Site Selection, Power Procurement & Permitting

Site selection used to ask 'where is the cheapest land near fiber and users?' — in the AI era it asks 'where can I energize 100 MW to multiple gigawatts in 24–48 months?', and getting that ordering wrong strands a billion-dollar campus on a slab that cannot be moved.

Power is not a commodity you buy — it is a queue position you win, a study you survive, and a long-lead transformer you order years before the slab is poured; the interconnection process, not the construction schedule, is the real critical path of an AI data center.

Energizing a site tells you whether you can run the machine; the power-cost structure tells you whether you should — because at 25–60% of opex, the price, shape, and curtailment terms of those electrons are the single largest controllable lever on lifetime return, and the cheapest headline rate routinely hides the most expensive bill.

Energy supply is a portfolio decision, not a procurement line: the structure you pick — grid-only, grid-plus-bridge, behind-the-meter island, or co-located hybrid — locks in your speed-to-power, your cost certainty, and a set of basis, shape, and regulatory risks that outlive the GPU generation it was meant to power.

When power is the binding constraint, on-site generation stops being a backup afterthought and becomes the primary energy-supply decision — and the fork you take (gas engine vs turbine vs fuel cell vs nuclear, bridge vs permanent, islanded vs grid-parallel) sets your speed-to-power, your fuel-supply critical path, your emissions exposure, and whether the asset you build to escape the queue becomes the stranded asset you cannot demote.

Fiber is a secondary screen for a training campus and a primary one for an inference fleet — but for both, the binding number is the same physics: roughly 5 microseconds of one-way delay per kilometer of glass, which decides whether two campuses can train as one machine and whether a metro can serve a sub-50 ms request, and the diligence that decides whether your power-and-land-perfect site is in fact functionally stranded.

Water is the second hardest siting gate after power: it decides your cooling architecture, your permit timeline, and whether you are first or last in line when a drought forces curtailment — so you screen for it before you sign the land, not after.

Land, soil, seismicity, and flood line are a secondary screen — they almost never win a site, but a missed liquefaction zone, an under-rated slab, or a parcel that floats in the 500-year floodplain will silently subtract 18 months and tens of millions before the first transformer arrives.

Permitting is not paperwork you process after siting — it is a parallel, multi-jurisdiction critical path where the two longest poles (grid interconnection and the air permit for backup generation) routinely outrun construction, so a project that breaks ground on schedule can still sit dark, and the only defense is to start the slowest approvals before you own the land.

Incentives are a second-order tiebreaker that decides nothing about whether you can build and almost everything about whether the build pencils — but the 10-year abatement you bank against a 20–30 year asset is the most reversible figure in your entire pro-forma, and 2026 is the year it started getting clawed back.

Social license is now a hard gate on the critical path, not a community-relations afterthought — the project that loses the room loses the rezoning, and the rezoning you lose strands an interconnection slot and a slab you have already begun to underwrite.

In 2026 the question is no longer just where you can get megawatts — it is whose chips you are allowed to plug into them, whose law can reach the bytes that run on them, and whether the answer changes the day an administration does; geopolitics has become a hard site-selection gate, not a footnote.

Every preceding siting chapter gave you one screen in isolation; this one is where they collide on a real map — the market you pick pre-loads your power, your permitting risk, your cooling physics, and your sovereignty constraints before you score a single parcel, so the scoring playbook exists to keep a tempting cluster from passing a gate it has already failed.