The Memory Shortage Runs Long
−3.7%
since Oct 16
The idea
I believe AI is going to end up everywhere, running on models that are hungry for memory and storage, and that the current shortage holds through 2030 or longer because new factories take years to build while demand keeps outrunning supply. Most of the basket sits in the companies making the scarce products: memory chipmakers, hard drive makers, and flash storage, with the equipment and testing suppliers who profit as those makers expand as steadier support. I hold a slice in cash and avoid putting everything on one name.
What it owns
- These companies make the AI memory chips in the shortest supply, so they earn the most while prices stay high. · Micron TechnologyIt makes the stacked memory that sits next to AI processors, including HBM, the most constrained type in the AI supply chain., SK HynixIt is the biggest supplier of HBM, the memory every large AI chip requires, and leads the market while HBM stays sold out.26%
- Memory makers pour their profits into new factories, and these companies sell the machines those factories need. · Lam ResearchIt makes the etching machines memory makers use to carve their chips, and gets paid when shortage profits flow into new factory capacity., Applied MaterialsIt sells a wide range of machines every chip factory needs, and gets orders as memory makers build out new capacity., ASML HoldingIt makes the only machines advanced enough to print chip patterns in new memory factories, so every memory capacity expansion requires its equipment., KLA CorporationIt makes the inspection tools that catch defects as chips are made, and stacked memory like HBM requires more checking than older chips.21%
- With flash so expensive, AI and cloud buyers are taking every big hard drive they can get. · Western DigitalIt makes the big hard drives that AI data centers are buying in volume as an alternative to scarce flash storage., Seagate TechnologyIt makes the highest-capacity hard drives, the lowest-cost way to store the massive volumes of data AI systems generate and access.20%
- AI memory chips need special testing and connection chips, a smaller niche that grows with every unit shipped. · RambusIt supplies the interface chips that move data between memory and processors on server modules, earning more as AI server memory loads increase., FormFactorIt makes probe cards that test memory chips before they are stacked into HBM, so higher HBM output means higher card volumes., Onto InnovationIt makes inspection tools for advanced packaging, the step where HBM gets stacked and attached to AI chips in volume., CamtekIt makes inspection systems used heavily in HBM production to catch flaws in stacked memory as output climbs.13%
- Flash chips are the other scarce half of memory, and their prices have climbed quarter after quarter. · SanDiskIt makes only flash memory, the fast storage that holds the huge datasets AI models train on and pull from during inference., Silicon Motion TechnologyIt makes the controller chips that run flash drives, earning fees on every drive regardless of whose flash memory is inside.12%
- Cash8%