
- memory equal partner with logic in ai chips
- vertical integration between memory and gpu makers
- new business models around bandwidth-as-a-service
- shift from commodity to specialized memory products
- geographic diversification
capital requirements
hbm investment scale:
- new fab: $15-20 billion for advanced memory fab
- conversion cost: $2-3 billion to convert dram to hbm
- r&d spending: 15-20% of revenue vs 10% for commodity
- equipment: 30% more expensive than standard dram tools
- time to market: 3-4 years from greenfield to production
limited to major memory manufacturers.
references
[1] anandtech. (2025). sk hynix reports that 2025 hbm memory supply has nearly sold out.
[2] sk hynix news. (2025). sk hynix begins volume production of the world’s first 12-layer hbm3e.
[4] digitimes. (2025, may 2). sk hynix says 2025 hbm production almost sold out.
[5] tom’s hardware. (2025). hbm supply from sk hynix and micron sold out until late 2025.
[6] semianalysis. (2025, august 12). scaling the memory wall: the rise and roadmap of hbm.
[8] tweaktown. (2025). sk hynix, samsung, micron are expanding hbm production.
[9] techpowerup. (2025). sk hynix ceo says hbm from 2025 production almost sold out.
[10] sk hynix news. (2025). sk hynix develops industry’s first 12-layer hbm3.
[11] investing.com. (2025, august 21). broadcom raises ai revenue forecast on strong custom chip demand.