AI data centers are building on-site power plants to bypass the overloaded grid and meet growing electricity demand, amid permitting and supply chain challenges
www.wsj.com/business/ene... @martenmickos : America should be adding about 80 gigawatts of new power generation capacity a year to keep pace with digital demands. It's currently building less than 65 gigawatts. — Last year, China added 429 gigawatts and USA 50. — www.wsj.com/business/ene...
Context & Ripple Effects
This is part of a widening power-access problem for AI infrastructure: related coverage has documented grid bottlenecks from data-center and crypto demand and warned that rising data-center load could strain North American reliability. Earlier pressure on an overtaxed grid makes generation capacity—not just computing hardware—a constraint on new facilities.
The move toward self-supply also aligns with reports of developers turning to aeroderivative turbines and diesel generation amid long grid waits and AI labs deploying on-site gas generation. The reported gap between needed and current US annual generation additions makes permitting and equipment availability central to whether this workaround can scale.
First-order effects
- AI data-center projects can reduce dependence on delayed grid interconnections by pursuing dedicated generation, but they assume new permitting, construction, and fuel- or equipment-procurement work.
- Power-plant permitting and supply chains become immediate gating items for data-center openings, shifting execution risk from utility connection queues to on-site generation delivery.
Second-order effects
- Developers competing for the same sites will face stronger incentives to secure powered capacity and generation equipment early, while utilities confront large loads that may arrive partly outside conventional grid-expansion timelines.
- Demand for turbines, generators, and associated project-development capacity may tighten further as more campuses emulate the use of on-site gas generators by AI labs, potentially raising the cost and duration of buildouts.
Third-order effects
- If self-powered campuses persist, AI infrastructure will increasingly be planned as a combined compute-and-energy project rather than a facility that simply buys grid service.
- The lasting constraint may shift toward the pace at which permits, generation equipment, and transmission upgrades can be delivered; on-site generation can bridge grid delays but does not eliminate the broader need for electricity infrastructure.
The trend: AI buildouts are turning reliable power access into a core determinant of where, when, and at what cost new compute capacity can be deployed.