Commentary: AI investment doubts rise, but chip demand keeps surging

The grid bottleneck isn't killing the AI thesis—it's forcing a structural shift from scale-driven to density-driven compute economics. When billions in capex freeze at permitting and interconnection stages, the cascade hits the tech stack's foundation: CoWoS advanced packaging transitions from a premium feature to a survival constraint, while HBM stacking is pushed toward 12- and 16-layer configurations to extract bandwidth from constrained die area. Demand hasn't vanished; it's been repriced by physics. On compliance, power permitting delays function as a hidden entry tax—invisible on any tariff schedule yet more effective than export controls at redrawing the geographic map of compute deployment. Firms are pivoting CAPEX from build to lease, converting sunk costs into recurring opex and systematically compressing margins for mid-tier cloud operators. Strategically, NVIDIA will likely accelerate in-house packaging to reduce foundry dependency. AMD exploits the window where rivals are bottlenecked by infrastructure to capture CoWoS allocation. Intel's 18A node lands precisely in that gap—while competitors wait for electrons, it pushes process nodes. Over the next 12 to 24 months, compute per watt replaces total compute as the industry's defining KPI. Deferred datacenter demand migrates toward edge inference and heterogeneous architectures. Power cost enters chip pricing as an explicit variable for the first time—the most profound cost-structure restructuring since Moore's Law.