SanDisk's Surge: Three Catalysts in Resonance, NAND Becomes a "Necessity," AI Revalues Storage
```
In the past three weeks, the storage sector has experienced a rare "perfect storm."
SanDisk's stock price has surged by over 100%, and all NAND-related stocks have moved upward. On the surface, this appears to be a typical storage sector rebound; but if we break down the technology and demand changes since the beginning of the year, it's more like a value re-evaluation triggered by the evolution of AI architectures.
From Nvidia proposing a new inference storage architecture at CES, to DeepSeek releasing its Engram model, and ClaudeCode accelerating the landing of "stateful AI Agents," three originally scattered technological paths are all pointing to the same conclusion at the start of 2026:
Storage is shifting from a "cost item" to a "core production factor" of AI.
Jensen Huang lights the first fire: Context becomes the bottleneck, storage must be restructured
The runaway growth of AI inference scale is forcing a restructuring of computing systems.
At CES2026, Nvidia CEO Jensen Huang systematically proposed the concept of ICMS (Inference Context Memory Storage) for the first time, giving a clear assessment: Context is becoming AI's new bottleneck, not computing power itself.
As model context windows grow from hundreds of thousands of tokens to the terabyte level, KVCache and context memory's squeeze on HBM is becoming unsustainable. On one hand, the unit cost of HBM3e is far higher than NAND; on the other hand, CoWoS packaging capacity also places a hard constraint on HBM supply.
Nvidia's solution is not to "pile on more GPUs," but to offload context from HBM.
In its new DGXVeraRubinNVL72SuperPOD architecture, besides compute and network racks, Nvidia has introduced—for the first time—a dedicated inference context storage rack. These racks, connected to the computing system via BlueField DPU and Spectrum-X Ethernet, essentially serve as "working memory."
From demand estimation, this is not a marginal change:
Each SuperPOD adds about 9.6PB of NANDPer NVL72 compute rack, incremental NAND is about 1.2PBIf 100,000 NVL72 racks in SuperPOD form ship in 2027, that's 120EB of new NAND demand
In a global NAND market with annual demand of about 1.1–1.2ZB, this means nearly 10% of incremental structural demand. More importantly, this demand comes directly from AI infrastructure, not traditional consumer electronics.
DeepSeek Engram: NAND is used as "slow memory" for the first time
If Nvidia solves the engineering architecture problem, DeepSeek's Engram model elevates NAND at the algorithmic level.
Engram's core breakthrough lies in deterministic memory access. Unlike MoE or dense Transformers' dynamic routing, Engram can, before computation starts, accurately determine which memory fragments need to be accessed based on input tokens, enabling prefetching in advance.
In traditional models, only ultra-low latency memory like HBM can support uncertain access paths; but Engram's deterministic prefetch effectively "masks" the latency gap between SSD and HBM.
DeepSeek's paper has verified:
An embedding table of 100 billion parameters can be fully offloaded to host memoryPerformance loss is less than 3%As model size increases, 20–25% of parameters naturally lend themselves to "offloadable static memory"
What does this mean?
It means NAND is no longer just "cold data storage," but is systematically included in the tiered memory system for the first time as AI's "slow RAM"—dedicated to carrying massive, low-frequency but indispensable knowledge bases.
In terms of cost, NAND's unit price is still significantly lower than DDR and HBM; once it becomes "indispensable" in model architecture, its strategic value in data centers will be repriced.
Morgan Stanley analyst Shawn Kim and his team believe DeepSeek has demonstrated a "doing more with less" technology path. This hybrid architecture not only relieves the real-world constraint on high-end AI compute resources, but also proves to the global market that storage-compute synergy may be more cost-effective than simply scaling up compute.
ClaudeCode: AI moves from "stateless" to "stateful," storage demand multiplies
The third catalyst comes from the application layer.
The rise of ClaudeCode marks the evolution of AI from "dialog tools" to long-running Agents. Unlike one-off text generation, code-writing AI needs to:
Repeatedly read and modify filesMultiple rounds of debugging and backtrackingMaintain conversations and states lasting days
The essence of this type of AI is a "stateful system" with long-term working memory.
And this working memory obviously cannot reside long-term in expensive GPU HBM.
The BlueField DPU + NAND combo provides a cost-controlled solution: An Agent's session state and historical context can reside in the NAND tier, rather than occupying compute resources.
This means, as AI Agent penetration increases, the storage demand function will decouple from inference call volume and instead be tied to "state duration"—a brand new growth logic.
Why SanDisk? Why now?
The simultaneous landing of three technological paths at the start of 2026 leads to a compelling conclusion:
Nvidia creates new application scenarios for NAND at the hardware architecture layerDeepSeek validates the feasibility of NAND as "slow memory" at the model layerClaudeCode amplifies long-term storage's rigid demand at the application layer
This is not a short-term boost from a single customer or product, but a signal of AI architecture transformation.
Against this background, SanDisk's stock performance is no longer just a mapping of the "storage cycle rebound," but a sign the market is rethinking one question: In the age of AI, what is real infrastructure?
When NAND is simultaneously driven by cyclical recovery, long-term demand, and structural revaluation, its pricing logic will naturally leap forward. This may precisely be the real reason behind SanDisk's meteoric rise.
Risk warning and disclaimerThe market contains risks, and investment should be prudent. This article does not constitute personal investment advice, nor does it take into account individual users' special investment goals, financial situations, or needs. Users should consider whether any opinions, views, or conclusions in this article match their specific circumstances. Investment decisions based on this are at your own risk.
```