SanDisk's spin-off from Western Digital dropped like a block in the chain.
Not a corporate restructuring. A signal. The memory industry is betting on a structural shift. One that could rewire the cost of storage for every AI inference server and every blockchain node. Chasing the alpha, one block at a time.
Here's the data I'm tracking: over the past 12 months, enterprise SSD prices have climbed 15-20%. But that's not the headline. The headline is that the NAND cycle—historically a brutal 2-3 year rollercoaster of boom and bust—might be flattening. And if it is, the implications for crypto infrastructure are massive.
Context: Why SanDisk matters for crypto.
SanDisk is the pure-play NAND giant born from the ashes of the Western Digital flash business. It shares fabs with Kioxia in Japan. It's a top-tier supplier of enterprise SSDs—the kind that power AWS, Azure, and Google Cloud. But also the kind that power Filecoin storage providers, Arweave nodes, and the growing fleet of AI inference servers that underpin decentralized compute networks.
For years, NAND was a commodity. A cyclical nightmare. When demand dropped, prices crashed, and margins evaporated. But the AI inference wave—and the parallel rise of decentralized storage—are introducing a new variable: persistent, high-capacity demand that doesn't peak and trough with consumer gadget sales.
From the front lines of the hype cycle, I've been watching this shift since 2024. When the Bitcoin ETF approval triggered a surge in institutional interest, I saw the same pattern. Now, it's NAND's turn. The question is: is this really a structural change, or just another cycle dressed in AI hype?
Core: The technical reality of NAND in the AI era.
Let's cut through the noise. The current generation of 3D NAND is at 218 layers for SanDisk/Kioxia's BiCS8. That's on par with Samsung and SK Hynix. No technological gap. But the real story is QLC – Quad-Level Cell NAND. For years, QLC was dismissed as consumer-grade, too slow for enterprise. But AI inference workloads are changing that equation.
Inference servers need capacity, not speed.
When you deploy a large language model, the model weights—hundreds of gigabytes to terabytes—are loaded into memory. But the long-term storage of those weights, plus the knowledge bases, logs, and checkpoints, sits on SSDs. Inference is read-intensive. It's not about random IOPS at 10K. It's about density per dollar.
SanDisk is betting big on enterprise QLC. And based on my audit of their product roadmap, they're aiming for 30+ TB SSDs in the E1.S form factor. That's a direct fit for the 1U inference servers that cloud providers and crypto miners are repurposing for AI.
But here's the hidden insight: the SanDisk-Kioxia relationship is a double-edged sword.
They share fabs in Yokkaichi and Kitakami. That means SanDisk's supply security is tied to a Japanese partner. On paper, it's a strength. But if Kioxia's financial health falters—or if geopolitical tensions squeeze Japan's semiconductor exports—SanDisk could face a supply bottleneck. The market isn't pricing this risk. I'd put a 6/10 confidence on this being a blind spot.
Capacity discipline: The new normal?
After the 2023-2024 crash, NAND manufacturers are terrified of oversupply. They're running at 85-90% utilization, not 95%. They're prioritizing margins over market share. SanDisk's capex is being held at 25-30% of revenue. That's lean. In a normal cycle, demand picks up, and they flood the market. But this time, they're holding back.
If AI inference demand continues to grow at 20%+ per year, supply will tighten. Prices will rise further. And that's good for SanDisk's stock—but it's a headwind for decentralized storage networks that rely on cheap NAND.
Contrarian: The structural shift might be overhyped.
Let me push back on my own thesis. I've been in this market since 2020, and I've seen cycles. The narrative that "AI kills the NAND cycle" is seductive, but it ignores a key variable: model compression.
Quantization and pruning are shrinking models.
LLMs are getting smaller. Mistral, Llama 3, and the upcoming models are being distilled. A 70B parameter model can be compressed to 4-bit, cutting storage needs by 75%. If inference moves to smaller, specialized models, the demand for high-capacity enterprise SSDs might not explode as expected. I've tested this myself—running a 7B quantized model on a consumer SSD. It works. The inference server isn't demanding terabytes per node.
And decentralized storage? It's still niche.
Filecoin and Arweave have real demand, but they're a fraction of the cloud market. The total storage capacity of the Filecoin network is around 20 exabytes. That's less than a single hyperscaler's annual NAND procurement. The crypto storage narrative is exciting, but it won't move the needle for SanDisk's revenue.
The real risk: overcapacity from China.
Yangtze Memory Technologies (YMTC) is locked out of advanced equipment by US sanctions. But they're still producing 128-layer NAND and selling it at a discount. If sanctions ease—or if YMTC finds a workaround—Chinese NAND could flood the market, crashing prices. SanDisk's high-end products would be safe, but the mid-range enterprise SSD market would bleed.
Takeaway: What to watch next.
I'm watching three signals. First, SanDisk's earnings call in Q3 2025. If they guide for capex above 30% of revenue, they're betting on sustained demand. Second, NAND contract prices. If they rise above 10% QoQ for two consecutive quarters, the cycle is tightening. Third, the adoption of QLC in enterprise. If AWS or Azure announces a QLC-based storage tier for AI inference, the thesis is confirmed.
For crypto infrastructure, the takeaway is clear: if NAND prices stay elevated, decentralized storage will become more expensive. That could slow adoption. But it could also drive innovation in more efficient storage algorithms—like proof-of-replication with better compression.
Surviving the winter to plant for spring. The sprint never stops, only the pace.
Signatures: - Chasing the alpha, one block at a time. - From the front lines of the hype cycle. - Surviving the winter to plant for spring. - Speed is the only currency that matters. - Pivoting when the chart says pause.