Appearance
HBM
HBM은 AI GPU와 고성능 가속기에서 중요한 고대역폭 메모리입니다. HBM 수요는 일반 DRAM보다 AI 서버/GPU 출하량과 GPU당 HBM 탑재량에 훨씬 직접적으로 연결됩니다.
Micron은 HBM portfolio를 next-gen AI systems, professional visualization workstations, high-performance computing을 가속하기 위한 제품군으로 설명하며, HBM4 36GB 12H, 11Gb/s 이상 pin speed, 2.8TB/s 이상 bandwidth 같은 product-level capacity/bandwidth 지표를 제시합니다.^[micron.com — High-bandwidth memory (HBM)]
Samsung Semiconductor도 HBM을 DRAM 제품군 안에서 다루며, HBM 세대와 capacity/bandwidth/product configuration을 주요 product attribute로 제시합니다.^[semiconductor.samsung.com — HBM | DRAM | Samsung Semiconductor Global]
Forecast thesis
HBM forecast의 핵심은 “AI accelerator 수요가 크다”가 아니라, accelerator unit × HBM content × generation mix × realization gate가 동시에 맞아야 실제 bit demand로 실현된다는 점이다. [[ai-infrastructure-memory]]는 수요 상단을, [[memory-supply-chain-equipment]]는 공급 실현 gate를, [[channel-pricing-signal]]은 shortage가 가격·allocation으로 전이되는지를 검증한다. ^[raw/articles/aisilicon-nvidia-amd-hbm-capacity-roadmap-2026.md] ^[raw/datasets/ai-rack-gpu-hbm-coupling-parameters.md] ^[raw/reports/pkg-cowos-allocation-nvidia-amd-2026.md]
| Forecast axis | What to measure | Strong raw anchors | Decision use |
|---|---|---|---|
| Unit demand | AI server/accelerator units, rack shipments, custom ASIC adoption | AI server +28% YoY, GB200/NVL72 shipment notes, hyperscaler capex | top-down demand range; not sufficient alone |
| Content uplift | HBM GB per accelerator and stack capacity | B200 192GB, GB300/Rubin-class 288GB, HBM4 36GB/48GB stack roadmap | largest sensitivity when generation mix changes |
| Generation mix | HBM3e vs HBM4, Blackwell/Blackwell Ultra/Rubin timing | Blackwell share, Rubin delay/qualification, HBM4 supplier certification | scenario split and confidence adjustment |
| Realization gate | CoWoS/package, memory test, power/cooling, customer timing | CoWoS allocation, 120kW/rack, grid interconnection, HBM wafer multiple | convert announced demand into installed/shipped HBM bits |
| Price/allocation signal | contract price, LTA, shortage severity | DRAM/server allocation, HBM fully booked commentary | validate whether unmet demand is real or order amplification |
What not to infer
- GPU count ≠ realized HBM demand: announced clusters must pass power, package, rack integration, and customer timing gates.
- HBM revenue growth ≠ bit growth: ASP/mix can inflate revenue even when bit growth decelerates.
- +70% YoY HBM growth deceleration ≠ weak demand: after +200% and +130% years, lower growth can still imply a much larger absolute bit base.
- NVIDIA-only view undercounts HBM: TPU/Trainium/custom ASIC and HPC/networking HBM buckets should be tracked separately.
- Capex is not immediate supply: HBM4 supply depends on wafer allocation, packaging/test, yield, and qualification timing.
기본 수요 구조
text
Realized HBM demand EB
= Σ(accelerator units × HBM GB/accelerator × attach rate × realization factor) / 1e9
realization factor = min(power-ready factor, CoWoS/package factor, HBM supply factor, customer timing factor)기존 단순 구조는 accelerator units × HBM GB/accelerator였지만, 344개 raw integration pass 이후에는 [[ai-infrastructure-memory]]의 변수표를 기준으로 unit / content / attach / realization을 분리해서 다룹니다.
더 세분화하면:
text
총 HBM 수요 = Σ(가속기 세대별 출하량 × 세대별 HBM 탑재량 × attach rate)[[hbm-demand]] simulator는 이 구조를 첫 단계로 단순화해 다음 입력을 직접 조작하게 합니다.
- AI accelerator 출하량
- HBM attach rate
- Accelerator당 HBM stack 수
- Stack당 HBM capacity
2026 source-backed signals
| Signal layer | Raw evidence | Model interpretation |
|---|---|---|
| Growth context | raw/datasets/hbm-bit-demand-growth-rate.md, TrendForce HBM forecast | 2024 +200%, 2025 +130%, 2026 +70% YoY는 simulator output sanity check로 사용한다. |
| Content step-up | raw/datasets/gpu-hbm-capacity.md, raw/datasets/ai-rack-gpu-hbm-coupling-parameters.md | B200/GB200 192GB → GB300/Rubin 288GB로 GPU당 HBM content가 +50% 상승한다. |
| Rack conversion | raw/datasets/ai-rack-gpu-hbm-coupling-parameters.md, raw/articles/oem-server-odm-gb200-nvl72-hbm-shipments-2025.md | GB200 NVL72는 72 GPU, 약 13.4TB HBM, 약 120kW/rack이다. GB300 NVL72는 약 20.7TB HBM/rack이다. |
| Custom accelerator | raw/articles/aisilicon-tpu-ironwood-trainium3-hbm-specs-2026.md | TPU/Trainium/ASIC도 HBM bucket에 포함해야 하며 NVIDIA-only unit count는 하방 편향이다. |
| Realization constraint | raw/reports/pkg-cowos-allocation-nvidia-amd-2026.md, raw/articles/power-us-grid-interconnection-queue-bottleneck.md | CoWoS/package와 power-ready capacity가 announced GPU demand를 installed HBM demand로 바꾸는 gate다. |
- TrendForce는 HBM 수요가 2025년 +130% YoY에서 2026년 +70% YoY로 둔화되지만, HBM3e mainstream과 HBM4 ramp가 여전히 핵심 소비 driver라고 설명합니다.^[TrendForce — TrendForce — 2026 HBM Demand +70% YoY; SK hynix 50% / Samsung 28% / Micron 22% Bit…]
- 2026 HBM bit output share는 SK hynix 50%, Samsung 28%, Micron 22%로 이동할 수 있으며, 이는 supplier capa allocation과 pricing power를 추적하는 데 중요합니다.^[TrendForce — TrendForce — 2026 HBM Demand +70% YoY; SK hynix 50% / Samsung 28% / Micron 22% Bit…]
- CoWoS allocation 자료는 2026년 global CoWoS wafer demand 약 1M units, NVIDIA 595k wafers(약 60%)를 제시합니다. Advanced packaging capacity는 HBM demand realization의 물리적 병목입니다.^[Astute Group (citing Morgan Stanley) — Advanced Packaging Demand Soars: Nvidia Secures 60% of CoWoS Capacity]
- Datacenter power 자료는 announced AI data center capacity와 실제 착공 사이의 차이를 강조합니다. 전력/그리드 병목은 GPU/HBM 수요 timing을 지연시키는 top-down discount factor입니다.^[Data Center Frontier / IEA / Sightline Climate — The Gigawatt Bottleneck — Power Constraints Define AI Data Center Growth (2026)]
주요 driver
- AI GPU/accelerator 출하량
- GPU당 HBM stack 수와 stack당 capacity
- HBM3E/HBM4 등 세대별 mix
- CoWoS/advanced packaging capacity
- 패키징/수율/공급 제약
- AI cluster buildout와 hyperscaler capex
- Datacenter power/cooling/grid interconnection
Simulator implication
Micron HBM4 36GB 12H 같은 product-level capacity 정보는 content anchor로 사용할 수 있지만, accelerator 출하량, HBM generation mix, CoWoS allocation, data center power availability는 별도 market scenario로 분리해야 합니다.^[micron.com — High-bandwidth memory (HBM)]
관련 문서
- [[hbm-demand]]
- [[dram]]
- [[bit-demand]]
- [[supply-demand-gap]]
- [[scenario-analysis]]
- [[structural-demand-model]]
- [[raw-source-map]]
Evidence checklist
| Forecast question | Primary raw evidence | How to use it |
|---|---|---|
| HBM bit growth | TrendForce — TrendForce — 2026 HBM Demand +70% YoY; SK hynix 50% / Samsung 28% / Micron 22% Bit… | Use +70% YoY 2026 HBM demand as growth-rate context, not a direct simulator default. |
| Per-GPU HBM content | NVIDIA / wccftech / IntuitionLabs (종합) — NVIDIA GPU 세대별 per-GPU HBM 용량 | Map accelerator generation to GB/GPU before multiplying by shipments. |
| Package bottleneck | Astute Group (citing Morgan Stanley) — Advanced Packaging Demand Soars: Nvidia Secures 60% of CoWoS Capacity | Use CoWoS allocation as a realization constraint for GPU/HBM demand. |
| Datacenter physical constraint | Data Center Frontier / IEA / Sightline Climate — The Gigawatt Bottleneck — Power Constraints Define AI Data Center Growth (2026) | Treat power/grid/cooling as a cap on how much accelerator demand can become installed demand. |
Citation notes
HBM demand pages should cite both product-level sources and bottleneck sources. A GPU shipment assumption without CoWoS/power context can overstate realized HBM bit demand.