AI Infrastructure

The build-out of infrastructure for artificial intelligence is among the largest capital deployment programmes ever undertaken by private industry. The largest cloud and AI companies have, individually, committed capital expenditure running into tens of billions of dollars per year — and collectively far more — directed at data centres, chips, networking and, increasingly, the power supply needed to run them. Individual AI data-centre campuses are now planned at gigawatt scale, a unit previously reserved for national electricity planning. Whether one views the demand forecasts as well-founded or exuberant, the physical commitments are real, and they are being made at speed.

Speed is where execution risk enters. The constraints are not abstract. Electricity: grid operators in the United States and Europe face connection queues measured in years, and AI load growth is colliding with electrification demand from other sectors. Equipment: transformers, switchgear and backup generation have lead times that have stretched substantially since 2021. Construction: the labour and contractor base capable of delivering mission-critical facilities at this scale is finite. Permitting and community acceptance: water use, noise and land impacts are generating organised opposition in multiple markets. And financing: a growing share of the build-out is funded through debt, special-purpose structures and long-dated commitments whose value depends on AI demand materialising broadly as forecast — an assumption that prudent boards should treat as a scenario, not a certainty.

This category investigates the delivery reality of the AI infrastructure build-out: which projects are actually being completed, at what cost and on what schedule; how power procurement is reshaping where capacity lands; how risk is being allocated between hyperscalers, developers, utilities and lenders; and where the gap between announced capacity and delivered capacity is widest. PIA's approach is documentary — filings, planning records, grid-operator data, regulator publications — and deliberately sceptical of announced-figure arithmetic. Announcements are cheap. Delivered megawatts are not. The investigations here track the difference.

AI Infrastructure — Frequently Asked Questions