AI Physical Stack Watchlist
Purpose
A map of physical AI infrastructure layers to watch for bottlenecks. This page is seeded from a pasted article and should be treated as unverified until each layer is checked against primary sources.
Layers from the source
| Layer / station | Why it matters in the source's framing | Example names mentioned |
|---|---|---|
| InP substrates | Substrate under lasers used in optical transceivers | AXTI |
| Laser tools | Equipment needed to create photonics/laser throughput | VECO |
| Photonics / CPO | Optical links connecting compute infrastructure | LITE, COHR, MXL, AAOI |
| Memory | High-bandwidth and other memory as an AI throughput constraint | MU, SNDK, SK Hynix |
| HBM / advanced packaging | Turns memory dies into usable bandwidth near accelerators | TSM, AMKR, ASE, KYEC |
| Inspection / metrology | Detects defects in advanced packages, bumps, bonds, substrates | ONTO, CAMT |
| Bonding / molding / attachment | Manufacturing steps that make HBM and advanced packages physical objects | BESIY, KLIC, TOWA |
| Retimers / connectivity | Signal integrity and connectivity fabric | ALAB, CRDO |
| Thermal / cooling | Enables higher rack density and package/system performance | Not specified in source |
| Power | Grid, generation, distribution, and datacenter power chain | Not specified in source |
| Construction / neoclouds | Builds and operates AI compute capacity | Not specified in source |
| Rare earths / connectors / storage | Other physical constraints in the bill of materials | Not specified in source |
Next diligence questions
- Which layer has the longest qualification cycle or least elastic capacity?
- Which companies have pricing power versus just volume exposure?
- Where are returns already discounting a perfect bottleneck thesis?
- Are the bottlenecks cyclical shortages, secular constraints, or temporary supply-chain tightness?
- How much of each name's revenue is actually tied to the bottleneck layer?
- What are the second-order beneficiaries if the current bottleneck gets solved?
Related pages
Dashboard follow-up
The dashboard provides the concrete 96-name universe and confirms the main physical-stack categories used in this watchlist. See ai bottlenecks dashboard for observed counts and data-quality flags.
Glass substrate follow-up
PhotonCap adds a more specific substrate branch to the physical-stack map: organic ABF to glass substrate transition, TGV formation, metallization, RDL lithography, bonding, and inspection. See glass substrate cycle.
Local source refs
raw/articles/ai-bottleneck-beta-dashboard-2026-05-11.mdraw/articles/ai-bottlenecks-dashboard-snapshot-2026-05-11.mdraw/articles/photoncap-glass-substrate-cycle-2026-05-08.md