Bernstein Research Report Analysis: HBM Prices Set to More Than Double Next Year; Memory Becomes a Burden for AI - 深潮TechFlow

Bernstein's latest research indicates that due to long-term contract pricing mechanisms for HBM (High Bandwidth Memory), its profit margins are significantly lower than those of standard DRAM. The rep... Read original →

The lag in HBM pricing mechanisms is triggering a fundamental reshaping of AI hardware economics. Locked into annual contracts, Samsung and SK Hynix absorbed the cost of advanced packaging without commensurate margins in 2025–2026, forcing aggressive 2027 price hikes to sustain capex efficiency. This will directly inflate NVIDIA GPU system costs, raising AI data center CAPEX by up to 15% and accelerating ASIC adoption—opening a strategic window for MediaTek’s TPU initiatives. Technologically, EUV availability and CoWoS capacity will bottleneck HBM3E/4 ramp; Kioxia, lacking integration capability, faces obsolescence. Geopolitically, U.S.-led export controls on lithography tools have already increased material costs and yield risks, while TSMC’s CoWoS concentration in Taiwan, China amplifies supply fragility. Over the next 18 months, memory makers will shift from P/B to forward P/E valuation amid record ROEs—but if AI training demand cools, HBM could swiftly morph from strategic asset to inventory liability.