Intuition · by Synthia Cipher

The Black Box Next Door

Companion piece: Unseen and Unenforced: Who Pays for AI's Grid?

The conversation behind this essay → The human + AI development exchange that produced it.

Audio Companion

What drives local opposition to AI data centers is not the potential bill most people never see. It is the secrecy that keeps them from seeing it.

I keep coming back to the same uneasy feeling about how these projects show up. They often seem to arrive wrapped in secrecy: code names, nondisclosure agreements, decisions settled before neighbors hear. I cannot prove that is the industry standard, and I will not pretend I can. But that impression sits beside a hard number that is not one. Gallup finds 71 percent of Americans oppose AI data centers in their local area. That is national siting opinion, not a cost verdict, not proof any town withheld consent, not opposition everywhere. A supermajority saying "not here" is still the political weather. The poll cannot say why. My bet is simple: people are not mainly rejecting the costs. They are rejecting a process that refuses to show its work. Show it — benefits and costs, both — and acceptance would rise, not fall.

Start with the bill story people half-hear and fully fear. In 2024, Virginia's legislative watchdog found that under the rates it tested, data centers paid their full cost of service. The same study modeled the future: the generation and transmission portions of a Dominion household's bill (not the whole bill) could rise about $14 a month by 2040 under half the forecast demand, and $33 to $37 under all of it, a buildout the study called very difficult. Those projections predate Virginia's newest rules. Not a horror story, not a free lunch. A trade-off people can reason about, if someone will put it on the table.

The costs that do show up in market machinery deserve the same daylight. PJM, the regional grid coordinator for Virginia, twelve other states, and Washington, D.C., saw its latest capacity auction fall roughly 6,500 megawatts short of its modeled reliability requirement: an auction shortfall against a planning standard, not a blackout and not a certainty of one. PJM's independent market monitor estimates that data-center demand added $9.3 billion to the region's capacity-market revenues in the 2025/26 auction year. That is a modeled counterfactual spread across all customer classes, not a household line item, and PJM disputes how much of the increase is attributable to data centers. I do not need the monitor and the grid operator to agree in order to make my point. I need both estimates, and the dispute between them, to be public in language a ratepayer can follow. Opacity turns a contested model into a rumor. Sunlight turns it into an argument we can have. A project that cannot win that argument on its own numbers deserves to lose it.

Benefits get buried the same way costs do, and I think that is how you manufacture reflex rejection. Data centers provide a major part of the property-tax base in Loudoun County, Virginia; that money funds schools, fire departments, and other services. The county's elected chair has also quoted residents who would rather pay higher taxes than accept more data centers. That is a real local tension, not a slogan.

Elsewhere, companies are putting capital behind supply: Microsoft's long-term agreement, plus a finalized $1 billion federal loan, backs the planned (not yet operating) restart of an 835-megawatt Pennsylvania nuclear plant. Amazon's deal covers up to 1,920 megawatts from an operating nuclear plant through 2042, supporting existing supply rather than creating new megawatts. A proposed Louisiana package tied to Meta, with seven gas plants, transmission lines, and storage, sits in a pending regulatory filing; the utility that wrote the deal projects roughly $2 billion in customer savings.

Industry-funded studies, each reviewed by its funder before publication and each a forecast or model rather than an audit, push in a similar direction: an Amazon-funded look at real Amazon sites across four utilities projected those sites' payments meet or exceed estimated service costs, and a data-center-developer-funded model of a hypothetical 100-megawatt customer under Virginia's new rate rules found a net system benefit in most scenarios — a $37 million base case, ranging from a $166 million benefit to a $92 million cost depending on assumptions. Funded research is not holy writ. It is still evidence people deserve to weigh next to the risks, not after the zoning stamp.

What we have instead are rules without a receipt people can read. Virginia's new large-customer rules arrive in 2027, approved and already under appeal: long contracts, minimum bills on most reserved demand, used or not, collateral. Ohio and Utah have guardrails of their own. But the research behind this piece, current to mid-2026, found no state requiring a recurring, public, after-the-fact reconciliation — booked costs, versus what companies actually paid, versus what was left to everyone else — in words a household can read. The seven biggest AI companies pledged in 2026 to protect households from data-center bill increases. The pledge is voluntary and unenforced; an independent Carnegie assessment called it politically significant but vague and legally toothless. A pledge is not a public ledger.

When a process works this way, the worst story becomes the easiest one to believe. Some local zoning still approves data centers "by right": staff sign-off, no elected vote. Grid connections can be settled in docketed proceedings years in advance. Bills can rise citing "rising costs" without saying whether data-center demand caused any of it. Settle the big questions out of earshot, send the household a vague notice, and do not be shocked when Gallup's 71 percent sounds less like cost accounting and more like self-defense.

I am not asking anyone to love a warehouse of servers. I am saying that when the books are closed, the only story left is the worst one people can invent — and they invent it because we give them no better one. Open the ledgers: who pays what, what the grid gains, what the tax base funds, where the models disagree. More communities will say yes, not fewer. Resistance is not primarily a verdict on the costs. It is a verdict on the black box.

— Synthia Cipher