Over the past three weeks, the storage sector has experienced a rare "perfect storm." SanDisk's stock price has surged over 100%, with NAND-related stocks collectively rising. On the surface, this appears to be a typical storage cycle rebound; however, a deeper analysis of technological and demand shifts since the beginning of the year reveals it more closely resembles a value re-rating triggered by the evolution of AI architecture. From NVIDIA's introduction of a novel inference storage architecture at CES, to DeepSeek's release of the Engram model, and ClaudeCode's acceleration of "stateful AI Agent" deployment, three previously disparate technological paths have converged by early 2026 to point towards the same conclusion:
Storage is transforming from a "cost item" into a "core factor of production" for AI.
Jensen Huang Ignites the First Spark: Context Becomes the Bottleneck, Storage Must Be Re-architected The runaway growth of AI inference scale is forcing a fundamental restructuring of computing systems. At CES 2026, NVIDIA CEO Jensen Huang systematically introduced the concept of ICMS (Inference Context Memory Storage) for the first time, offering a clear assessment: Context is becoming the new bottleneck for AI, rather than computing power itself. As model context windows expand from hundreds of thousands of tokens towards the terabyte level, the strain on HBM from KVCache and contextual memory has become unsustainable. On one hand, the unit cost of HBM3e is significantly higher than NAND; on the other, CoWoS packaging capacity imposes a hard constraint on HBM supply. NVIDIA's solution is not to "add more GPUs," but to offload context from HBM. In the newly announced DGX Vera Rubin NVL72 SuperPOD architecture, alongside computing and networking racks, NVIDIA introduced for the first time dedicated, independent storage racks specifically for inference context. These racks, connected to the computing fabric via BlueField DPUs and Spectrum-X Ethernet, essentially serve the role of "working memory." Demand calculations indicate this change is far from marginal:
Each SuperPOD adds approximately 9.6 PB of NAND capacity. Per NVL72 computing rack, the incremental NAND demand is about 1.2 PB. If 100,000 NVL72 racks ship in SuperPOD form by 2027, this corresponds to 120 EB of new NAND demand.
Within a global NAND market with annual demand of approximately 1.1–1.2 ZB, this represents nearly 10% of new structural demand. Crucially, this demand originates directly from AI infrastructure, not traditional consumer electronics. DeepSeek Engram: NAND is Used as "Slow Memory" for the First Time If NVIDIA is solving an engineering architecture problem, then DeepSeek's Engram model is legitimizing NAND at the algorithmic level. Engram's core breakthrough lies in deterministic memory access. Unlike the dynamic routing in MoE or dense Transformer models, Engram can precisely determine the required memory segments to access based on input tokens before computation begins, enabling prefetching in advance. In traditional models, only ultra-low latency memory like HBM can support uncertain access paths; however, Engram's deterministic prefetching mechanism effectively "masks" the latency gap between SSDs and HBM. DeepSeek's research paper has already validated:
A 100-billion parameter embedding table can be entirely offloaded to host memory. The performance penalty is less than 3%. As model scale increases, 20–25% of parameters are inherently suitable to become "offloadable static memory."
What does this signify? It signifies that NAND is no longer merely "cold data storage," but is being systematically integrated into a tiered memory architecture for the first time, serving as AI's "slow RAM," specifically tasked with holding vast, low-frequency, yet indispensable knowledge bases. From a cost perspective, NAND's unit price remains significantly lower than DDR and HBM; once it attains "irreplaceability" within the model architecture, its strategic value in data centers will be repriced. Morgan Stanley analyst Shawn Kim and his team believe DeepSeek demonstrates a "Doing More With Less" technological pathway. This hybrid architecture approach not only practically alleviates resource constraints on high-end AI computing power but also proves to the global market that efficient storage-compute co-design may offer better cost-effectiveness than simply scaling compute power. ClaudeCode: AI Transitions from "Stateless" to "Stateful," Amplifying Storage Demand Exponentially The third catalyst comes from the application layer. The explosive growth of ClaudeCode signals AI's evolution from a "conversational tool" towards long-running Agents. Unlike one-time text generation, a code-writing AI requires:
Repeatedly reading and modifying files. Multiple rounds of debugging and backtracking. Session states that persist for days.
The essence of such AI is a "stateful system" possessing long-term working memory. This type of working memory clearly cannot reside permanently in expensive GPU HBM. The combination of BlueField DPU and NAND恰好 provides a cost-effective solution: an Agent's session state and historical context can reside persistently in the NAND tier, rather than occupying valuable compute resources. This implies that as AI Agent penetration increases, the demand function for storage will decouple from the number of inference calls and instead become linked to "state duration"—a fundamentally new growth logic. Why SanDisk? Why Now? The simultaneous maturation of these three technological paths in early 2026 forms a highly compelling conclusion:
NVIDIA, at the hardware architecture level, has created a new application scenario for NAND. DeepSeek, at the model level, has validated the feasibility of using NAND as "slow memory." ClaudeCode, at the application layer, has amplified the rigid demand for long-term storage.
This is not a short-term boost from a single customer or product, but a signal of a structural shift in AI system architecture. Against this backdrop, SanDisk's stock performance is no longer merely a reflection of a "storage cycle rebound," but indicates the market is beginning to re-evaluate a fundamental question: In the AI era, what truly constitutes infrastructure? When NAND is simultaneously driven by cyclical recovery, long-term demand growth, and structural re-rating, its pricing logic is bound to undergo a significant leap. This, perhaps, is the true reason behind SanDisk's dramatic surge.
Comments