The spin-off of Sandisk from Western Digital in late 2024 was a structural event that most crypto analysts dismissed as a legacy storage play. They saw a commodity memory manufacturer, not a barometer of the AI-driven shift in data architecture. But as a macro watcher who has spent years mapping liquidity flows across asset classes, I see something else: the NAND cycle is undergoing a regime change, and the market is mispricing the second-order effects on crypto-native storage networks.
Let me be clear from the start. This is not a bullish thesis for Filecoin or Arweave. Quite the opposite. The same AI inference boom that is lifting Sandisk's enterprise SSD margins is revealing the fundamental fragility of decentralized storage solutions. The narrative that 'AI needs crypto storage' is a consensus that has not been stress-tested by the cold mathematics of NAND economics.
Context: The NAND Cycle Meets AI Inference
Traditional NAND cycles are driven by supply discipline and consumer demand. Every 2-3 years, oversupply crashes prices, then a demand recovery (smartphones, PCs) pulls prices up. The 2023-2024 period was a textbook downcycle: NAND prices fell over 50%, forcing all major players (Samsung, SK Hynix, Micron, Kioxia/Sandisk) to cut production. The recovery began in late 2024, driven by enterprise SSD demand from AI training clusters.
But the new variable is AI inference. Unlike training, which requires massive HBM bandwidth and GPU compute, inference is latency-sensitive and storage-intensive. A single inference server might load a 700GB model into DRAM, but the model weights, KV cache, and user data cycles require terabytes of high-speed NAND. The shift from training to inference is a shift from compute-bound to storage-bound workloads. This changes the demand profile for NAND: it becomes more elastic, more persistent, and less cyclical.
According to industry data, enterprise SSD revenue grew over 20% year-on-year in Q1 2025, with AI-related storage accounting for the majority of the increase. Sandisk, now independent, is positioned to capture this trend with its BiCS8 218-layer NAND and enterprise QLC SSDs. The company's focus on density and cost-per-bit aligns perfectly with inference workloads that prioritize capacity over write endurance.
Core Analysis: The Quantitative Case for a Structural Shift
Let me quantify the shift. In a pre-AI data center, storage was a static line item. Today, inference servers consume 5-10x more storage per server than traditional cloud servers. The average AI inference node requires 8-16TB of NVMe SSD capacity. With global AI inference server shipments projected to grow from 1.5 million units in 2024 to 5 million units by 2027, the incremental NAND demand is staggering.
Using a simple model: if each inference server requires 12TB of NAND, and 5 million servers are deployed by 2027, that is 60 exabytes of additional NAND demand. For context, the entire NAND industry shipped roughly 600 exabytes in 2024. This represents a 10% boost to total demand, concentrated in the highest-margin enterprise segment. Liquidity is the pulse; policy is the brain. Here, the policy is AI adoption, and the pulse is NAND pricing.
The bull case for Sandisk is straightforward: as a pure-play NAND manufacturer, it will benefit from both volume growth and price stability. The company's capital expenditure discipline (maintaining 25-30% capex-to-revenue) relative to the investment cycle suggests that supply will remain tight through 2026. Storage prices are expected to rise 20%+ in 2025, and enterprise SSD premiums are widening.
But the crypto market has drawn a false parallel. The narrative that 'AI inference will drive demand for decentralized storage' is seductive but mathematically flawed. Let me explain why.
Contrarian Angle: The Decoupling Illusion
Crypto storage tokens like Filecoin and Arweave are built on the assumption that enterprise users will pay a premium for verifiable, decentralized storage. The logic is that AI models and data must be censorship-resistant and immutable. This is a powerful narrative, but it ignores the economics of NAND.
The enterprise SSD market is dominated by hyperscalers (AWS, Azure, GCP) who buy NAND at scale. They are not paying retail prices for crypto storage. The cost of storing 1TB on Filecoin is approximately $0.10 per TB per month, but the hidden costs—retrieval latency, redundancy, and computational overhead for proofs—make it significantly more expensive than centralized cloud storage when you factor in the cost of compute for AI inference. An inference server cannot wait 10 minutes for a file to be retrieved from a decentralized network. Value is a consensus, not a fundamental truth. The market has assigned value to decentralized storage based on a narrative of trust, not on the technical reality of inference latency.
Furthermore, the AI industry is moving toward model compression and quantization. Techniques like 4-bit quantization reduce model size by 4x, meaning the storage demand per inference server may actually decrease over time. The same NAND that is now in short supply could see a demand pullback if inference becomes more efficient. This is the second-order effect that the crypto storage narrative ignores: the elasticity of AI storage demand is asymmetric.
I have seen this pattern before. In 2021, I published a forensic audit of BAYC wash trading, proving that 60% of volume was artificial. The market ignored the data until the floor price collapsed. Today, the same dynamic is playing out in crypto storage. The on-chain metrics show that Filecoin's deal-making is growing, but the deals are overwhelmingly for cold storage—archival data, not active inference workloads. The value proposition for AI inference is hot storage, and decentralized networks cannot compete on latency or cost.
Takeaway: Positioning for the Real Cycle
The AI inference boom is real, and it is reshaping the NAND cycle. But the beneficiaries are centralized hardware manufacturers like Sandisk, not crypto storage tokens. The next two years will see a decoupling: NAND prices rise, Sandisk margins expand, and crypto storage tokens struggle to justify their valuations as the market realizes that inference workloads will never migrate to decentralized networks.
For the crypto investor, the question is not whether to buy Filecoin, but whether to hedge against the inevitable correction in storage token narratives. The structural shift in NAND is a signal: follow the hardware, not the hype. The liquidity is flowing into enterprise SSDs, not into proof-of-replication circuits. Macro always wins. And the macro is telling us that the only storage that matters for AI inference is the kind that arrives in milliseconds, not minutes.