Tweet

Posted June 2, 2025. Captured 2026-04-24 as a retroactive add.

Thesis

Long-form AI video generation (15-min to 1-hour outputs) is a storage/memory forcing function. Each generated video is gigabytes; training requires petabytes of video data. When video generation matures from clips to long-form, memory and storage demand go parabolic.

Tickers implied

• SSD / NAND: SNDK (Sandisk), MU (Micron HBM + NAND), Samsung (private)
• HDD: STX (Seagate), WDC (Western Digital)
• HBM / DRAM: MU, Samsung, SK Hynix

Cross-Reference

• memory.json watchlist already tracks MU, STX, WDC
• SNDK is in monster-discoveries and mentioned by @ParadisLabs and @aleabitoreddit
• Converges with Paradis/aleabit thesis that storage is the underweighted AI supply chain layer

Why Track Zephyr

Author previously surfaced the memory long-thesis ahead of the mid-2025 run. Posts are terse — one-line thesis plus chart. Signal-to-noise is high for infrastructure-level takes. Worth watching for forward calls on compute/storage cycle transitions.

Integration

• Consider adding SNDK to memory.json watchlist (currently only MU/STX/WDC)
• Pair with @aleabitoreddit and @ParadisLabs takes on storage to build a fuller memory-supercycle perspective