Today's AI-driven demand for memory chips may extend this industry cycle longer than history suggests, because production cannot keep pace.
High bandwidth memory (HBM), essential for AI accelerators, will consume 30% of the three largest makers' production capacity in 2027 while producing only 13% of memory bits, starving conventional DRAM supply even as AI servers demand more of it.
Although manufacturers are investing heavily—SK Hynix approved $38.3 billion in new South Korean facilities, and Micron is expanding U.S. capacity—long equipment lead times mean supply growth (projected at 24% in 2027) may still fall short of demand, keeping prices and margins elevated.
What happened
Memory chip makers are expanding capacity rapidly, but production growth is lagging AI-driven demand. TrendForce expects DRAM bit supply to increase about 24% in 2027, yet HBM (high bandwidth memory, critical for AI accelerators) will consume roughly 30% of production capacity while delivering only 13% of memory bits—leaving less room for conventional DRAM just as AI servers need more of it.
Why it matters
Memory booms typically collapse when oversupply pushes prices down, but this cycle may differ. Manufacturers cannot ramp fast enough; long equipment lead times and the shift toward HBM are keeping conventional DRAM supply tight and profitable. Samsung holds 38.5% of server DRAM market share, SK Hynix 28.8%, and Micron 22.4%, and all three are reporting historic margins and profit growth (Micron's revenue jumped 346% year-over-year to $41.5 billion in Q3 fiscal 2026; SK Hynix's operating profit rose 557% year-over-year in Q2 fiscal 2026).
What to watch
The turning point arrives when DRAM supply growth finally outpaces demand. TrendForce projects NAND supply will loosen in the second half of 2027, but DRAM capacity could remain tight through 2027—the signal that will mark the end of this extended boom.
Ask the AI about this article →
Memory chip cycles have historically followed a predictable boom-bust pattern: high prices drive manufacturers to add capacity, oversupply emerges, and prices collapse. Today's boom appears different because AI-driven demand for both HBM and conventional server DRAM is outrunning even aggressive capacity expansion. The crux lies in production physics. HBM is essential for AI accelerators but ties up disproportionate manufacturing capacity—TrendForce estimates it will consume 30% of the three largest makers' production lines in 2027 while accounting for only 13% of the memory bits produced. This efficiency gap starves conventional DRAM production just as AI servers themselves require more server DRAM, creating a structural capacity shortage that typical supply responses cannot quickly close.
All three dominant manufacturers—Samsung, SK Hynix, and Micron—are posting historic financial results, signaling how tight the market has become. Micron's revenue jumped 346% year-over-year to $41.5 billion in Q3 fiscal 2026, while SK Hynix's operating profit surged 557% year-over-year in Q2 fiscal 2026. Most tellingly, conventional server DRAM has been more profitable to produce than HBM since Q1 2026, because DRAM prices have risen faster while HBM prices are locked into longer-term contracts. This profit reversal accelerates reinvestment: SK Hynix approved $38.3 billion in new South Korean facilities, and Micron is expanding U.S. capacity. Yet TrendForce projects DRAM bit supply will grow only about 24% in 2027—likely insufficient to offset the capacity consumed by HBM expansion and surging server demand.
AI-summarized, only the topics you pick — one digest a day via Email, Slack, or Discord.
Free · takes 30 seconds · unsubscribe anytime
Ask AI anything about this article. Q&As are published on this page for other readers too.
