HBM / Memory
HBM/DRAM bandwidth beside accelerators, plus enterprise NAND/HDD capacity. US-listed tickers or ADRs only.
Supply-chain view
How data moves through the memory layer
Materials & wafer fabs
Silicon, process materials and fabrication capacity
HBM · DRAM · NAND · HDD
Bandwidth close to accelerators plus capacity storage
Accelerators & AI servers
Memory feeds compute and retains working datasets
Training · inference · retention
Model execution and the storage behind it
Icons mark layer roles — not company logos. Position only; not a price chart.
Why this layer matters
Larger models need more bandwidth, not just more FLOPs, so HBM demand scales with compute rather than trailing it. The chain runs accelerator → HBM → DRAM → wafer → equipment. Separately, training data and checkpoints still need NAND and HDD capacity behind the compute tier — that is a different constraint than HBM scarcity.
What is the bottleneck?
HBM capacity is committed far in advance and the qualified HBM supplier list is short. That bandwidth constraint is not the same as enterprise NAND/HDD capacity. This layer lists US-listed tickers or ADRs only — major HBM makers without a US listing are omitted here even when they matter to the stack. Equipment and packaging for HBM sit in other layers.
A structural observation about supplier concentration and lead times — not a view on any security.
Is HBM bandwidth shortage the same as HDD/SSD capacity?
Short answer: no. Open the answer page for the split, why SK hynix / Samsung are omitted under the shared coverage rule, and links to equipment / foundry.
Read the answerWho sits in this layer?
HBM / DRAM (bandwidth)
Company
MUHBMDRAMNANDMicron Technology
Role
Supplies high-bandwidth and conventional memory used beside accelerators and inside servers.
Evidence
HBM and DRAM supplier for AI accelerators.
AI demand path
Larger accelerator clusters increase demand for memory bandwidth and capacity.
NAND / SSD (fast capacity)
Company
SNDKNANDFlash storageEnterprise SSDSandisk
Role
Supplies enterprise solid-state drives (eSSDs) for data-center workloads.
Evidence
NAND flash memory supplier.
AI demand path
More AI data and checkpoints increase the need for fast, persistent storage.
HDD (mass capacity)
Company
WDCEnterprise HDDCapacity storageWestern Digital
Role
Supplies high-capacity hard drives used for datacenter datasets and retained model data.
Evidence
Enterprise hard drives for datacenter capacity storage.
AI demand path
Dataset growth raises demand for economical capacity behind the compute tier.
Company
STXEnterprise HDDMass-capacity storageSeagate Technology
Role
Supplies mass-capacity hard drives for datacenter and enterprise data retention.
Evidence
Enterprise hard drives for datacenter capacity storage.
AI demand path
AI datasets and generated content add long-lived data that still needs capacity storage.
The other side
Bandwidth demand does not automatically become memory scarcity. HBM comes from a small number of DRAM makers, and the capacity they add can catch up with demand; memory has a long history of cycles in which new supply overshoots. Designs that need less memory per unit of compute — better compression, on-chip SRAM, smaller models for inference — would also weaken the link.
What could invalidate this picture?
The picture weakens if HBM lead times shorten and capacity no longer has to be committed far in advance, if accelerator shipments stop being limited by memory, or if new architectures cut the memory needed per unit of compute. It strengthens if lead times stay long while model sizes and context lengths keep growing.
Coverage boundary
Coverage boundary: US-listed tickers or ADRs only (same rule as TSM / ASML). Important suppliers without a US listing are omitted even when they matter to the stack. That is a coverage rule, not a quality ranking.
- SK hynix — Major HBM supplier; no US primary listing in this map's scope.
- Samsung Electronics — Major memory maker; no US primary listing in this map's scope.
Checked 2026-10-09. Not exhaustive.
Same track
Descriptive positioning only. The positioning label says whether a company's business is exposed to a layer, never whether to buy it. Layers with no exposure are not recommendations to add any.