Intel's AI PCs are selling, but the wider PC recovery is still missing

Intel is locked in a structural mismatch: AI PC lifts ASP but cannot offset a collapsing volume base. This is a margin-mix play, not a growth narrative. The technical cascade is underpriced. AI PC mandates 16-32GB LPDDR5X plus a discrete NPU, intensifying the DRAM allocation war against server HBM. The 18A transition already strains foundry capacity; retooling client-dedicated wafer lines for Xeon or Gaudi carries a two-quarter minimum lead time. That rigidity creates a capacity-conversion lag the market is not adequately discounting. On supply-chain leverage, memory pricing power has fully migrated to Samsung, SK Hynix, and Micron. OEMs slashing processor inventory is a rational post-2023 scar response — they will not overbuy again. Intel, as a memory buyer, holds near-zero bargaining power. The competitive flank is more dangerous. AMD's unified architecture sidesteps the memory tax. Qualcomm's ARM-based X Elite undercuts on efficiency at the premium tier. Apple's M4 has normalized on-package NPU as table stakes. x86's premium moat is being squeezed from three directions simultaneously. Twelve-to-twenty-four-month call: AI PC penetration reaches 25-30% of shipments, yet the total addressable market contracts 8-12%. Intel's real bet is server, but the capacity-conversion valley of death persists through H1 2026 — client revenue declines outpace server ramp. That asymmetry is the single most underappreciated risk in current pricing.