Metalligence
Preparing the page.
Loading the latest research and market context.
Metalligence
Loading the latest research and market context.
Artificial Intelligence
An examination of how shifting infrastructure constraints, critical-path dependencies, and geographic factors determine the speed and location of AI compute deployment.
Core judgement
Artificial intelligence infrastructure is entering a phase in which physical constraints extend beyond semiconductor availability to electricity, grid connections, equipment, cooling, construction, and network capacity. This report argues that the principal limitation is not necessarily global resource scarcity, but the ability to deliver interconnected infrastructure at the required location and time. Using public datasets, regulatory evidence, industry research, and scenario analysis, the study identifies grid interconnection, local power capacity, and electrical equipment as the strongest current constraints. It introduces Time-to-Compute as a central analytical variable, emphasizing that infrastructure deployment depends on the longest critical path rather than individual component availability. The research also examines how efficiency improvements, expanding inference workloads, and geographic relocation can redistribute bottlenecks instead of eliminating them. Water availability, material processing concentration, permitting, and digital sovereignty create additional location-specific constraints. Scenario stress tests and counterfactual analysis explore how resolving one constraint may expose another, reshaping AI infrastructure development and global industrial geography.
Key findings
1. Grid Interconnection and Equipment Are the Leading Constraints
The report identifies grid access, local power capacity, and electrical equipment delivery as the strongest current bottlenecks, with IEA evidence indicating grid development timelines of 5–15 years compared with approximately 1–3 years for data centers.
2. Compute Demand Growth Outpaces Hardware Efficiency Improvements
Analysis of publicly available datasets finds that median AI training compute approximately doubles every 0.49 years, compared with 1.43 years for hardware peak performance per watt, although the two measures are not directly comparable as electricity-demand forecasts.
3. Time-to-Compute Determines Infrastructure Deployment Speed
Across 83 observed data center sites, the median interval from the first publicly recorded construction milestone to initial operation was 773 days, illustrating the importance of infrastructure delivery timelines while excluding earlier development and permitting stages.
4. Water and Materials Present Distinct Forms of Supply Risk
Water availability can impose location-specific constraints, while the concentration of copper refining, aluminum smelting, and silicon production indicates supply-chain vulnerabilities rather than evidence of imminent global resource exhaustion.
5. Geographic Relocation Creates New Network and Sovereignty Constraints
Full research
Public previews do not expose original DOCX or approved PDF files. Entitled clients can access the full research through the protected reader as it is introduced.
Client sign in →Relocating AI infrastructure may reduce certain electricity, land, and construction pressures but introduces additional challenges involving network latency, connectivity, data residency, export controls, and national sovereignty.
6. Resolving One Bottleneck May Expose Another
Scenario stress tests and counterfactual exercises suggest that alleviating the most restrictive infrastructure constraint can shift pressure toward other components, supporting the report's framework of rotating bottlenecks without constituting empirical forecasts.