The market is finally asking the right question. Not whether SanDisk can survive the NAND cycle, but whether the cycle itself has been structurally broken. The answer is a qualified yes, but the qualification matters more than the thesis.
Over the past six months, institutional flows have rotated into SanDisk with a velocity normally reserved for crypto-native narratives. The price action is not a reflex of spot shortages. It is a recalibration of the asset class. NAND is no longer a commodity memory play. It is being revalued as AI infrastructure — a high-barrier, long-duration, quasi-utility asset.
Let me be clear: I have seen this pattern before. In 2020, when DeFi yields were being priced as sustainable returns, I argued they were liquidity subsidies. The market ignored the structural flaw until the correction. Today, SanDisk’s revaluation is built on a more solid foundation, but the risk of narrative overshoot is real. The difference is that the underlying demand driver — AI inference scaling — is grounded in silicon reality, not speculative tokenomics.
I will dissect this thesis through five lenses: technology, supply chain, capacity discipline, demand structure, and geopolitical risk. Each lens reveals a piece of the revaluation puzzle. Together, they form a coherent narrative that justifies a higher valuation multiple — but only if the market internalizes the structural shift, not just the price momentum.
Hook: The KV Cache Bottleneck
In March 2025, a major cloud provider disclosed that its large language model inference cluster was spending 40% of its memory budget on KV cache storage. The cache was overflowing DRAM capacity, forcing the system to spill to SSD. The latency penalty was 12 milliseconds per spill. That is an eternity in inference time.
This is not an edge case. It is the new normal. Every inference request generates a KV cache that grows linearly with sequence length. As models move to 1M+ token contexts, the cache becomes a memory monster. The standard solution is to offload cold cache layers to high-bandwidth NAND. This is not a nice-to-have. It is a necessity for economically viable inference.
SanDisk, through its joint venture with Kioxia, is the dominant supplier of enterprise SSDs that sit in this exact latency tier. The company’s investor day presentation explicitly called out "KV cache offload" as a demand vector. The market has finally connected the dots.
Context: The NAND Revaluation Framework
SanDisk is a pure-play NAND IDM. It designs, fabricates, and integrates NAND flash into SSDs. The company was spun off from Western Digital in 2024, inheriting a 50% stake in the Kioxia-WD joint venture. This gives it access to the BiCS 3D NAND technology roadmap, currently at 218 layers (BiCS8).
The NAND industry has historically been a textbook cyclical commodity: oversupply, price collapse, capacity cuts, recovery, repeat. The cycle length is 3–4 years. The last trough was 2023, when industry revenue fell 35%. The recovery began in 2024, driven by AI demand. What is different this time is the nature of the demand.
Previously, NAND growth was driven by mobile storage and PC SSDs — price-sensitive, upgrade-driven markets. Now, the marginal buyer is the hyperscale cloud operator, procuring enterprise SSDs in multi-billion-dollar long-term contracts. These buyers are not price-sensitive in the traditional sense. They are capacity-constrained. They need to store petabytes of training data, checkpoint files, and increasingly, inference cache.
This shift changes the demand elasticity. Price increases no longer trigger immediate demand destruction. They trigger budget reallocation within the cloud provider’s infrastructure spend. The demand curve becomes more inelastic.
From my experience auditing ICO tokenomics in 2017, I recognized a similar pattern: when a token’s utility is tied to a scalable computation (like gas), demand becomes less elastic. The difference is that NAND has actual physical production constraints, not just code-defined supply caps.
Core: The Structural Supply Discipline
SanDisk’s management has been remarkably disciplined on capacity. The joint venture with Kioxia has not announced any major new fab beyond the existing Yokkaichi and Kitatani facilities. CapEx as a percentage of revenue is running at 15–18%, well below the 25–30% peaks of the 2017–2018 cycle.
This is not accidental. The memory industry learned a painful lesson: overbuilding destroys value. The 2023 crash was brutal enough to force a structural shift in capital allocation. SanDisk is now prioritizing free cash flow over market share. This is exactly what the market wants to see.
Let me quantify this. The industry’s bit supply growth has decelerated from 35% per year in 2018 to approximately 15% in 2025. Meanwhile, AI-driven demand growth is running at 20–30% annually. The gap between supply and demand growth is positive for pricing. This is the foundation of the revaluation thesis.
But there is a nuance. The 3D NAND layer count race is not the only competitive dimension. Interface speed, power efficiency, and reliability matter more in the enterprise segment. SanDisk’s BiCS8 at 218 layers is competitive with Samsung’s 236-layer V-NAND and SK Hynix’s 238-layer product. The 12–18 month layer gap is not a competitive disadvantage because hyperscale buyers qualify multiple suppliers. The moat is in the system integration — the controller, the firmware, the validation.
SanDisk has a strong internal controller design team. Its enterprise SSDs use proprietary firmware that optimizes latency, endurance, and power consumption. This is not easily replicated. Chinese competitor YMTC has caught up in layer count (232 layers) but lacks the enterprise ecosystem validation. The barrier to entry is not just manufacturing; it is the decade-long relationships with cloud providers.
Based on my 2022 analysis of crypto derivatives hedging, I learned that structural supply constraints are only valuable if demand is non-discretionary. In NAND, the demand is becoming non-discretionary for AI inference. That is the key insight.
Contrarian: The Decoupling Trap
The consensus view is that NAND has decoupled from its commodity cycle. I am skeptical of the word "decoupled." Decoupling implies a permanent break. What we are seeing is a temporary shift in demand composition that may revert if AI capex slows.
The risk is real. Hyperscale cloud providers are spending heavily on AI infrastructure, but their return on investment is not yet proven. If AI model improvements plateau or enterprise adoption disappoints, the storage budget will be cut. NAND demand will not disappear, but the growth rate will decelerate. The supply discipline that underpins the revaluation could break if manufacturers panic and increase capacity.
There is a second risk: the Kioxia relationship. SanDisk’s manufacturing depends entirely on the joint venture. If Kioxia is acquired by SK Hynix or merges with Micron — both plausible scenarios — SanDisk could lose its fab access. The company would then need to build its own fabs, a multi-year, multi-billion-dollar endeavor. The valuation today does not price in this dependency risk.
A third risk is technological substitution. HBM is becoming more affordable. If HBM capacity expands enough to absorb KV cache entirely, the need for NAND offload diminishes. This is a long-term risk, but the market is pricing in a 5–10 year demand cycle. Technology cycles can turn faster than expected.
I saw this dynamic play out in the DeFi yield farming frenzy of 2020. Investors treated temporary liquidity mining yields as structural revenue. When the incentives stopped, the yields collapsed. The NAND revaluation is better founded, but it is not immune to the same pattern of narrative overshoot.
Takeaway: Positioning for the Next Cycle
The revaluation of SanDisk from cyclical memory to AI infrastructure is analytically sound but fragile. It depends on three assumptions: sustained AI capex growth, continued supply discipline, and no disruption to the Kioxia partnership.
If these assumptions hold, SanDisk deserves a valuation multiple closer to a semiconductor infrastructure company (15–20x forward earnings) than a memory commodity player (8–10x). The current multiple is around 12x, leaving room for re-rating.
If any assumption breaks, the correction will be violent. The market is pricing in a smooth transition. It never is.
Liquidity is the only truth in a vacuum of trust. In this market, trust is placed in the AI narrative. The liquidity is flowing into storage. For now, that flow is justified. But I will be watching the Kioxia headlines and the hyperscale capex guidance. Code does not lie, but incentives often do. The incentive for SanDisk management is to keep the narrative alive. The incentive for investors is to verify the structural shift before the next cycle arrives.
Yield without basis is just delayed liquidation. The basis here is real demand. But the delay is the cycle. Time will tell if the market has correctly priced the duration.
Stability is a feature, not a market condition. The NAND market is stable today. That is exactly when the seeds of the next downturn are sown. I have seen this pattern in every cycle, from 2017 ICOs to 2020 DeFi to 2022 Terra. The structural revaluation narrative is convincing. It is also the most dangerous type of narrative — because it sounds like a permanent truth.
I will continue to hold a position in SanDisk, but I am hedged. The macro environment is supportive, but the structural skepticism I built over 18 years of market analysis tells me to never fall in love with a thesis. The market will prove me right or wrong. Either way, I will learn.
Code does not lie. But markets often do.