Siting Risk Is the New Compute Bottleneck — and It Belongs in Your Capacity Plan

For three years, the implicit assumption behind every AI capacity roadmap was that the hard part was silicon. Get the GPUs, and the rest — land, power, water, the local zoning board — would sort itself out. That assumption just broke in public.

The constraint moved

New York’s legislature passed a one-year moratorium on new large data centers, the first statewide ban of its kind, now awaiting Governor Hochul’s signature. The bill defines “large” as any facility with a peak draw of at least 20 megawatts, and it forces proponents to fund and hold public hearings three months before approval while the state’s environmental agency produces a report on electricity, water, land, and pollution impacts. The Verge’s reporting notes the New York grid operator is currently reviewing 24 proposals totaling more than 9,000 megawatts. That is the pipeline the moratorium freezes.

Out west, Kevin O’Leary’s “Stratos Project” in Utah got the same treatment from the other direction. Under pressure from Utah Senate President J. Stuart Adams, who demanded a 75% cut, O’Leary agreed to shave nearly 20,000 acres off the footprint. The remaining site is still larger than Manhattan, and the fight over water diversion to the shrinking Great Salt Lake isn’t over. The point isn’t the acreage. The point is that a sitting state legislator could unilaterally renegotiate the physical scale of a marquee compute project, and did.

These are not NIMBY footnotes. They are the binding constraint reasserting itself. You can buy H100s with a purchase order. You cannot purchase community consent, and you cannot provision around a moratorium.

Where the capital is voting with its feet

Watch where money goes when permission gets expensive. Blackstone-backed AirTrunk just committed $30 billion to build 5GW of capacity in India by 2030 — a country actively courting the buildout with tax exemptions for foreign cloud providers running through 2047, and a proposed 3GW facility in Maharashtra alone. India’s projected jump from ~1.5GW to as much as 8GW isn’t just market growth; it’s a regulatory arbitrage. Modi shakes hands with the CEO; a Long Island business association gets a one-year freeze. Capital reads that difference instantly.

The other escape hatch is stranger. Google agreed to pay SpaceX $920 million a month for compute — roughly 110,000 NVIDIA GPUs — mirroring a $1.25 billion/month Anthropic deal for Colossus resources. When the marginal terrestrial megawatt comes with a three-month hearing requirement and an environmental impact report, leasing someone else’s already-sited capacity (or eyeing orbit) starts to pencil out. The compute is fungible. The location, increasingly, is not.

What this means if you plan capacity

Siting risk has been treated as someone else’s problem — a real-estate or government-affairs line item that surfaces, fully resolved, as “we have a region.” That decoupling is no longer safe. If your 2027 roadmap assumes a specific cluster lands in a specific place on a specific date, that date now carries political variance you probably haven’t modeled.

Concretely, the things that used to be invisible inputs are now the volatile ones. A capacity plan that looks like this is lying to you:

capacity_plan:
  gpus_ordered: 32000
  delivery: 2027-Q1
  region: us-east-grid-3   # assumed available
  power_mw: 240            # assumed grantable

The honest version treats placement as a probabilistic dependency with its own failure modes, the same way you’d model a vendor SLA:

capacity_plan:
  gpus_ordered: 32000
  delivery: 2027-Q1
  siting:
    primary: us-east-grid-3
    status: permit_pending      # not yet granted
    moratorium_risk: high        # NY-style freeze
    interconnect_queue_months: 30
    water_constraint: true
    fallback: [india-mumbai, leased-colossus]
    political_dependency: state_legislature

The differences that matter: an interconnection queue measured in years, not weeks; a moratorium_risk flag that can zero out a region with one signature; and an explicit fallback list, because single-region siting is now single-point-of-failure planning. If your only Plan B is “appeal to the board,” you don’t have a Plan B.

Engineering implications, not just spreadsheet ones

This pushes back into architecture. Designs that assume one fat, co-located training cluster are betting on exactly the kind of 20MW-plus site that now triggers public hearings. Workloads that can tolerate geographic distribution, asynchronous checkpointing across regions, or bursting into leased capacity hedge against a freeze in any single jurisdiction. The 90-day termination clause in the Google–SpaceX deal is a tell: even the biggest buyers want optionality, not lock-in to one supply of compute.

It also raises the value of efficiency in a way that hype cycles obscured. When you could always add another megawatt, squeezing more tokens per watt was a nice-to-have. When the next megawatt requires a three-month hearing and an impact report on the Great Salt Lake, every percentage point of utilization, every batch you don’t recompute, every model you can serve on smaller hardware is effectively a capacity expansion you didn’t have to permit. Resource discipline is no longer just a cost story; it’s a regulatory-exposure story.

The new question to ask

The reflexive question of the GPU era was “can we get the chips?” The question for the next phase is “where will they be allowed to run, and who can say no?” New York can say no. A Utah state senator can say no. The Long Island Association can warn that a blanket ban “prevents case-by-case evaluation” and still lose. Meanwhile India and SpaceX are saying yes, loudly, and capital is responding.

If you build or plan AI systems, siting risk now deserves a row in your dependency graph, an owner, and a fallback — the same seriousness you’d give an upstream API that might rate-limit you. Treating it as infrastructure’s problem is how a roadmap quietly becomes fiction. The wall AI just hit isn’t made of silicon. It’s made of permits, water tables, and people who get a vote.