The market is repricing SanDisk. Not because of a new NAND layer count. Not because of a surprise earnings beat. But because the narrative is shifting from a cyclical memory play to an AI infrastructure necessity. Over the past 90 days, the stock has ripped 40% while the broader NAND index barely moved. The divergence tells me one thing: institutional money is betting on a structural re-rating, not a cyclical upswing.
I’ve been through three NAND cycles. I’ve seen the boom-bust of 2017, the inventory glut of 2019, the COVID-driven demand spike of 2021, and the brutal collapse of 2023. Each time, the market treated NAND as a commodity—prices determined by supply discipline and demand from PC and smartphone replacements. But this time, the narrative is different. SanDisk is being sold as the “storage layer for AI inference.” The phrase “KV Cache Necessity” appears in every sell-side note. The question is: is this just repackaged hype, or is there genuine infrastructure-level demand?
Let’s dissect the data.
Context: The NAND Bloodbath and the Phoenix
NAND is a capital-intensive, cyclical business. The industry has a 3-4 year cycle. The last peak was in 2021, followed by a massive oversupply in 2022-2023. Prices collapsed by 60% from peak to trough. SanDisk’s parent company Western Digital saw its storage revenue drop 45% in fiscal 2023. Then came the spin-off in early 2025. SanDisk emerged as a pure-play NAND IDM, with a joint venture (JV) with Kioxia (formerly Toshiba Memory) for manufacturing. The JV operates two fabs in Japan: Yokkaichi and Kitakami. The technology node is BiCS6 at 162 layers, with BiCS8 (218 layers) ramping.
But here’s the catch: SanDisk is not the technology leader. Samsung is on 236-layer V-NAND, SK Hynix is on 238-layer, and Micron is on 232-layer. SanDisk/Kioxia has a 12-18 month lag in layer count. That’s not a crime—NAND competition is not just about layers. It’s about cost per bit, I/O speed, reliability, and system integration. SanDisk’s strength lies in enterprise SSDs (eSSDs) and controller firmware. They own the entire stack from NAND die to finished SSD. That’s their moat.
Core: The AI Story – Replacing the Commodity Narrative
The market’s current thesis is that AI inference will require massive amounts of NAND for KV cache offloading, model checkpointing, and RAG (Retrieval-Augmented Generation) vector databases. The argument goes: as LLMs scale to longer contexts (128K, 1M tokens), the KV cache becomes a memory bottleneck. HBM is too expensive for cold data. DRAM is too volatile. NAND-based SSDs with PCIe 5.0/6.0 interfaces can serve as a “high-capacity, low-power overflow layer.” This is not a theoretical concept. Major hyperscalers are already deploying tiered storage for inference: hot data in HBM, warm data in DRAM, cold data in NVMe SSDs. SanDisk is positioning its enterprise SSDs as the default cold layer.
Let’s quantify the demand. Hyperscaler capex on AI servers is projected to grow 30%+ in 2025-2026. Each AI server, on average, contains 8-16 TB of NAND. That’s roughly 4x the storage density of a traditional server. The total addressable market for enterprise SSDs in AI is estimated at $15-20 billion by 2027, up from $8 billion in 2024. SanDisk’s eSSD revenue, which I estimate at 35-45% of total revenue, is growing at 20%+ annually. But the real kicker is the shift from spot pricing to long-term agreements (LTAs). The article mentions “new long-term commercial agreements with major customers.” This is a structural change. Historically, NAND was bought on spot or short-term contracts. Now hyperscalers are locking in volume for 2-3 years at fixed prices. This provides revenue visibility and reduces the cyclicality of the business.
In my experience, this is the single most important factor for the re-rating. If SanDisk can convert 50% of its revenue from spot to LTAs, the earnings volatility drops significantly. The stock moves from a 10x P/E (cyclical) to a 20x P/E (infrastructure). That’s a 2x multiple expansion. The market is pricing that in.
But there’s a catch. The LTAs are not unconditional. They are negotiated with clauses that allow repricing if NAND costs drop. The hyperscalers have significant bargaining power—they can dual-source from Samsung, SK Hynix, or Micron. SanDisk’s product is not unique. The switching costs are low. So the LTAs provide visibility, but not necessarily pricing power. This is a nuance the market is ignoring.
Contrarian: The Risks Buried in the Narrative
Let’s talk about the elephant in the room: SanDisk’s dependence on Kioxia. The JV is the only source of NAND wafers for SanDisk. If Kioxia decides to merge with SK Hynix (as rumored) or change the JV terms, SanDisk is left without a factory. Building a new NAND fab takes 3-4 years and $10-20 billion. SanDisk’s current cash flow is insufficient to fund that. The JV agreement is the single largest risk in the thesis.
Second, the technology gap. SanDisk is 1-1.5 years behind Samsung and SK Hynix in layer count. While layers aren’t everything, they drive cost per bit. If Samsung starts shipping 300-layer NAND in 2026, SanDisk’s cost structure will be at a disadvantage. The company’s answer is to focus on high-value eSSDs where reliability and firmware matter more than cost. But that’s a niche. The hyperscalers are price-sensitive. They will push for cost reductions.
Third, the advanced packaging bottleneck. SanDisk does not own CoWoS or other advanced packaging capacity. For a future “high-bandwidth flash” (HBF) product, they would need to rely on TSMC or OSATs. This limits their ability to differentiate in the AI memory hierarchy. HBM is the gold standard for AI; NAND is a budget substitute. The market is pricing HBF as if it’s a sure thing, but it’s still in the concept stage.
Fourth, the counterparty risk. The article mentions “self-custody strategies” and “exchange solvency” in the persona’s background. In the crypto world, we know that trust is a variable. In the NAND world, the counterparty is the hyperscaler. If a major customer decides to drop SanDisk’s SSDs for a cheaper alternative, the LTAs will be renegotiated. The revenue visibility is not guaranteed.
Takeaway: The Trade vs. The Investment
The short-term momentum is supported by supply discipline and AI narratives. NAND prices are still rising. SanDisk’s Q2 2025 earnings (reporting in August) will likely show strong revenue growth and margin expansion. The stock could run another 20% in the next 3 months.
But the long-term investor must watch three things: (1) Kioxia’s independence, (2) the ramp of 218-layer with meaningful cost reduction, and (3) the conversion of LTAs into actual profit. If any of these break, the re-rating narrative collapses.
I’m not a buyer at these levels. The risk/reward is skewed to the downside. The market is pricing in a perfect outcome: AI demand continues to grow, supply stays disciplined, Kioxia stays in the JV, and SanDisk’s technology gap closes. That’s a lot of assumptions.
Liquidity vanishes. Lessons remain.
Data over drama.
Calculate. Execute. Repeat.