SK Hynix's NAND Lag Is a Quiet Threat to AI Data Centers

SK Hynix Is Winning HBM but Losing NAND

SK Hynix has built its reputation on high-bandwidth memory, dominating the HBM market that powers AI accelerators from NVIDIA and AMD. But behind that spotlight, a quieter competitive erosion is underway — one that could reshape the company’s position in the memory chip industry and create friction for AI data center supply chains.

According to industry sources cited by Chosunbiz, SK Hynix’s 176-layer NAND accounted for roughly 50% of its NAND production in the second quarter of 2026. That means half of its NAND output relies on a process generation that competitors have already moved past. Samsung Electronics had shifted its mainstay to 236-layer NAND, with that generation accounting for over 60% of its Q2 output. Micron’s 232-layer products held 53% of its mix, with 276-layer at 33%. China’s Yangtze Memory Technologies (YMTC) put 232-layer at 68% of its production.

The generational gap is not trivial. NAND flash capacity scales with vertical stacking depth — more layers mean more bits per wafer and lower unit costs. A producer stuck on older layers must run more wafers and more fab time to deliver the same output, eroding profitability as prices rise. In a market where NAND prices have been climbing, the penalty for lagging in layer count is paid twice: higher costs per bit and lower capture of price upside.

What the Market Share Data Actually Says

The consequence appeared clearly in revenue market share figures. SK Hynix, including its SoliDIM sub-brand, saw its NAND share fall from 21.4% in Q4 2025 to 16.4% in Q1 2026 — a near 5 percentage point drop in a single quarter. Over the same period, YMTC surged from 11.5% to 16.2%, nearly closing the gap. The timing matters. NAND prices were rising during that window, which should have favored the producer with the most efficient generation mix. Instead, SK Hynix lost ground to a rival running newer process nodes at scale.

An industry source noted that in a price-increasing cycle, higher-generation product share amplifies margin gains because the price uplift combines with lower per-bit costs. SK Hynix’s decline, therefore, is not simply a volume problem. It is a structural cost disadvantage rooted in slow世代 transition.

SK Hynix itself highlighted in a recent earnings call that its 321-layer product reached the largest share of NAND production in Q1. But that figure is reported on a production capacity basis — a measure of installed capability, not shipped volume. Analytics firms such as Omdia, which track actual market production volumes, show a substantially different picture. The discrepancy suggests SK Hynix is promoting roadmap progress even as its revenue numbers reflect an older product mix still dominating the line.

Capital Allocation Is the Real Story

The generation lag did not emerge overnight. It reflects a deliberate capital allocation choice. SK Hynix’s NAND equipment spending has fallen for three consecutive years — from $3.5 billion in 2023 to $3.0 billion in 2024 and $2.9 billion in 2025. Meanwhile, DRAM and HBM lines have absorbed the bulk of remaining investment. The strategy is understandable in the short term. HBM demand from AI training and inference workloads has pushed margins and pricing to levels that dwarf NAND. Every dollar directed toward HBM fab expansion yields a higher immediate return than a dollar spent on NAND node transition.

But the trade-off is real. While SK Hynix compounds its HBM advantage, rivals are consolidating NAND leadership. Samsung’s ability to concentrate production on a single generation — 236-layer — has allowed it to drive yield and cost down faster than a multi-generation operation can. SK Hynix, by contrast, runs 176-layer, 236-layer, and newer nodes in parallel, which fragments line efficiency and slows yield improvement on any single product.

Micron and YMTC are executing the same concentrated strategy. Micron’s shift from 232-layer to 276-layer is well underway, and YMTC has already locked in 232-layer as its dominant output. Kioxia and SanDisk, while still carrying significant older generation share, accelerated their BiCS8 (218-layer) transition from 9% in late 2025 to 24% by Q2 2026 — a nearly threefold increase in two quarters.

What This Means for AI Data Center Supply Chains

The implication for AI infrastructure is subtle but meaningful. Data centers do not run on HBM alone. NAND storage sits alongside memory tiers in every rack, and the cost-per-gigabyte of that storage directly shapes total data center economics. When NAND producers with newer generations gain share, they gain pricing leverage and margin room. When a major producer like SK Hynix lags, it cedes that leverage to competitors.

SK Hynix’s NAND shortfall also creates a supply concentration risk. If Chinese demand for NAND continues growing — and YMTC’s share gain suggests it will — the company faces a narrower set of alternative suppliers for bulk storage procurement. Samsung is the primary alternative, but Samsung is simultaneously deepening its own NAND capacity. That dynamic compresses negotiation options for buyers who previously relied on SK Hynix as a competitive check on Samsung pricing.

The company’s own commentary on the impact of rising NAND prices on its earnings has been measured. An industry source indicated that the current price-upcycle effect on SK Hynix’s results remains limited, precisely because its product mix does not yet capture the full upside. If the price cycle extends — as some analysts project through 2027 — the gap between SK Hynix’s potential and its actual performance could widen further.

The Question Ahead

SK Hynix is not in crisis. Its HBM business is strong, and its 321-layer roadmap is real. But the data from Q2 suggests the company is managing a trade-off that will test its long-term positioning: chasing the highest-margin memory segment while letting a second major product line fall behind.

The memory chip industry rewards concentration. Samsung, Micron, and YMTC are all betting on it. SK Hynix is betting on HBM instead. The question is whether that bet can sustain its market position if NAND share continues to erode — and whether AI data center operators will notice the shift until it is too late to diversify their supply base.