2027年将迎来史上最严重内存荒？

On July 10, 2026, the day SK hynix listed on Nasdaq, CEO Kwak Noh-Jung warned that the global memory chip industry would face its most severe supply shortage in history in 2027, with demand outstripping capacity possibly lasting beyond 2030. Earlier, Micron CEO Sanjay Mehrotra said it is unclear when supply will catch up with rising demand. NVIDIA founder Jensen Huang predicts the AI memory shortage will persist for years, with SK hynix remaining NVIDIA's largest memory supplier.

The near-simultaneous warnings from three industry leaders mark a sharp departure from typical memory industry cycles of the past decades. Previously, rising demand prompted expansion, then capacity release drove prices down, typically over 2-3 years. But this time, AI is not only creating new demand but also compressing supply—the boom in HBM is reshaping the supply-demand dynamic.

Starting in 2023, HBM demand surged rapidly. The combined HBM wafer capacity of Samsung, SK hynix, and Micron increased from about 123,000 wafers per month at end-2023 to roughly 331,000 wafers per month at end-2025, and is projected to reach about 668,000 wafers per month by end-2027. By then, HBM may account for 35% of the three companies' total DRAM wafer capacity. However, HBM is not ordinary DRAM: it requires larger chip area, through-silicon vias, wafer thinning, backside processing, and complex stacking, making it more difficult to manufacture with lower yields. For example, for HBM3E 12-layer stacking, the same wafer produces about three times more bit output from ordinary DRAM than from HBM; with HBM4, the gap may widen to nearly four times.

This means the more aggressively manufacturers shift to HBM, the tighter ordinary DRAM becomes. SemiAnalysis estimates that from 2026 to 2027, the commodity DRAM supply gap may persist at about 7%; the HBM gap may widen from about 6% in 2026 to about 9% in 2027. This is a rare dual shortage: AI servers compete for HBM, while PCs, smartphones, and general servers compete for the remaining DRAM capacity.

The supply side cannot respond quickly. Building new advanced memory fabs requires billions of dollars and years of construction; clean rooms, equipment installation, qualification, and yield ramping all take time. New capacity additions in 2026 are mainly from Samsung's P4, SK hynix's M15X, and Micron's A3, and the latter two will still prioritize HBM. Truly large-scale new capacity may not be released until the second half of 2027.

Prices are therefore accelerating. Reports estimate that in Q1 2026, DDR5 and LPDDR5 contract prices rose 70% and 35% quarter-on-quarter respectively; from Q4 2025 to Q4 2026, DRAM prices could double again. The super cycle is a double-edged sword: memory manufacturers see improved profits, but component costs for smartphone, PC, and AI server manufacturers continue to rise, forcing end users to absorb pressure through price increases, spec downgrades, or reduced shipments.

What truly determines when the cycle ends is not just demand slowing, but the ramp-up of new fabs, migration to 1b and 1c nodes, HBM4 yield improvement, and whether manufacturers lose capital discipline again. At least for now, this AI-driven memory cycle may be larger and longer-lasting than in the past.

Constraints include: new capacity release timing in the second half of 2027, when HBM may account for an even higher share of capacity; if HBM4 yield improves slower than expected, the squeezing effect on ordinary DRAM will continue. Uncertainties include: whether AI demand growth slows, whether manufacturers over-expand leading to the next glut, and geopolitical impacts on supply chains.