The NAND flash memory cycle is entering a phase that has analysts deeply divided — not about whether demand is real, but about how long the good times last, who captures the value, and whether today’s exuberance is pricing in risks that haven’t shown up yet. We brought together our three Korea Perspectives analysts to hash it out.
Let me be clear: I’m not dismissing the eSSD demand story. The structural shift is genuine. As AI workloads evolve from pure training toward inference and agentic AI architectures, the compute bottleneck is migrating downstream — from GPUs and HBM into networking, CPUs, and increasingly, storage. Machine learning training cycles are reportedly consuming roughly 30% of their time on data I/O operations alone, and some model architectures are pushing that figure toward 65% of epoch time. That is a real and pressing constraint, and enterprise SSDs are sitting directly in that critical path.
But here is where I get cautious. The memory industry’s history is a graveyard of premature cycle extensions. What we are observing right now is a bifurcated market: server DRAM and high-end eSSD are in genuine shortage, while consumer-facing NAND — smartphones, retail SSDs, laptops — is facing visible demand softness as end prices climb. When memory prices run this hard this fast, consumer electronics OEMs start stretching replacement cycles. That demand destruction at the consumer end has historically been the first crack in the foundation before it spreads upward. The current consensus view that the memory upcycle peaks in the second half of 2027 means we are already in the phase where forward-looking PER multiples on names like Samsung Electronics start to compress, even as headline earnings look strong.
There is also the SOCAMM wildcard. NVIDIA’s moves to reshape how memory interfaces with its next-generation platforms introduces genuine architectural uncertainty for both DRAM and NAND suppliers. The AI infrastructure buildout is accelerating, yes — but it is also becoming more proprietary, more vertically integrated, and frankly more concentrated in the hands of a few hyperscalers who are not shy about using their purchasing power. Korean memory companies are not setting the terms here; they are responding to them.
I will not argue with the macro cycle analysis, but I will argue with the investment conclusion it implies. The market has already priced in a great deal of the AI eSSD story. What it has not adequately priced, in my view, is the structural shift in the composition of NAND revenue — specifically the dramatic mix shift toward enterprise SSDs and away from commodity consumer NAND.
Look at what SanDisk’s most recent quarterly results are telling us. When you plot nearly a decade of quarterly margin data, this cycle looks genuinely different. Guided gross margins approaching 67% and operating margins near 55% are not numbers this industry has seen at this stage of a cycle. Historically, SanDisk — and by extension WDC — was not a serious player in the data center eSSD segment. For years after the 2016 acquisition, the enterprise SSD business was underinvested. What we are seeing now is a belated but forceful pivot toward hyperscaler demand, and the margin profile reflects a product mix that commands genuine pricing power. That is not noise; that is a fundamental change in business quality.
The more interesting value question to me is the relative valuation between HDD and NAND-based storage. Cold data storage in data centers still overwhelmingly runs on HDD — roughly 80% of total data center storage capacity. NAND-based SSDs hold about 20%. As AI inference workloads create persistent demand for fast-access hot and warm data, that ratio is structurally shifting. Seagate’s decade-long chart of quarterly revenues makes the HDD industry’s stagnation painfully obvious — with the brief exception of the 2011–2012 Thailand flood shortage and the recent AI training bump. The long-term value capture is accruing to NAND, and the market is still not fully reflecting the enterprise SSD moat being built by the survivors of this consolidating industry.
My concern is not the demand story — it is supply discipline. Samsung and SK Hynix are both aggressively expanding wafer capacity, with some reports citing monthly expansion targets that have nearly doubled from initial estimates, with securities houses now modeling potential 15-wafer-per-month additions at the high end. If both Korean majors push capacity simultaneously while chasing HBM economics, the NAND supply picture gets uncomfortable faster than the bull case assumes.
Both of my colleagues are thinking about this the right way, just at the wrong time horizon for actual money-making. Here is what I see on the ground: the eSSD demand cycle is not a future story — it is a present story that is still being underestimated in market positioning.
The datapoint that gets me excited is the HPE and Dell earnings pattern. Enterprise server and storage customers are accepting significant price increases without flinching. The constraint is supply, not demand — they said it explicitly. When traditional enterprise server buying combines with hyperscaler eSSD buildout in the same upcycle, you get an unusual double-demand wave that the industry has not seen in this configuration before. This is not 2021 consumer electronics NAND. This is infrastructure NAND driven by capex budgets that have multi-year commitments behind them.
The NAND cycle graph going back to the early 2000s tells a useful story. NAND cycles have historically lagged DRAM cycles and have often been more violent on both the upswing and the downswing — partly because the industry consolidated more slowly and partly because consumer end-demand was more price-elastic. The iPod Nano moment in 2005 that ignited the first real NAND supercycle is instructive: a single killer application unlocked years of structural demand growth. AI inference is that moment, scaled by orders of magnitude. KV cache storage, vector databases, inference logs — these are not discretionary; they are infrastructure requirements for AI services that are now commercializing at scale.
For Korean investors specifically, the play right now is less about pure memory OEMs — their margin story is getting crowded — and more about the equipment and materials suppliers enabling the physical layer shift. As NAND layers go higher and chips stack vertically, the laser processing and advanced packaging equipment requirements grow structurally. The companies enabling those process steps are running at high utilization now and have genuine forward visibility.
Synthesis: A Cycle With Structural Legs, But Discipline Will Define the Winners
What emerges from this debate is a picture of a NAND flash cycle that is genuinely different in character — driven by durable eSSD infrastructure demand rather than consumer electronics whims — but not immune to the classic risks of over-expansion and demand bifurcation. The margin profiles and guidance data from the major NAND suppliers suggest pricing power that this industry rarely sustains. The structural shift from HDD to SSD in data centers is real and durable. But the 2027 cycle peak consensus, combined with aggressive Korean capacity expansion plans, means the window for unambiguous upside is narrowing. For investors in Korean memory and related equities, the question is not whether the eSSD wave is real — it clearly is — but whether you are buying the structural story at a price that respects the cyclical risk hiding underneath it.