🚨 Citrini Research Latest Report: The AI Memory Revolution Replacing DRAM with Flash

AMD Acquires MEXT to Optimize Flash Memory Performance Nearing DRAM, Apple Promotes "LLM in a Flash" Edge Solution—Two Giants Moving Simultaneously, This Is Not a Coincidence, It Signals a Structural Shift in AI Memory Architecture.
The Only Core Data Point: Flash Memory Costs Only 1/55 of DRAM. *HBM Accounts for 25% of DRAM Capacity, "Memory Tax" Pressure Has Reached a Critical Point. Through Controller Optimization, NAND Stacking, and Cell Mode Adjustments, Flash Is Becoming a Genuine Alternative for Edge AI Capacity and Bandwidth. This Is Also the Theoretical Basis for the Recent Continuous Surge in Storage Stocks Represented by SanDisk ($SNDK ).
---
The essence of this logic is not "Flash replacing DRAM," but that the memory hierarchy for AI inference is being restructured.
Low-frequency KV cache, model weights, and edge data during inference can be shifted from expensive HBM/DRAM to NAND Flash/SSD layers. Flash is not replacing HBM but is absorbing overflow demands in capacity—AI storage needs are so large that a multi-layer architecture is required to handle it together.
---
Four levels of positive feedback transmission:
① NAND Manufacturers (Most Direct)
High-capacity NAND, enterprise SSDs, QLC NAND are the purest directions.
US Stocks: SNDK, WDC, MU, Kioxia
Logic Purity: SNDK / WDC / Kioxia > MU
② SSD Controllers (Most Durable)
What truly brings Flash close to memory experience are controllers, firmware, and NVMe architecture optimization—not just component price hikes.
US Stocks: SIMO, MRVL
Taiwan Stocks: Phison
③ CXL / PCIe High-Speed Interconnection
Data moves from Flash to computing units, and the pathway itself is valuable.
US Stocks: ALAB, RMBS, CRDO
④ Edge AI Storage Upgrades
Apple's LLM in a Flash: Model weights reside in iPhone/Mac flash, driving the terminal storage from 256GB to 1TB, boosting NAND demand.
Beneficiaries: MU, Samsung, Kioxia, Phison
---
A-share Mapping:
Most Direct → Jiangbolong, Baiwei Storage, Demingli
Core of Architecture Upgrade → Lianchuang Technology (CXL + Memory Interface)
Domestic Platform Mapping → GigaDevice, Beijing Junzheng
---
Two Conclusions:
Short-term Flexibility: SNDK, WDC, SIMO, Phison, Jiangbolong, Baiwei Storage
Medium-Long Term Certainty: MU, MRVL, ALAB, Lianchuang Technology