How We Got Here: The AI Race Left The Spreadsheet

The early AI boom was tidy. Models lived in blog posts. Scaling laws lived in charts. “Infrastructure” was something you mentioned in a funding deck to look serious.

Then data centers showed up in actual places with actual electricity bills, and the public discovered the fun part of “innovation”: it has a physical footprint. Suddenly the argument isn’t “is AGI real,” it’s “why is my substation a celebrity now?”

Throttle Point #1: Voters Are Turning Zoning Into Compute Policy

MultiState’s rundown of post‑Maine dynamics reads like the early days of fracking politics: first it’s a local hearing, then it’s a ballot measure, then it’s an industry PR campaign with the emotional tone of a hostage note.

After Maine Gov. Janet Mills vetoed LD 307 (a state data-center moratorium), MultiState notes that communities are increasingly moving to ballot measures as an alternative way to regulate development. One highlighted example: an Ohio initiative effort that would prohibit construction of data centers requiring 25MW or more of power, with signature collection targets and a real deadline. At the local level, the article points to measures like Port Washington, Wisconsin requiring voter approval before tax incentives for data centers can be granted, plus various bans/referenda attempts across states.

Mother Jones frames the broader pattern: moratorium bills and grid‑strain politics are spreading beyond town meetings and into state legislatures, with the AI infrastructure argument turning into a straight-up question of who pays, who benefits, and who gets to say “no.”

Throttle Point #2: The Electrical Supply Chain Is Out Of Stock (For Years)

Even if every community rolled out the red carpet for new capacity, the supply chain still has a bottleneck that does not care about your quarterly guidance.

Data Center Knowledge, citing a Wood Mackenzie report, describes the US data-center electrical equipment market scaling from roughly $20B (2026) to $65B (2030), driven by rising demand for transformers, switchgear, and power distribution systems as hyperscale construction accelerates. The same piece cites US data-center capacity expected to scale from about 24GW to 100GW between 2026 and 2030—meaning data centers stop being a “customer segment” and start being the reason factories reorganize their lives.

And then comes the line that should be printed on every “AI will eat the world” slide: substation transformer lead times stretching to 160+ weeks in 2026 (up from ~140 weeks in 2023). That’s not a delay. That’s a new planning horizon.

Wood Mackenzie’s forecast also includes transformer demand rising from about 1,500 units annually to more than 9,000 by the end of the decade. If you were wondering whether the buildout is “real,” here’s your answer: it’s real enough to bend manufacturing capacity.

The Non-Obvious Thing: Compute Is Becoming A Financeability Problem

Put the two throttle points together and you get a pattern that doesn’t show up in model evals:

• Politics creates uncertainty. Ballot measures and moratoria don’t just slow approvals—they make timelines non‑deterministic. That raises the cost of capital.
• Hardware creates delay. Even “approved” projects stall while waiting for equipment with multi‑year lead times.
• Uncertainty multiplies. The longer the queue, the more time you give opponents (and elections) to catch up.

This is how infrastructure gets regulated in practice: not by a single “AI law,” but by a stack of constraints that turns “we want to build” into “we can’t close financing until the transformer contract is real and the referendum isn’t.”

The Singularity Soup Take

What to Watch