A single data point from the NAND forward market is telling a story the blockchain storage sector hasn’t yet priced in: the AI inference boom is rewriting the supply-demand curve for flash memory, and the decentralized storage networks that depend on it are about to face a structural shift.
Context: The NAND Industry and the SanDisk Signal
NAND flash is the physical substrate of virtually every decentralized storage network — Filecoin, Arweave, Storj, Sia. The cost of storage, the latency of retrieval, and the durability of data all trace back to the 3D NAND layers stacked inside SSDs. For years, the industry has been a textbook cyclical commodity: boom when demand outstrips supply, bust when oversupply crashes prices. The 2022-2023 downturn was brutal, with combined losses of over $20 billion across major NAND producers.
Then came the SanDisk spin-off from Western Digital in early 2025. This was not just a corporate restructuring. It was a declaration that the storage market is bifurcating: consumer NAND (phones, PCs) is a mature, low-growth business, while enterprise SSD — driven by AI — is becoming a structurally different asset class. SanDisk, now independent, inherits the crown jewel of the WD-Kioxia joint venture: the BiCS8 218-layer NAND, a technology that sits at the absolute frontier of the industry, alongside Samsung’s 300-layer and SK Hynix’s 300-layer stacks.
But the real story is not the layer count. It is the demand driver. AI inference — not just training — is consuming NAND at an unprecedented rate. Every inference request on a large language model requires loading the model weights (hundreds of gigabytes to terabytes) from high-capacity SSDs into DRAM, and the KV cache also demands storage. The inference server fleet is scaling faster than training clusters, and each server now carries 30-50TB of enterprise SSD capacity. This is a new marginal buyer for NAND, one that is less price-sensitive than smartphone OEMs or cloud hyperscalers running general-purpose workloads.
Core: How AI Inference Is Reshaping the NAND Cycle
The traditional NAND cycle operated on a 2-3 year rhythm: oversupply, price collapse, industry losses, capacity cuts, price recovery, then oversupply again. The 2024-2025 upcycle began in Q4 2024, with NAND contract prices rising 5-10% per quarter, and enterprise SSD prices climbing even faster. Industry utilization rates recovered to 85-90% by mid-2025, a level that signals healthy demand but not yet full throttle.
What is different this time is the composition of demand. Based on industry data, enterprise SSD now accounts for 25-30% of NAND revenue, and its growth rate of 20%+ is double that of the overall market. The driver is AI inference, not just training. Training requires HBM and high-bandwidth memory, but inference requires massive, reliable storage. The shift from training to inference is a structural change: inference is a scaling law of its own, with token counts growing exponentially, and each token is backed by storage reads. This is not a one-time capex spike; it is a recurring operational need.
SanDisk, with its BiCS8 QLC (Quad-Level Cell) NAND, is positioned to capture this demand. QLC, which stores four bits per cell, offers lower cost per gigabyte but slower write speeds and lower endurance. For read-intensive inference workloads — where the model is loaded once and read repeatedly — QLC is ideal. SanDisk has already launched enterprise-grade QLC SSDs targeting AI inference, and the technology is being validated by major cloud providers.
But there is a hidden tension here. SanDisk shares its wafer fabs in Yokkaichi and Kitakami, Japan, with Kioxia. This is a co-opetition structure: they co-develop and co-manufacture, but they compete in the enterprise SSD market. Both are selling AI inference-grade SSDs to the same hyperscalers. This dual role creates a fragility in the supply chain. If Kioxia decides to prioritize its own brand or if the joint venture faces financial strain, SanDisk’s access to its own wafer supply could be constrained. The market has not priced this risk.
From a capacity perspective, the NAND industry is showing unprecedented capital discipline. The 2023 losses were so severe that all major producers — Samsung, SK Hynix, Micron, Kioxia, and SanDisk — are cautious about expanding capacity. The new Fab in Kitakami is a phased build with conservative capital expenditure. The industry’s capex-to-revenue ratio of 25-35% is lower than the historical average. This means that even if AI demand continues to grow, supply will be constrained, potentially keeping NAND prices elevated for longer than the traditional cycle would suggest.
For blockchain storage networks, this is a double-edged sword. Higher NAND prices mean higher hardware costs for storage miners. The Filecoin model, for example, requires miners to commit sealed sectors on high-capacity SSDs. If the cost of those SSDs rises, the return on investment for mining decreases, potentially reducing network participation. On the other hand, the narrative of “decentralized storage for AI data” becomes more compelling as centralized cloud providers face rising costs — but only if the blockchain networks can offer competitive pricing.
Contrarian: The Over-Optimism Trap
There is a flaw in the AI-inference-driving-NAND thesis. It assumes that the demand for storage at inference time is linearly correlated with the growth of AI usage. But the reality is more nuanced. Model compression techniques — quantization, pruning, distillation — are rapidly reducing the size of models without sacrificing accuracy. The industry is already seeing a shift from 70B models to 8B or 1.5B models that run on-device or in edge servers. The inference server of the future may not need 50TB of SSD; it might need 5TB, because the model is compact and the KV cache is optimized.
If this compression trend accelerates, the marginal demand for NAND from AI inference could plateau as early as 2026. The storage industry has a history of overestimating new demand drivers. In 2018, the “bitcoin mining uses NAND” narrative was disproven. In 2021, the “metaverse will drive storage” narrative fizzled. The current AI inference narrative is more robust, but it is not immune to technological disruption.
Furthermore, the SanDisk-Kioxia relationship is a time bomb. The joint venture was originally formed to share the massive capital costs of NAND fabrication. But now that SanDisk is independent, the strategic interests of the two companies may diverge. Kioxia is rumored to be exploring an IPO, and a public company would be under pressure to maximize its own brand sales, potentially at the expense of SanDisk’s wafer allocation. This is a governance failure waiting to happen — a classic case of misaligned incentives in a shared infrastructure.
Another contrarian angle: the blockchain storage token market (FIL, AR, STORJ) is already pricing in a rosy scenario for NAND supply. The market cap of these tokens has risen significantly in 2025, partly on the AI narrative. But the actual economics of decentralized storage networks are still challenged by latency, retrieval speed, and reliability. The NAND cycle’s current upswing will increase the cost of provisioning new storage nodes, which could squeeze margins and reduce the incentive to mine. This is a headwind that the token market is ignoring.
Takeaway: The Convergence Is Real, But the Timing Is Uncertain
The convergence of AI inference and blockchain storage is not a fantasy. It is a measurable trend in the NAND market. Every line of code writes a history of power, but every layer of NAND writes a history of cost. The governance of decentralized storage networks must adapt to the reality that their most critical input — NAND flash — is becoming a structurally different commodity, influenced by AI demand and constrained by cautious capacity expansion.
We didn’t account for the fact that the supply chain for decentralized storage is as much a political commodity as a technical one. The SanDisk-Kioxia relationship, the capital discipline of the industry, and the potential for model compression to deflate the narrative are all factors that need to be monitored. Governance isn’t just about protocol-level voting; it’s about understanding the physical infrastructure that underpins the network.
The next 12 months will be a test. If NAND prices continue to rise while model compression accelerates, the blockchain storage thesis will face a stress test. If, instead, AI inference demand proves sticky and NAND supply remains constrained, then the decentralized storage tokens that successfully navigate the cost environment will emerge as dominant players. Truth emerges from transparency, not from silence. The NAND market’s opacity is a risk for DePIN, and it is time for the blockchain community to start reading the signals from the semiconductor world.