- The author is bullish on memory as an agent-inference bottleneck: decode depends on memory bandwidth, while capacity determines context length and concurrent users.
- Caching can sharply reduce inference costs—DeepSeek Flash charges $0.006 per million cached input tokens versus $0.30 uncached—supporting demand for SSD and NAND-based context storage.
- AI adoption and multi-agent use could multiply memory demand; AWS reportedly has capacity blocks sold out for 2026 and largely for 2027, while Micron has contracts through 2031 with take-or-pay commitments.
- The author sees rising earnings and falling forward P/Es as evidence the memory thesis is working, and is bullish on NAND suppliers such as SanDisk, but supplies no explicit ticker.
Investment pitches
filter:
74