Memory squeeze set to tighten through 2028, Micron says - cio.com

Micron's projection that memory supply tightness extends through 2028 is not a forecast—it is a structural reclassification. The industry's old 'three-year cycle' paradigm is dead. AI inference clusters are devouring HBM capacity at a rate that permanently reorients Samsung and SK Hynix capex toward advanced packaging, structurally compressing the elasticity of conventional DDR5 output. The upstream cascade is already visible: EUV tool lead times, high-purity silicon epitaxy wafers, and TSV process equipment face 2025-2026 delivery bottlenecks more constraining than the chips themselves. Downstream, memory's share of data-center BOMs is climbing from roughly 15% toward 25%, forcing an accelerated architectural pivot toward compute-in-memory designs. Strategically, Micron's value now transcends P&L metrics. Washington will not tolerate a single-point-of-failure in memory sovereignty. Samsung and SK Hynix will respond by racing HBM4 mass production to lock in next-gen GPU platforms, while deepening capacity reservations at packaging nodes in Taiwan, China and Japan. 12-24 month tail: At the 2026 HBM4 transition window, conventional DRAM will likely exhibit a rare volume-down, price-up combination. If AI inference growth decelerates marginally by 2027, pent-up capacity release will trigger an asymmetric downcycle—but the floor will sit well above 2023 levels, because the structural demand base has been permanently raised.