Coverage of AI increasingly concerns substations, water permits and county planning meetings. The shift is not a change of subject but a consequence of what serving these models physically requires.
Density changed the electrical problem
Conventional server halls draw a moderate amount of power spread across a large floor area.
Accelerator racks concentrate far more power into the same footprint, so a building of unchanged size can require several times the electrical supply.
That turns a facility from an ordinary commercial connection into something closer to an industrial load, and industrial loads are negotiated with the grid operator rather than simply requested.
Grid connection is the long pole
Interconnection queues in many regions run for years, because each new large load requires study and often new transmission equipment.
Operators respond by siting near existing generation, by contracting directly with power producers, or by installing their own generation on site.
Each of those choices carries a local consequence, which is how these projects enter local politics rather than staying in trade press.
Cooling turns into a water question
Removing concentrated heat is done either with water evaporation, which is efficient and consumptive, or with closed systems that use more electricity and little water.
The choice is usually made on cost and local conditions, and in water-stressed regions it becomes the point of public objection.
Liquid cooling delivered directly to the chip is displacing air cooling in new builds, which reduces total demand without removing the question.
Local tax arrangements draw scrutiny
Facilities are frequently attracted with property and sales tax abatements agreed before construction.
Because these buildings employ relatively few people once operating, the promised local benefit is smaller than the headline investment implies.
Residents who see rising electricity costs and limited employment increasingly contest the agreements, and several jurisdictions have begun attaching conditions.
Siting decisions constrain what gets built
The practical effect is that capacity appears where power is available rather than where demand is, which is why clusters form in a small number of regions.
Distance from users adds latency, so some workloads must sit closer regardless of cost, splitting infrastructure into large remote training sites and smaller distributed serving ones.
The pace of model deployment is therefore partly set by permitting timelines, which is an unusual dependency for a software industry and a durable one.