Pax Silica Can't Build a Substation

Pax Silica is a customs and tracking framework. Launched this week, it expedites the movement of AI supply-chain goods across borders for roughly two dozen partner nations ^[1]. It says nothing about power grids, water systems, or local zoning boards. In the United States, 3,969 AI data centers have been announced. Eight hundred and two are under construction. The other 80% are stalled — by shortages of specialized labor, building materials, and TSMC chips, and by grid constraints that no trade agreement can resolve ^[2]. At the township level, the resistance is concrete and costly. In a single quarter — April through June 2025 — US communities blocked or delayed $98 billion in AI data center proposals across 11 states ^[3]. The $16 billion Stargate data center in Michigan proceeded only after developers sued the township to overturn a rezoning denial and paid $14 million in community benefits ^[4]. In Ohio, the state Supreme Court ordered the village of Ashville to restore a ballot referendum on a proposed data center, ruling that the village council's declaration of an "emergency" to bypass public voting was improper. The ruling signals that the emergency-bypass tactic developers have used to fast-track projects may not survive judicial review ^[5]. At the state level, the ceiling hardens into policy. New York became the first state to enact a statewide moratorium on certain AI data center developments, and Governor Kathy Hochul is additionally repealing sales tax exemptions to deliberately increase costs for developers ^[4]. In Florida, AI data centers have become a central flashpoint in the gubernatorial race, with candidates split between regulated growth and outright moratoriums ^[6]. California, Maine, and South Carolina are actively moving to ban or moratorium data center construction ^[7]. The Ohio EPA reversed a proposed blanket wastewater permit for data centers after receiving over 7,000 public comments and bipartisan pressure ^[8]. And in Virginia — the state once cited as proof the buildout was proceeding — the Department of Environmental Quality concluded the Potomac Aquifer cannot support new industrial water withdrawals, while Goldman Sachs analysis shows grid constraints in the PJM Interconnection are actively pushing data center developers out of the state ^[9][10]. At the federal level, the ceiling is deliberate abdication. The White House stated its position plainly.

Communities know what's best for them, and the Administration is clear that local infrastructure decisions remain with states and localities. — Franklin D. Roosevelt

Congress pulled its own streamlining legislation, the Protect American AI Act, after Representative Thomas Massie opposed it within his own party ^[11]. The congressional spectrum on data centers now runs from Senator Josh Hawley's bipartisan bill requiring independent power sources to Senator Bernie Sanders and Representative Alexandria Ocasio-Cortez proposing a total moratorium — with no faction arguing for federal preemption of local authority ^[11]. House Energy and Commerce Chair Brett Guthrie captured the uncertainty.

I'm not sure if it's going to be a federal law, or if each local community can do it themselves. — Brett Guthrie

At the supply-chain level, the ceiling is arithmetic. Eaton Corporation reports a US data center infrastructure backlog of 307 gigawatts — representing 15 years of work at 2025 build rates ^[12]. That means transformers, switchgear, and cooling systems. Even the workaround — AI hyperscalers building behind-the-meter gas turbines, as Meta is doing with 10 company-funded plants generating 7.5 gigawatts at its Hyperion center in Louisiana — hits its own wall: GE turbines are sold out through 2029 ^[13]. US manufacturing faces a critical shortage of skilled technicians at a seven-year high in Ohio, with automation simultaneously eliminating the entry-level roles that traditionally trained new workers ^[14]. The same ceiling appears in India, where OpenAI is scouting a 1-gigawatt Stargate data center and where AMCHAM and Applied Materials have jointly identified the electricity grid as the binding constraint ^[15][16]. AMCHAM Director General Ranjana Khanna made the point directly.

We urge energy policymakers to consider that the electricity grid is the single largest constraint on India's AI ambitions. — Ranjana Khanna

India's AI data center capacity is projected to surge from 1.5 gigawatts to 13.5 gigawatts by 2031-32 — a ninefold increase demanding an additional 40 to 45 terawatt-hours of electricity by 2030 across five major cities ^[17]. The grid to deliver that power has not been built. Even if every political objection vanished tomorrow — every zoning board, every state moratorium, every ballot referendum — the 307-gigawatt equipment backlog means the physical supply chain for transformers, switchgear, and cooling systems would take a decade and a half to clear. The ceiling is not diplomatic. No trade agreement can build a substation, lay a water main, or compel a township to rezone.