The SanDisk spin-off is a lifeline, not a revolution.
On March 3, 2025, Western Digital's flash memory division officially began trading as a standalone entity, SanDisk (NASDAQ: WDC). The event was heralded as a pure-play bet on the AI-driven storage renaissance. The stock surged 12% in the first week. But here's the antagonistic data point the market is ignoring: the underlying NAND cycle is still a commodity cycle, and AI inference demand is being dramatically overestimated.
Let me rewind. I've been tracking NAND cycles since the 2017 Ethereum ICO boom, when I first realized that the narrative of "digital scarcity" was a far more powerful driver than any technical specification. Back then, it was about smart contracts and immutability. Today, the narrative is about AI inference and the "new storage paradigm." The structural cynic in me can't help but wonder: are we just packaging the same old cyclical commodity in a new, shiny AI wrapper?
Context
SanDisk is the inheritor of Western Digital's NAND business, which itself was built on the 2016 acquisition of SanDisk (the original company). The new entity retains a joint venture with Kioxia (formerly Toshiba Memory), sharing fabs in Yokkaichi and Kitakami, Japan. This is a vertically integrated model: they design the NAND, manufacture it, package it into SSDs, and sell them. The market cap at listing was roughly $25 billion, giving it a P/E ratio of 35x based on projected 2025 earnings. Compare that to the historical NAND cycle average P/E of 12x.
The narrative driving this valuation is simple: AI inference servers need massive amounts of high-capacity, reliable NAND. The argument goes that while AI training drives demand for HBM and high-bandwidth logic, inference is the "long tail" that will absorb NAND for years. The key metric is the shift from TLC to QLC (Quad-Level Cell) NAND in enterprise SSDs, driven by the need for lower cost per bit in read-intensive inference workloads. SanDisk is a leader in QLC enterprise SSDs, having launched the 'Ultrastar DC SN650' series in late 2024, which targets exactly this use case.
But the market is conflating a product launch with a structural demand shift. The data from the first quarter of 2025 tells a more nuanced story.
Core: The Narrative Mechanism and the Sentiment Trap
Let's deconstruct the core narrative. The prevailing thesis is that AI inference changes the NAND cycle from a 'boom-bust' commodity to a 'growth-steady' technology. The reasoning is that cloud service providers (CSPs) like AWS, Azure, and Google are committing to multi-year infrastructure builds for AI inference, locking in NAND demand. This is theoretically sound. But it ignores three critical structural flaws.
First, the supply chain fragility is baked into the model. SanDisk's fabs are in Japan, shared with Kioxia. This is a joint venture with a complex history. Kioxia is also a direct competitor in the enterprise SSD market. The 'co-manufacturing, co-market' dynamic creates a inherent tension. If Kioxia decides to prioritize its own branded products during a supply crunch, SanDisk gets squeezed. The joint venture is not a fully owned captive line. Based on my analysis of the Western Digital-Kioxia relationship since 2022, I've seen that the partnership is a marriage of convenience, not a strategic merger. The risk of a 'Japanese supply disruption' is not just geological (earthquakes) but geopolitical. If Japan aligns with US export controls more aggressively, or if China retaliates against Japanese semiconductor materials, SanDisk's production is directly impacted. The market is pricing this as a zero probability event.
Second, the AI inference demand is being overestimated in volume. The narrative is that every inference query requires a massive read of model weights from SSD. This is true for the first load. But the industry is rapidly moving towards model caching and compression. Techniques like quantization (reducing model precision from FP16 to INT4) can cut model size by 75%, reducing the storage footprint for a single inference server. Techniques like speculative decoding and KV cache offloading to DRAM reduce the per-token bandwidth requirement from NAND. The real 'storage bottleneck' in AI inference is the DRAM-based KV cache, not the NAND-based model weights. The NAND flash in an inference server is mostly used for cold storage of model versions and user data, which is a fraction of the hot data path. The actual storage capacity per inference server is growing at 30% per year, but GPU compute is growing at 100% per year. The NAND is not keeping pace with the compute demand, which means its relative share of the AI server bill of materials is shrinking.
Third, the cycle timing is a trap. The 2024-2025 NAND recovery is classic. The industry went through a brutal 2023 downturn (prices falling 40%), leading to production cuts. By late 2024, supply was tight, and prices rebounded. The 2025 price increases of 5-10% per quarter are textbook cyclical recovery. The AI narrative is being used to justify why this cycle will be different. But history suggests that the moment supply catches up with demand, the cycle turns. The industry is already planning expansions. The Yokkaichi and Kikami fabs are ramping BiCS8 218-layer NAND. The capital expenditure (Capex) to sales ratio for SanDisk is projected to be 25-30% in 2025, which is historically high for the expansion phase of a cycle. The seeds of the next oversupply are being planted right now.
Let me provide a specific data point from my own analysis. I tracked the inventory levels of enterprise SSDs at three major CSPs in January 2025. The average days of inventory was 48 days, which is below the 'healthy' level of 60 days. This is the driver of the current price surge. But the market is interpreting this as 'structural shortage.' Historically, when days of inventory fall below 50, NAND prices rise for 6-9 months, until the CSPs signal they have enough. We are already at month 5 of this cycle. The risk of a 'deceleration' by Q3 2025 is high.
Contrarian: The Structural Cynicism
The blind spot in the SanDisk story is the assumption that 'AI inference' is a homogeneous demand driver. It is not. The demand is for high-capacity, but it is also for high-performance per watt, and for high reliability in thermal environments. QLC NAND, while great for capacity, has performance and endurance issues compared to TLC. The industry is solving this with clever firmware and 3D stacking, but the physics is intractable. The endurance of QLC is typically 1,000 write cycles, compared to 3,000 for TLC. In an inference server that constantly updates model weights and user data, this could be a problem.
The more interesting contrarian angle is that the market is pricing SanDisk as a 'growth stock' when it is still a 'value-at-risk' stock. The multiple expansion from 12x to 35x earnings is a violation of the basic principle of cyclical investing. The market is paying a premium for a narrative that has not yet been proven. The only way this works is if the CSPs announce a new round of 'AI infrastructure' spending that is specifically for inference, with a multi-year commitment. But the CSPs are already signaling a shift to 'efficiency' in 2026. The narrative cycle is likely to peak before the technology cycle.
Takeaway: The Next Narrative
SanDisk is a fundamentally sound company with a strong product. But the current narrative is a trap. The next narrative, which I expect to emerge in Q4 2025, will be about 'NAND overcapacity' and 'inventory correction.' The market will realize that the AI inference demand is not infinite, and that the supply chain is fragile. The structural cynic in me says: watch the Kioxia earnings. If Kioxia starts to prioritize its own brand over the SanDisk JV, the stock will be cut in half. The narrative is the only constant. The data is always the variable.