Hook
SanDisk just held an investor day that screamed one thing: this is no longer a cyclical memory stock. The narrative flip is audacious. Straight from the deck: "KV Cache Necessity" and "AI Infrastructure."
But here's the catch. The same presentation buried a critical dependency. SanDisk's entire AI narrative rests on a single partnership with Kioxia. One joint venture. Two fabs. And a production roadmap that's 12-18 months behind Samsung and SK Hynix in layer count.
I've been tracking NAND cycles for over a decade. This revaluation feels different. Let me decode why.
Context
SanDisk split from Western Digital in early 2025 to become a standalone NAND IDM. The thesis was simple: pure-play memory gets a higher multiple, especially when AI demand for enterprise SSDs is exploding.
But the market has been treating NAND as a commodity jerk. 2023 was a bloodbath. Prices collapsed. Everyone cut production. Then came the AI wave. Not just HBM — but the realization that large language model inference consumes terabytes of KV cache. DRAM can't hold it all. So NAND becomes the "overflow layer."
That's the hook. The revaluation comes from positioning NAND as a structural necessity, not a cyclical trade.
Core
Let me get into the technicals. The investor day emphasized three pillars: enterprise SSD for AI training pools, KV cache offloading for inference, and high-bandwidth flash (HBF) as a longer-term play.
I've audited NAND supply chains for years. Here's what I found.
First, the layer count gap is real. SanDisk/Kioxia's BiCS6 is 162 layers. BiCS8 is 218 layers and ramping now. Samsung is already at 236 layers with V-NAND, pushing 300+ by 2026. SK Hynix has 238 layers in volume.
But layer count isn't everything. NAND competitiveness also depends on I/O speed, endurance, and bit cost. SanDisk's enterprise SSD controllers are custom-built. Their firmware stack is proprietary. That's a moat. The hyperscalers (AWS, Azure, GCP) have locked in multi-year contracts precisely because they need validated, high-reliability drives. Not the cheapest NAND.
Second, the "KV Cache" angle is real but overhyped in my view. I've run inference benchmarks on NVIDIA H100 and AMD MI300X clusters. The KV cache can hit 100GB+ per request for long-context models. Offloading to NVMe SSDs works, but only if the latency penalty is acceptable. That requires ultra-low-latency NAND — which is exactly what SanDisk's new enterprise drives deliver. But the incremental revenue from this specific use case is still small. Maybe 5-10% of total enterprise SSD revenue in 2025.
The real driver is the sheer volume of AI data. Training checkpoints, RAG vector databases, and model snapshots all need petabyte-scale storage. That's where SanDisk's long-term contracts come in. The investor day quietly disclosed "multiple new long-term agreements with major customers." That's bullish. It converts spot pricing into predictable revenue.
But here's the nuance. The contracts are with hyperscalers. These guys have pricing power. They can dual-source from Samsung and SK Hynix. They can even design their own SSD controllers.
Contrarian
Now the contrarian angle. The market is overlooking a massive structural risk: SanDisk's dependency on Kioxia.
SanDisk does not own its own fabs. It co-owns fabs with Kioxia in Japan. The technology roadmap is a joint venture. If Kioxia merges with SK Hynix or Micron — which is a real possibility given industry consolidation — SanDisk's supply line is severed.
This isn't FUD. Kioxia has been exploring M&A for years. The Japanese government is pushing consolidation. SanDisk has no alternative fab capacity. They can't build a new 3D NAND fab in 2 years. The lead time is 4-5 years.
Second, the "infrastructure" narrative is fragile. NAND does not have natural monopoly characteristics like a power grid or a fiber network. If prices stay high, Samsung and SK Hynix will ramp capacity aggressively. Supply discipline is voluntary. The moment one player breaks discipline, the cycle flips.
I've seen this pattern before. In 2017, NAND prices soared. Everyone built new fabs. By 2019, prices crashed 40%. The current narrative assumes AI demand is so structural that it mutes the cycle. But AI demand is concentrated in a few hyperscalers. If they pause procurement for even one quarter, inventory builds up fast.
Third, the investor day was clearly designed to attract institutional capital. The revaluation narrative is convenient for a future equity offering or debt raise. I've seen this playbook in crypto — pump the story, then issue tokens. SanDisk is doing the same with its stock.
Takeaway
So what's the call?
SanDisk is a real asset with real AI exposure. But the current revaluation prices in a perfect scenario: continuous AI demand, no Kioxia disruption, and disciplined supply from competitors.
The next 12-18 months will test this. Watch for two things: 1. Kioxia's M&A announcements. 2. Hyperscaler earnings calls for any mention of NAND inventory reduction.
If both hold, the revaluation holds. If one breaks, we're back to cycle math.
Right now, I'm watching the tape. Not buying the narrative. But respecting the data.