Midnight MarketsV1.0
⚠ Not financial adviceNot financial advice. Independent research, not investment recommendations. The author and website make no representation as to accuracy or completeness. Do your own due diligence before acting on anything you read here. Read the full disclaimer →
Choke 04 · Sector I

The Memory Stack — HBM & Memory

HBM · NAND/SSD · Memory Metrology · Process Materials

Modern AI is memory-starved. The bottleneck isn't compute anymore — it's feeding the compute.

Every NVIDIA GPU has a tower of memory chips glued right next to it . This is called HBM (high-bandwidth memory). Without enough HBM, the world's GPUs literally sit idle waiting for data to arrive. Three companies make HBM — Micron (US), Samsung, SK Hynix (both Korea). Around them sit a layer of specialty toolmakers (Camtek, Onto for inspection; Entegris for materials; TOWA for sealing) and storage companies (SanDisk for SSDs that buffer the data). Six names own this whole layer.

Why this is a chokepoint

Wall Street obsesses over GPU shipments. The actual rate-limiter on AI training is HBM volume, and HBM volume runs through these six companies.

6 names on the watchlist

3D metrology — inspecting every HBM stack

Camtek makes the inspection machines that measure every layer of a stacked memory tower (HBM) down to the nanometer. The source thesis is direct: 'tool of reference at every HBM/CoWoS player.' Every memory stack at every leading manufacturer goes through a Camtek tool.

Only Western HBM maker with U.S. capacity

Micron is the only American HBM maker (Samsung and SK Hynix are Korean). HBM is the towers of memory glued onto every GPU — and there's a structural shortage. Micron's HBM3E is qualified at NVIDIA. The anchor of the whole memory supercycle thesis.

3D imaging tool — sole-qualified at HBM leaders

Onto's 3D imaging tools are the only ones qualified at two of the three HBM makers . As HBM stacks get taller (HBM3 → HBM4 → HBM4E), inspection time goes up and Onto's revenue scales linearly. Direct beneficiary of the memory supercycle.

Highest-beta NAND/SSD play for AI

SanDisk makes NAND flash and enterprise SSDs — the medium-fast storage layer right next to the GPU. AI training generates massive intermediate data ('checkpoints') that gets dumped to SSDs constantly. Highest-beta way to play the AI storage cycle.

HBM compression molding — sleeper monopoly

Tiny Japanese company with a functional monopoly on the molding step that seals every HBM stack . Almost nobody outside the industry knows this name. Pure-play on HBM volume — every memory stack the industry ships pays a small toll here. OTC.

Silent picks-and-shovels for every fab

Entegris supplies the ultra-pure chemicals, gases, and filtration systems every chip fab needs — including the HBM lines. Boring, defensive, and impossible to replicate without years of qualification. Quiet beneficiary of the entire fab buildout.

Sector sources
  • Dylan Bristot, AI Bottlenecks · whatllm.org · May 2026