Technology — 2026-04-19
SK Hynix chairman warns HBM/DRAM shortage may persist to 2030 as AI demand outstrips capacity
SK Hynix chairman
Analysis
Converging reporting and hyperscaler contract behavior, with Google and Microsoft pursuing five-year DRAM supply deals, shift AI-capex planning from cyclical-squeeze into structural-constraint framing. Fab capacity allocated to HBM4 and HBM4E for Nvidia Vera Rubin crowds out conventional enterprise and automotive DRAM, making non-AI memory prices very likely to sustain 15–25% year-over-year increases through 2026, with automotive OEM BOM exposure crystallizing by late Q3. Samsung's Pyeongtaek P5 offline until 2028 removes any credible capacity relief before then; whether DPA-class US industrial action to reshore HBM-grade fab becomes viable before the 2027 election cycle is the live policy question. Chey, however, carries a disclosed commercial interest: the 2030 scarcity horizon lands alongside locked-in HBM3E price hikes, and the claim may be as much contract-positioning as a physics-constrained forecast.
Converging reporting and hyperscaler contract behavior, with Google and Microsoft pursuing five-year DRAM supply deals, shift AI-capex planning from cyclical-squeeze into structural-constraint framing. Fab capacity allocated to HBM4 and HBM4E for Nvidia Vera Rubin crowds out conventional enterprise and automotive DRAM, making non-AI memory prices very likely to sustain 15–25% year-over-year increases through 2026, with automotive OEM BOM exposure crystallizing by late Q3. Samsung's Pyeongtaek P5 offline until 2028 removes any credible capacity relief before then; whether DPA-class US industrial action to reshore HBM-grade fab becomes viable before the 2027 election cycle is the live policy question. Chey, however, carries a disclosed commercial interest: the 2030 scarcity horizon lands alongside locked-in HBM3E price hikes, and the claim may be as much contract-positioning as a physics-constrained forecast.