The AI Constraint Relay: How the Compute Build-Out Displaces Physical and Financial Limits

A supply-side risk assessment of the AI compute build-out, method-first rather than a crash call. Version 5 corrects US Bureau of Reclamation 24-Month Study attributions, documenting the vintage mismatch across traces (July 2026 Min Probable, May 2026 Max Probable; canonical August file remains August 2025) while confirming that the 3,490 ft minimum power pool breach remains deferred past 2026 with the July 2026 Probable Minimum trace projecting 3,503.65 ft at year-end (roughly 13.7 ft clearance, 36 percent unregulated inflows, 6.00 maf releases). Version 4 reframed the core thesis from a static 'barbell' to 'the constraint relay', documenting continuous hydraulic displacement and moving bottlenecks across the infrastructure stack. When logic availability scaled, the constraint shifted to grid interconnections; price collars at PJM ($325 per MW-day) converted price signals into a 6,831 MW physical volume deficit; volume deficits triggered administrative rationing (PJM BYOP rules) and pushed developers to on-site generation; behind-the-meter generation hit multi-year turbine manufacturing backlogs (GE Vernova, Siemens Energy) and single-crystal superalloy blade casting queues (Howmet); substation energisation and switchyard builds hit an inelastic, non-tradable 5-year apprenticeship pipeline for certified high-voltage electricians and linemen; and HBM wafer reallocation displaced price spikes onto commodity DRAM (+54 to +116 percent) and consumer electronics. The assessment audits non-relocatable tail risks (Taiwan Strait, licensed rare earths, Spruce Pine high-purity quartz) alongside financial underwriting and single-campus reinsurance accumulation caps (15 to 40+ billion USD TIV). Sourced and dated constraints in section 2 carry the argument; a companion reproduce.py tabulates and cites every figure from supply_constraints.csv. Independent analysis, not investment advice. Disclosure: produced with assistance from Claude Opus 5.0 (Anthropic) and Gemini 3.7 (Google DeepMind / Alphabet), both of which appear as subjects in this analysis.