Decision brief / Cached
L4WORKER CACHEHBM / Memory
Recommended decision
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.
MUMicron TechnologySupplies high-bandwidth and conventional memory used beside accelerators and inside servers.
SNDKSandiskSupplies enterprise solid-state drives (eSSDs) for data-center workloads.
WDCWestern DigitalSupplies high-capacity hard drives used for datacenter datasets and retained model data.
STXSeagate TechnologySupplies mass-capacity hard drives for datacenter and enterprise data retention.
Why this path
- Larger accelerator clusters increase demand for memory bandwidth and capacity.
- More AI data and checkpoints increase the need for fast, persistent storage.
- Dataset growth raises demand for economical capacity behind the compute tier.
- AI datasets and generated content add long-lived data that still needs capacity storage.
Risks to monitor
- 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.
- Cached public research only. Not a personal recommendation. Research, not investment advice.
Action plan
- /ai-map/memory
- NO TRADE EXECUTED