Unseen and Unenforced

Audio Companion

A Virginia household's electricity bill goes up. The notice cites rising costs but not whether new regional data centers caused any increase, or how much. A data-center project may be miles away; its grid connection may have been approved years ago in a proceeding identified by a docket number — its official case number — rather than a name. The filings in those proceedings discuss data-center demand at length. The pieces of an answer are scattered across those case files. No rule found in this research requires anyone to assemble them where a household could read them.

Households did not ask for a better explanation of a higher bill. They wanted the bill not to rise because of a private company's new demand.

The largest AI companies offer an admirably simple answer: we will pay our own way. Seven — Amazon, Google, Meta, Microsoft, OpenAI, Oracle, and xAI — signed a March 2026 public pledge to protect household customers from exactly this. That sounds fair. Questions remain:

Which parts of that promise have become rules a company can be held to? Which costs do those rules cover? And how would anyone outside the deal know whether it was kept?

In late 2024, Virginia's watchdog found two truths that travel together. First: the rates it tested allocated costs appropriately — including to data centers, which were paying their full cost of service. Second: the same study modeled a different future. For a household served by Dominion, Virginia's largest utility, and using 1,000 kilowatt-hours a month, the generation and transmission portions of the bill — generation means producing power; transmission means moving it over high-voltage lines — could rise about $14 a month by 2040 if utilities served half of the study's unconstrained demand forecast — growth if nothing limits it, a buildout the study called difficult — and about $33 to $37 if they served all of it — which it called very difficult. Fair under the 2024 rules the study tested; exposed in the future it modeled. Both halves belong in view.

The pledge is voluntary, with no built-in consequence if a company breaks it. A promise, standing where a rule should be. In Virginia — the state where this bill arrives — it does not stand alone: the state had already begun putting enforceable rules beneath parts of what the companies would later promise.

Deciding whether to build is entirely distinct from deciding who pays for it. Here is the strong case for building—because even the most heated version of this debate tends to understate it.

The grid needs more dependable power. In PJM — the regional grid coordinator for Virginia, twelve other states, and Washington, D.C. — the latest capacity auction, which PJM uses to secure commitments for future power availability, fell about 6,500 megawatts short of the region's modeled reliability requirement. That is an auction shortfall against a planning standard — not a blackout or a certain future one; changing forecasts, plants retiring later than expected, and later efforts to secure supply may narrow the gap. Planners and the federal government still take it seriously. The shortfall is a case for new supply, new lines, and faster connections for the power plants waiting in line — not, by itself, a case for adding another giant customer. The pro-build answer is that large customers' money can help finance what the grid already needs.

The companies are making substantial commitments, but their deals differ. Microsoft's long-term agreement and a finalized $1 billion federal loan back the planned 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, not new megawatts. A proposed Louisiana package tied to Meta — seven gas plants, high-voltage lines, storage — is in a pending regulatory filing; the utility that wrote the deal projects roughly $2 billion in customer savings. Planned, operating, proposed: all real, none interchangeable. Some of that equipment may serve the wider grid long after any single contract ends. A power plant is not a private extension cord; it can outlive the deal that built it.

Ratepayers are the households and businesses that pay electricity bills. Running nearly flat-out, a data center can spread the grid's fixed costs — costs that do not rise and fall with every unit of electricity sold — across enormous sales, benefiting others if its payments exceed service costs. Two studies model that. One, funded by Amazon and reviewed by Amazon before publication, examined real Amazon sites across four utilities and projected that payments meet or exceed estimated service costs. Another, funded and reviewed before publication by a data-center developer, modeled a hypothetical 100-megawatt customer under Virginia's new tariff — the binding rate-and-rule sheet approved by the regulator — and found a net benefit to the system in most scenarios: a $37 million base case, ranging from a $166 million benefit to a $92 million cost, depending on assumptions. Neither is an audit. Both are evidence that honest public accounting could vindicate the companies.

Host communities can also gain substantial tax revenue: data centers provide a major part of the property-tax base in Loudoun County, Virginia — funding schools, fire departments, and other services. A community can look at that trade with open eyes and decide it is worth making.

The U.S. government treats advanced AI as an economic and national-security competition. If China or another rival builds power and computing faster, delay may carry a real cost — a serious strategic possibility, though nobody has shown China will spend without limit. Some builders make a larger wager: enough computing power could produce systems far beyond human ability and compress decades of medicine, science, and invention into a few years. That belief is speculative and not presently testable. For those who accept that wager, delay can look like postponing future cures and discoveries — and building fast, at almost any cost, like a moral duty.

Despite headlines predicting mass cancellations, there is no evidence of a nationwide investment collapse: development and planned spending remain enormous even as individual projects slip. Virginia and one big Ohio utility have adopted rules meant to keep the costs of very large electricity customers — large loads — from shifting to others. In June 2026, federal regulators ordered the six regional grid operators they oversee to justify or reform large-customer connections, toward faster studies, clearer cost reporting, and agreements that keep stranded costs from shifting. Stranded costs are costs left without the expected payer and at risk of being charged to others. PJM proposes cost-assignment rules for very large loads. Across governments and parties, the direction is not "stop." It is "make them pay more of their own cost, and keep building."

Held together, it is a serious case.

All of that can be true, and the payer question still remains.

The shortage points two ways. PJM and the federal government see a reason to build faster. PJM's independent market monitor — the watchdog that checks the market's math — calls for temporary connection limits: do not add giant new customers until matching supply exists. The planning shortfall is real. It does not choose the policy by itself.

The announced nuclear deals do not yet cover the demand. One independent review's rough scale check found that if every announced nuclear project succeeded, the output would cover less than one-fifth of projected data-center electricity use in 2035. "Companies are buying power" is true. "The power problem is handled" is not yet supported.

Experts dispute data centers' share of the capacity-market cost increase. The market monitor estimates that their demand raised the region's capacity-market revenues — the payments suppliers receive for committing power to be available — by $9.3 billion in the 2025/26 auction year, and by $23.1 billion cumulatively across three auction years ($9.3, then $7.3, then $6.5 billion). PJM disputes how much of that increase the monitor attributes to data-center demand, also citing plant retirements, slow generator connections, and rule changes; another analysis puts about half the increase on demand growth generally. The fight is over the share, not the mechanism: very large new demand meeting tight supply raises the price of keeping enough power available. Any estimate should show its assumptions. It is not a verdict.

A community benefit is not community permission. Loudoun County's data centers pay a major share of its property taxes — and its elected chair has quoted residents saying they would rather pay higher taxes than accept more data centers. The benefit is real. So is the opposition. The tax money lands in the host county while grid costs can land on households across the utility's territory or the whole region. Those publics overlap but differ.

Even a promise of enormous future benefit does not make households the default payer.

If the public may be asked to pay, it needs three rungs of one ladder: a rule, a receipt, and a say.

A promise can be changed. A binding rule can be enforced.

The rule starts by sorting costs honestly. Costs traceable to one customer — a campus substation, a dedicated line — follow that customer. Genuinely shared grid costs follow a published rule based on who benefits and who caused the need; that ordinary allocation is how utility regulation has always worked, not a subsidy. Market effects — prices rising for everyone when supply tightens — cannot be invoiced to one project afterward. Policy can address them up front through modeled charges based on projected effects, requirements to bring matching supply, or curtailment terms, which provide for reducing use under specified conditions. A public share means something narrower: a deliberate contribution beyond ordinary allocation that leaders ask other ratepayers or taxpayers to make. That choice belongs to the third rung.

The rule should cover any very large new customer — factory, crypto mine, or data center — based on what it does to the grid, not its industry. Projects sharing a site should count together, so a giant load cannot evade the rule by splitting into pieces.

Virginia adopted a substantial part of this rule in November 2025, in Dominion's rate case — the formal proceeding in which the regulator reviews the utility's rates and rules. Customers with at least 25 megawatts of demand, running at least three-quarters of the time, enter their own rate class — a group subject to the same rate rules — in January 2027: GS-5. New customers generally sign fourteen-year agreements. Minimum bills are based on most of their signed-up demand — 85 percent for distribution, 85 percent for transmission, 60 percent for generation, subject to specified exceptions — whether used or not. In plain language: these customers must pay for most of the capacity they reserve even when their actual use falls short, and collateral — money or a guarantee securing payment — backs the obligation. Distribution means local delivery, transmission means high-voltage delivery, and generation means power supply.

What remains open: Dominion has filed alternatives for dividing transmission costs as these customers multiply — that proceeding is pending — and generation-cost allocation comes later. The state's watchdog had pointed toward this fix — a separate class, changed allocations. This is rung one: an adopted, enforceable order, not yet evidence of its outcome.

Ohio moved earlier on stranded-cost protection, in one big utility's territory: a tiered minimum-payment formula rising toward 85 percent of reserved capacity for the largest data centers, with protections for early departure or nonpayment. The regulator's stated reason was to keep other customers from carrying the cost of equipment built for data centers. In Utah, a large-load contract needs state commission approval, granted only on express findings that the customer bears its incremental costs — the extra costs created by serving it — and existing ratepayers bear none of them.

A long-term power contract and an approved, effective tariff are real obligations. A public promise with no enforcement is neither; an independent assessment called the seven-company pledge politically significant but vague and legally toothless. The dividing line is simple:

Can someone hold the company to it if the company changes its mind?

Minimum bills and collateral protect against a customer that shrinks, defaults, or leaves; separate rate classes keep identified costs in the right column. None shows a household what growth added, what companies paid, or what remained. Even the states with the best rules have no receipt. Nor can a rule erase a tighter market. Pennsylvania fought for and won a region-wide cap on capacity prices, yet the market monitor estimates that data-center demand still added about $13.8 billion to capacity-market costs across the two capped auctions, region-wide — the same $7.3 and $6.5 billion auction years already counted inside the $23.1 billion total above. A rule assigns what it can but cannot promise a flat bill or show its work. That is why the second rung matters.

The receipt has two ledgers, not one total.

The audited ledger compares the same assets, years, and repayment schedule. It shows the booked costs — the costs recorded in the utility's accounts — assigned to large loads; the money actually collected from them for those costs; and the residual, meaning any balance left to each other group of customers. That residual can be zero or negative — negative meaning the large customers paid more than their assigned costs, which would be their best evidence. Audited does not mean beyond dispute: the booked costs and collections are checkable, while the published rule deciding which shared costs belong to which group remains open to argument.

The modeled ledger estimates what system costs and market prices would have been without the new customer demand, showing assumptions, range, and effects on each customer group. Because that world did not happen, its numbers can be updated and argued with but not audited like an invoice. They are never netted against real payments — never used to offset those payments and produce one net figure. This is why Virginia's watchdog and the region's market monitor can both be right: one tested how actual rates allocate actual costs; the other estimates a market that never happened. Separate ledgers let the public check what is checkable and argue about what is arguable.

This accounting arrives twice. Before the decision comes the price tag: a forecast of costs and payments, clearly labeled as a forecast, built on the same assets, years, and payment schedules the later audited ledger will use — alongside the separate modeled estimate of bill impact. After service begins comes the receipt: the forecast is replaced with audited booked costs and actual collections, on a schedule, and the model is rerun. Regulators should true up only the amounts the tariff allows to be adjusted later — and explain model error rather than billing it to anyone.

Does this exist? In pieces. The Amazon study and developer-funded Virginia model are forecasts of service costs measured against payments — useful evidence, not completed examples of either ledger: they provide neither after-the-fact booked costs and collections nor the separate market-effects estimate. Federal regulators seek public reporting of which grid upgrades serve large loads and who pays; PJM's backstop proposal would assign some new costs to large loads. But no source found in this research provides the finished product: recurring, regulator-required, after-the-fact public accounting that compares the same assets, years, and payment schedules and shows booked large-load costs, money collected, and each customer class's residual, in words a household can read. The pieces exist in rate cases, cost studies, and bill line items. The hard part is consistent causation, asset-life, and time-period boundaries — then publication. The state consumer advocate — the office whose job is to argue for household customers, where one exists — must be able to test it, including confidential parts: where inputs must stay confidential, the advocate publishes how it checked them, what it concluded, and why each redaction stands.

Imperfect transparency is better than total secrecy when it comes to figuring out who is really paying for massive power grid upgrades

Two decisions sit inside "a say." The host community needs an informed, reasoned process on land use. And any deliberate public share needs recorded authorization by a body answerable to the people who will pay.

Start with the first, because "why did nobody ask us?" has a real answer: there is no single approval. Different bodies decide different pieces. In Virginia and the PJM region:

Some local zoning allows data centers by right, meaning staff can approve a project under existing zoning without any new vote by the elected board. Other states arrange the pieces differently, but nowhere in this research does one body weigh land, grid, household bill, and national strategy together. The grid crosses borders; no town can be asked to settle a multistate grid plan by itself. That division is defensible. It also creates a democratic problem. Each body can truthfully say that the largest moral choice belongs to someone else. The public can speak at hearings, elect officials, and file comments. But there is no single moment when the whole bargain is placed before it.

Gallup found that 71% of Americans oppose construction of AI data centers in their local area. That is a national measure of siting opinion, not cost opinion; it is not proof that any particular community withheld consent, and it is not opposition to data centers everywhere. The powerlessness is structural: fragmented authority, and by-right approvals that never reach an elected vote.

For siting, a real say need not mean a veto. It means being told in advance, being shown the evidence, having someone whose job is to challenge it, and receiving a recorded decision with reasons — from the body that holds the power.

The reform is sequencing, not a referendum: the vote comes after the numbers. For a project needing an elected land-use vote, the enforceable payment rule must already be in place, and the state consumer advocate must test the published price-tag forecast and the separate modeled bill-impact estimate before the board votes. Early, conditional land review is fine; a detailed grid study may require a developed proposal. The board votes only on what it controls, numbers in hand. Virginia has taken a partial step — a 2026 law requiring utilities to share limited grid information for certain very large projects needing discretionary local approval.

Where zoning is by right, the elected checkpoint disappears; a sequencing clearance would fill the gap. A statute would require a designated state regulator, before the utility's final commitment to serve the project, to confirm: the site is legally eligible; the grid study is complete; the tariff or service contract is enforceable; and the public price-tag forecast and the separate modeled bill-impact estimate have been tested by the consumer advocate. The regulator's role would be ministerial — confirm the conditions are met, and return incomplete applications to be fixed. It would not reopen vested land rights — land-use rights the project has already secured under law — and could not direct PJM or the federal regulator. If lawmakers want the power to reject an otherwise compliant project, they should call that a veto and give it a standard, a deadline, and an appeal. These tools exist: Utah requires commission approval of large-load contracts, with findings that costs will not shift to existing customers. No jurisdiction found in this research combines all four elements in one clearance.

The third rung matters most when leaders want the public to share the cost. National leaders may decide building faster serves the whole country — because of economic and national-security competition, or the largest hopes attached to AI — and ask the public to help pay. A country may make that choice openly: how much, at most, for how long, with an end date; what public benefit it buys; who carries the loss if the project fails. The approving body must answer to the payers. A county board cannot volunteer households across a state or region. A claimed national benefit does not by itself justify charging one state's households — national funding through a law is the cleanest match, and a narrower payer group needs an open finding of benefit by a body answerable to it.

A tight market raising prices for everyone is not, by itself, a subsidy. Subsidy by stealth begins when officials leave an assignable, customer-specific cost — or a deliberately chosen public share — sitting inside general rates without a visible decision.

The standard is not "no public investment." It is "no public subsidy by stealth."

None of this promises a flat bill, and none of it turns a model into an invoice. It makes the uncertainty visible instead of leaving it buried.

There is an ending the current system does not require: before officials commit the public, the household sees the price tag and the bargain in plain view. This is what the project is expected to add. This is what the company is bound to pay. This is the public share being requested. This is what leaders say the country will gain. Is that trade worth making?

The answer may be yes. It may be no. The point is that the answer should not be hidden inside the bill after the decision is already over.

The rule assigns the bill. The receipt shows the result. The say covers anything the public is deliberately asked to carry.

What this is: An explainer — a structural argument about how data-center grid costs get decided and checked, built on regulator orders, statutes, and market-monitor filings read in the original, as of July 11, 2026.

Confidence: Medium-high. The rules and orders cited were read directly. The cost estimates are contested models, and the essay labels them that way. Its "nothing like this exists" claims are limited to what this research searched, and corrections are invited.

What would change our mind: A jurisdiction shown to already require the recurring, after-the-fact public accounting this essay says is missing — or Virginia's new framework producing filings that let outsiders match booked costs, collections, and what remained for each customer group.

What would prove this essay wrong

Sources

Virginia: SCC Final Order, Case PUR-2025-00058 (Nov. 25, 2025) · SCC data-center fact sheet · JLARC Report 598, "Data Centers in Virginia" (Dec. 2024) · Va. Code § 15.2-2209.5

PJM and its market monitor: PJM 2027/28 auction results · Market monitor, Analysis of the 2027/2028 Base Residual Auction, Part A (Jan. 5, 2026) · Monitor's FERC comments on the capped auctions · PJM reliability-backstop proposal (June 30, 2026)

Other states: AEP Ohio data-center tariff · Utah Code § 54-26-302

Pledges and deals: White House Ratepayer Protection Pledge (Mar. 2026) · DOE loan for the Pennsylvania nuclear restart · Talen–Amazon agreement (SEC filing) · Entergy–Meta announcement

Studies and assessments: E3, "Tailored for Scale" (Dec. 2025, Amazon-funded) · E3, "Beyond the Headlines" (June 2026, developer-funded) · Carnegie Endowment, "Beyond the Hype" (June 2026) · SemiAnalysis on the cancellation narrative

Polling, local, federal: Gallup: Americans oppose AI data centers in their area · Loudoun County data-center FAQ · FERC's June 2026 large-load action · DOE, "Powering America's AI Future"

Signal & Noise is written under the pen name Synthia Cipher. AI tools draft and critique; the human author owns the editorial judgment, final wording, and published claims. If something here is wrong, the fault is the author's, not the algorithm's.