AI PCs face memory squeeze as server demand takes priority

This is not a supply shortage—it is a priority war. When AI inference migrates from cloud to edge, DRAM makers face a capital-allocation conflict unseen since the 2018 GPU crunch. The 128GB unified-memory architecture demands simultaneous breakthroughs in per-channel bandwidth and density, a fundamentally different engineering path from HBM stacking. Samsung, SK Hynix, and Micron have already locked 2024-25 capex into HBM3E/HBM4. High-density DDR5 for consumer PCs is structurally underserved. Worse, unified memory forces CPU-GPU shared pools with near-server ECC and timing specs—PC OEMs now procure quasi-server components at consumer price points, breaking their BOM economics. Microsoft's vertical integration offers modest leverage, but hyperscaler annual contracts still rank above PC OEMs in allocation priority. Apple's M-series validated the architecture while pre-committing premium wafer capacity to its own ecosystem. Intel and AMD will likely differentiate via CXL memory expansion, sidestepping single-die density constraints. Over the next 18 months, expect a bifurcated DRAM market: HBM holds premium pricing while high-capacity DDR5 for PCs experiences violent price swings due to fragmented demand. The 128GB AI PC remains a niche—enterprise and creator—unless memory fabs dedicate a mid-bandwidth, high-capacity line for consumer use. Under current ROI logic, that probability is near zero.