Where should AI compute live?
The best site is a compromise among power, fiber, latency, cooling, regulation, tax treatment, labor, and the workload that will run there.
Geography becomes binding when the cheapest power is too far from users, the best fiber is too power-constrained, or the fastest permit is in the wrong thermal climate.
Training and inference want different maps
Training can live farther from end users because the job is internal, long-running, and batch-like. Inference often wants to sit closer to demand because latency changes product quality.
That split creates two geographies: giant training campuses near power and cooling advantages, plus inference capacity distributed closer to users and enterprises.
Fiber is the second utility
AI sites need high-capacity fiber for model checkpoints, data movement, replication, and user traffic. A cheap-power site without enough connectivity can still be a poor compute site.
The strongest regions combine power access, fiber routes, land, permitting familiarity, and an ecosystem of operators who know how to build.
Latency is product-specific
A coding agent can tolerate more latency than a voice assistant. A real-time video model has a different envelope from overnight document processing. Physical AI adds still another requirement when control loops touch the world.
The right geography starts with the product experience, then works backward to where capacity can sit.
Politics prices the site
Large AI loads change local grids, water plans, tax bases, and public politics. A state that offers fast permitting but socializes too much cost may trigger backlash. A state that gives clear large-load tariffs may become more attractive, not less.
Regulatory predictability is part of capacity.
The map will specialize
Expect training, inference, storage, disaster recovery, and enterprise-serving regions to diverge. One mega-campus will not serve every job equally well.
The winners will treat geography as architecture: put each workload where its binding constraint is easiest to satisfy.