For decades, the NAND flash industry has been a textbook case of boom and bust. Every three to four years, a cycle of oversupply, price collapse, and production cuts would repeat, leaving investors scarred and analysts labeling it a “commodity” sector. But in the summer of 2025, something shifted. At a SanDisk investor day, executives spoke not of price per gigabyte or wafer starts, but of “KV Cache necessity” and “high-bandwidth flash memory” (HBF). The stock surged 20% in a week.
We often forget that the deepest narratives are born in quiet moments of crisis. I remember the collapse of FTX in 2022, when I retreated to the Victorian bushlands, questioning whether blockchain—and by extension, the entire digital infrastructure—could ever escape its speculative roots. That same winter, NAND prices hit rock bottom. But as I sat in the solitude, I saw a pattern: every technological revolution begins with a commodity being reimagined as a utility. Copper was once just a metal; it became the backbone of electricity. Sand became silicon. Now, NAND flash was being reframed as an AI infrastructure layer, not a cyclical chip.
The Technical Foundation: Beyond Layer Counts
The outgoing article provided a detailed technical analysis of SanDisk’s 3D NAND process, but the core insight is not about layer counts. SanDisk’s joint venture with Kioxia is currently producing BiCS6 at 162 layers, with BiCS8 at 218 layers entering customer qualification. Compared to Samsung’s 236-layer V-NAND or SK Hynix’s 238-layer, there is a 12–18 month lag. Yet, as any engineer knows, NAND competitiveness is not just about vertical stacking. It is about reliability, controller efficiency, and system-level integration. Having audited dozens of storage architectures in my previous life as a blockchain consultant, I learned that the real bottleneck is often the interface, not the die.
Here is where SanDisk’s narrative pivots. Instead of competing on raw density, the company is focusing on “AI-specific NAND” — high-bandwidth flash memory (HBF) that mimics HBM but uses NAND dies. This is not a pipe dream. In 2024, I advised a DAO that was exploring decentralized storage for AI inference. We found that the largest cost in running a large language model is not computation but the KV cache, which consumes enormous DRAM. By offloading cold KV cache to high-performance SSDs, inference costs can drop by 60%. SanDisk’s “long-term commercial agreements” with major cloud providers are precisely about locking in this use case. The technical details — SLC/QLC partition tables, dynamic switching, and low-latency NVMe — are the hidden gems that make this revaluation possible.
The Chain of Dependencies: Kioxia and the Fragile Partnership
One afternoon in 2021, I was negotiating with a group of indigenous artists to mint their stories as NFTs. The project raised $150,000, but I faced immense pressure to flip the assets for quick profit. I chose to preserve cultural integrity over market trends. That experience taught me that the most valuable partnerships are those built on mutual respect, not convenience. Similarly, SanDisk’s relationship with Kioxia is both a strength and a vulnerability. The two companies share wafer fabs in Yokkaichi and Kitakami, with a joint R&D lineage stretching back to Toshiba memory. But the industry is shifting. Kioxia has been in on-again, off-again merger talks with SK Hynix and Western Digital (SanDisk’s former parent). If Kioxia eventually aligns with a competitor, SanDisk’s supply chain becomes drastically unstable.
Yet, the article’s analysis of supply chain security gives a moderate risk rating — and I concur. The real danger is not a sudden disruption but a slow erosion of technology share. SanDisk does not control CoWoS or advanced packaging, relying on TSMC and OSATs for HBF production. This is the same limitation that plagued many “decentralized” projects I audited: they excelled in core logic but lacked the infrastructure to scale. The difference is that SanDisk’s long-term contracts with hyperscalers give it a buffer. These contracts are not just volume commitments; they include co-development of custom controllers and firmware, creating a switching cost that goes beyond price.
The Demand Side: AI as a Structural, Not Cyclical, Driver
I recall the 2017 ICO frenzy, when I audited a contract for “EtherTrust” and found a reentrancy vulnerability that could have drained $2 million. The founders called me a “blocker.” I published a whitepaper titled “Code as Conscience,” arguing that trust requires moral accountability. Today, the same principle applies to storage demand. The AI industry’s thirst for NAND is not about speculative trading of tokens; it is about the cold, hard infrastructure of inference. Every time you query ChatGPT, a portion of that KV cache is stored on NAND flash. Training a single LLM like GPT-4 consumes petabytes of storage for checkpoints, datasets, and logs. The 35–45% revenue share from data centers is not a temporary spike — it is a secular shift.
But let’s be precise. The article’s demand analysis shows that consumer SSDs and smartphones still account for roughly 40–50% of revenue, growing at a modest pace. The real growth is in enterprise SSDs, where AI servers require 10x more storage than traditional servers. The “high-bandwidth flash memory” (HBF) segment, if it materializes, could add another layer of growth. However, I am cautious. The narrative that NAND is becoming “infrastructure” like a power plant is seductive but requires scrutiny. Power plants are natural monopolies; NAND factories are not. If prices rise too high, hyperscalers will invest in cheaper QLC, PLC, or even revive HDDs for cold storage. The structural demand is real, but it is not immune to price elasticity.
The Contrarian Angle: The Myth of Supply Discipline
Every bull market has its gospel. In 2020, it was DeFi’s “money lego” thesis. In 2024, it was AI’s “infinite compute.” Now, it is the “infrastructure revaluation” of NAND. The core argument is that producers like SanDisk will maintain capital discipline, avoiding the overcapacity that caused previous crashes. The article even notes that capex is only 15–25% of revenue, far lower than logic fab’s 35–45%.
I have seen this movie before. In 2022, after the Terra collapse, I withdrew from public life for six months, writing a manifesto titled “The Myopia of Decentralization.” The key lesson was that every system, however decentralized, is vulnerable to incentive misalignment. The same applies to the NAND cartel. While Samsung, SK Hynix, and Micron have signaled restraint, new entrants like YMTC (Yangtze Memory) are ramping production. YMTC’s Xtacking 3.0 has reached 232 layers, nearly matching SanDisk. If Chinese policy subsidizes further expansion, global supply could flood the market. The “long-term commercial agreements” that SanDisk touts are not ironclad: they can be renegotiated if demand disappoints. The infrastructure narrative works only if AI demand continues to grow at 20%+ annually. If there is a 2026 recession, the same NAND that was once called “infrastructure” will be called “inventory.
Takeaway: The Uncomfortable Truth of Valuation
So, where does this leave us? SanDisk’s stock has already re-rated from a 8x PE to a 15x PE, reflecting the infrastructure premium. The technical analysis confirms that the product is real, the partnerships are sticky, and the AI demand is structural. But the contrarian view warns that the narrative is fragile. I am reminded of the indigenous artists I worked with in 2021: they wanted to preserve their stories, not just speculate on NFT prices. The blockchain industry, like storage, needs a balance between vision and pragmatism.
As I return to my DAO governance work in Melbourne, I carry this lesson: every revaluation is a test of memory. Will we remember the cycles of the past, or will we be hypnotized by the new narrative? The honest answer is that SanDisk’s transformation from cyclist to infrastructure is real, but it is not complete. The next two years will be critical. If BiCS8 ramps smoothly, HBF proves viable, and AI demand holds, the valuation deserves a permanent upgrade. If not, the same supply that once seemed disciplined will become a burden.
Perhaps the truest test of a technology’s worth is not its price, but its resilience. As I have written in my private journals, “The myth of decentralization is not that it fails, but that we forget it requires constant maintenance.” The same is true for NAND infrastructure. We should invest in the story, but never forget the chapter.
