Scaling the Unsustainable?: Technical Architecture, Energy Sustainability, and Scalability Constraints in Multinational AI Data Center Megaprojects

The contemporary expansion of artificial intelligence (AI) rests on a wager that computational scaling laws will keep converting additional compute into additional capability. This dissertation asks whether the physical infrastructure required to sustain that wager is technically and environmentally scalable. Using a quantitative multi-case technical sustainability assessment, the study evaluates six of the most consequential multinational AI data center megaprojects announced through 2026: the Microsoft-OpenAI partnership in Wisconsin, the Oracle-OpenAI compute agreement in Abilene, Texas, the Stargate consortium, Meta's hyperscale expansion in Louisiana and Ohio, the aggregate hyperscaler capital-expenditure surge, and the NVIDIA-OpenAI systems partnership. A six-component analytical framework, comprising energy demand modeling, cooling and power usage effectiveness analysis, carbon-footprint estimation, water-footprint analysis, hardware architecture and scalability analysis, and bottleneck or constraint analysis, is applied to each project using exclusively public data from the U.S. Energy Information Administration, the U.S. Environmental Protection Agency, the U.S. Geological Survey, the Electric Reliability Council of Texas, corporate sustainability disclosures, and semiconductor specifications. The analysis finds that five site-assigned facilities alone imply approximately 7.2 gigawatts of new electrical demand and 44.2 terawatt-hours of annual consumption, generating roughly 18.3 million metric tons of location-based carbon dioxide each year on grids that remain, on average, roughly three-quarters fossil-fueled. Although liquid cooling reduces power usage effectiveness to between 1.08 and 1.25, deployment scale overwhelms these efficiency gains, empirically confirming a Jevons paradox dynamic in which per-operation efficiency rises while absolute resource consumption climbs. Grid interconnection and generation emerge as the systemic binding constraint, most acutely in Texas, where reserve margins are projected to turn sharply negative. The study concludes that efficiency improvements are necessary but insufficient to render frontier compute sustainable, and that absolute resource accounting must anchor infrastructure planning. It contributes a replicable, evidence-status-aware framework and completes a companion trilogy alongside governance (Pokorny, 2026a) and economic-geopolitical (Pokorny, 2026b) analyses.