Pricey local AI machines arrive as memory costs threaten PC shipments

The memory price spike is not cyclical—it is structural. AI compute demand is cannibalizing DRAM/HBM wafer capacity at Samsung, SK Hynix, and Micron, starving the consumer DDR5 pool. Gartner's shipment downgrade is the symptom; the real signal is a permanent repricing of compute density across the PC stack. This cost pressure is accelerating architectural divergence. Apple's unified-memory Mac Studio is the only economically viable path to local 70B-parameter inference at consumer price points. Nvidia's RTX Spark bets on discrete GPU plus high-bandwidth VRAM, but the PC driver stack and thermal design remain immature—this is brand positioning, not volume. AMD's Ryzen AI Max NPU ceiling caps it at lightweight inference. Upstream, memory makers have seized pricing power; OEM BOM structures are permanently altered. Strategic intent matters more than product specs. Nvidia is anchoring the 'local AI requires a discrete GPU' narrative to set up RTX 50-series consumer cards. Apple is making local LLMs the default macOS story, forcing rivals to fight inside Windows. AMD's window is under 12 months—if NPU ecosystem momentum does not materialize, Ryzen AI Max becomes a transitional SKU. 12-to-24-month outlook: 32GB becomes the AI PC entry floor, 64-to-128GB the mainstream. Cloud inference's marginal cost advantage erodes as local hardware amortizes, accelerating enterprise de-clouding for inference workloads. Memory prices will partially normalize in H2 2026 with new capacity, but the AI PC ASP floor has shifted irreversibly upward.