The AI infrastructure boom is clashing with physical constraints, transforming industry economics. While inference costs per token have dropped ~10x, global AI-driven data center power demand is projected to reach ~950 TWh by 2030—roughly the current annual consumption of Japan [5] [13] [16].
The Jevons Paradox in AI
Increased hardware efficiency is triggering the Jevons paradox: cheaper per-token costs drive demand faster than savings can offset, causing total energy consumption and infrastructure expenditure to climb despite model optimizations [5] [16].
The New Bottleneck: The Grid
Power availability has replaced chip supply as a primary constraint. With U.S. data center power demand expected to rise ~755% by 2029, utilities face critical grid interconnection delays of three to seven years [3] [9]. Consequently, large-scale capital is shifting downstream toward power infrastructure companies and on-site generation solutions [3] [10].
Economic Outlook
Enterprise ROI remains pressured by rising DRAM/HBM costs—up ~90% in early 2026—and the necessity of navigating constrained power markets [7] [10]. As companies transition from pilot to production, "cost per token" has become the defining metric for operational viability in an era of persistent physical scarcity [5] [11].
Sources
- Cost per token: the metric defining AI factory economics – Schneider …
- The Coming AI Blackout: $10 Trillion Data Center Surge Threatens …
- AI Demand Is Still Supply-Constrained: Why the Infrastructure Boom …
- AI infrastructure bottlenecks and power constraints 2026
- International Energy Agency projects AI data center power demand …
- US Utilities Plan $1.4T for AI Data Centers: 27% Capex Surge [2026]
- Tokens per megawatt-hour: the energy economics of frontier AI
- AI Energy Statistics 2026: Data Center Power & Facts
- Goldman Sachs Warns: AI Energy Demand Equivalent to "Building Another …
- Goldman Sachs warns: AI's energy demand is equivalent to "recreating …