The Latent State

The power bill: Who pays for the AI buildout?

Wed Nov 04 2026

#AI#data-centres#electricity#utilities#infrastructure#regulation#finance

In February 2026, AEP Ohio published the result of a useful experiment. Data-centre developers had asked it for roughly 30,000 megawatts of electricity capacity. Then the utility required them to pay for engineering studies, sign binding agreements and provide financial security. Of that initial pool, 5,642 megawatts reached the contract stage. AEP Ohio, February 13, 2026

The difference was larger than the utility's entire historical peak demand.

Those withdrawn or deferred requests were not necessarily dishonest. A developer can investigate several sites, wait for a customer, change the size of a campus or decide that a connection will take too long. But a utility has to decide when a request becomes sufficiently real to justify ordering equipment and building lines. The consequences of making that decision too early can remain on customers' bills long after the developer has moved on.

That is one way the AI investment boom reaches people who have never bought an AI subscription. Another arrives through wholesale power markets. A third sits inside the rules that divide the cost of a transmission line among customers.

A data centre can pay every bill it receives and still make everyone else's electricity more expensive. Whether it does depends on what gets built, how quickly supply responds, which costs regulators assign to the customer, and who remains responsible when the original plans change.

The documents that answer those questions are rarely the ones announcing the investment. They are service agreements, capacity-auction reports, transmission plans and orders with names such as Rider T1.

Table of Contents

A request becomes a construction programme

AEP Ohio's February update separates three stages that are frequently collapsed into a single headline about electricity demand.

Stage in the initial queue Requested or committed capacity What the developer had done
Expressions of interest before the new tariff About 30,000 MW Asked for service
Projects proceeding to paid engineering studies 13,022.7 MW Requested and paid for a study
Projects signing under the new tariff 5,642 MW Accepted binding financial and legal commitments

The utility separately reported 12,219 MW of contracts signed before the tariff, bringing its total contracted data-centre pipeline to 17,861 MW, scheduled progressively through 2035. Its historical peak across all customers had been approximately 8,000–10,500 MW. These are contracted future projects, not operating consumption. AEP Ohio's filing summary

Calculated from the rounded starting figure, about 19% of the initial 30 GW pool reached the new-contract stage. That is a measure of this particular screening process, not a national cancellation rate. The old contracts belong to a separate group; adding them back to the initial queue would muddle the comparison.

flowchart TD
    accTitle: AEP Ohio's initial data-centre queue under its new tariff
    accDescr: About 30 gigawatts of initial requests narrowed to 13.023 gigawatts of paid studies and 5.642 gigawatts of binding new contracts. Previously signed contracts are shown separately.
    Requests["Initial requests: about 30 GW"] --> Studies["Paid engineering studies: 13.023 GW"]
    Studies --> New["Binding new contracts: 5.642 GW"]
    Old["Previously signed contracts: 12.219 GW"] --> Total["Total contracted pipeline: 17.861 GW"]
    New --> Total
    Total --> Timing["Scheduled progressively through 2035"]

Source: AEP Ohio, February 2026. The arrows show stages and aggregation, not electricity already being delivered.

The tariff changed the price of reserving a place in the queue. Under AEP's published process, a developer must control its proposed site and supply a location, expected demand and expansion schedule. Study fees range from $10,000 to $100,000. Its pre-energisation agreement requires reimbursement of 100% of buildout costs if the customer cancels or delays the project by more than twelve months before the target connection date. AEP Ohio's connection process

Ohio's consumer advocate summarises the approved tariff's continuing protection this way: new large data centres must pay for at least 85% of contracted capacity, for up to twelve years, even when they use less. This is a minimum capacity-payment obligation; it should not be described as buying and wasting 85% of the forecast kilowatt-hours. Ohio Consumers' Counsel, case 24-0508-EL-ATA

There was a commercial disagreement over those terms. The regulator's July 2025 announcement records competing settlements: one backed by Microsoft, Amazon, Google and the Data Center Coalition, among others; another backed by AEP Ohio, commission staff, the Consumers' Counsel and other parties. The commission adopted the latter. PUCO's July 9 announcement

The important result is visible without guessing anyone's motives. Once the request carried a substantial obligation, the planning number became much smaller. A contract helped establish what the developer actually intended to buy.

The scale is real. The forecasts are conditional

The underlying growth is substantial. Berkeley Lab's 2024 US Data Center Energy Usage Report estimated that American data centres consumed 176 terawatt-hours in 2023, or 4.4% of US electricity. Its 2028 scenarios ranged from 325 to 580 TWh, equivalent to 6.7–12% of projected national consumption. Berkeley Lab's report summary

The International Energy Agency's updated 2026 outlook puts global data-centre consumption at 485 TWh in 2025, rising to 950 TWh in 2030 in its central projection. It explicitly says bottlenecks make the more aggressive near-term growth scenarios less likely. IEA, Key Questions on Energy and AI

Measure Historical estimate Future estimate What the comparison covers
US data-centre electricity, Berkeley Lab 176 TWh in 2023 325–580 TWh in 2028 A range of equipment and operating scenarios
US share of electricity, Berkeley Lab 4.4% in 2023 6.7–12% in 2028 Depends on both data centres and the rest of the economy
Global data-centre electricity, IEA's 2026 outlook 485 TWh in 2025 950 TWh in 2030 Central projection, not a count of signed power contracts

These estimates cover data centres generally, including conventional cloud services. They cannot all be attributed to generative AI. They also cover different geographies and years, so the global and American figures should not be used as successive points in one growth series.

There is a second distinction worth making before reading any announcement: power and energy are different quantities. A megawatt measures the rate of electricity use. A megawatt-hour measures how much electricity is used over time. A gigawatt is 1,000 megawatts; a terawatt-hour is one million megawatt-hours.

For illustration, a facility drawing 1 GW at an average of 90% of that level throughout a non-leap year would consume:

1,000 MW × 8,760 hours × 90% = 7,884,000 MWh, or 7.884 TWh.

That is arithmetic under an explicit utilisation assumption. An announced 1 GW campus may instead open in phases, operate below its ultimate size or reserve more connection capacity than it initially needs. Installed computing capacity, contracted grid capacity and actual meter readings can therefore tell three different stories.

Utilities must plan before all the meters are running. That makes forecasts unavoidable. It also makes the rules for accepting a forecast unusually consequential.

Three costs can arrive on the same bill

Electricity bills collect several different services. The US Energy Information Administration identifies generation, transmission, distribution, financing and operating costs among the factors affecting prices. It also points to fuel, weather and regulation. A bill increase is not, by itself, evidence of a data-centre subsidy. EIA's explanation of electricity prices

For this investigation, three mechanisms matter:

Cost What is being purchased How a new data centre can affect others
Energy Electricity produced and delivered over time Additional demand can require more expensive generation and raise market prices
Capacity A commitment that resources will be available when needed A higher demand forecast can raise the price of securing future availability
Network and other fixed infrastructure Transmission lines, substations and, where applicable, utility-owned generation Construction costs can be assigned to the customer, spread among classes, or left behind after a departure

A payment that addresses one row does not automatically address the other two. Funding a dedicated substation does not undo a regional capacity-price increase. Signing a contract for renewable electricity does not decide who pays for an unrelated transmission upgrade. A long service agreement helps with cancellation risk but cannot make fuel prices stand still.

flowchart TD
    accTitle: How data-centre demand can reach other customers' bills
    accDescr: Actual demand can affect energy markets, forecast demand can affect capacity procurement, and construction can enter regulated cost recovery. These are separate mechanisms, with different protections.
    Project["Data-centre project"] --> Actual["Actual electricity demand"]
    Project --> Forecast["Demand included in forecasts"]
    Project --> Build["New plants, lines and substations"]
    Actual --> Energy["Energy-market prices"]
    Forecast --> Capacity["Capacity procurement and prices"]
    Build --> Allocation["Regulatory cost allocation"]
    Energy --> Bills["Retail bills, through tariffs and supply contracts"]
    Capacity --> Bills
    Allocation --> Direct["Charges assigned to large customer"]
    Allocation --> Shared["Costs allocated among customer classes"]
    Shared --> Bills

The diagram shows possible routes. It does not imply that every project raises prices or that every jurisdiction uses a capacity market.

The most misleading version of the debate asks only whether a technology company pays for the electricity it uses. The more useful examination follows each cost from the decision that creates it to the account that ultimately collects it.

PJM: the price of keeping capacity available

PJM coordinates the grid across parts or all of thirteen states and the District of Columbia. Its capacity auctions buy commitments for future delivery years, which run from June through May. The quoted unit is dollars per megawatt-day of accredited capacity. It is not the price of a megawatt-hour of electricity.

The recent auction sequence shows how sharply the cost of availability changed:

Delivery year Main PJM clearing price, $/MW-day Qualification
2024/25 28.92 Historical comparison
2025/26 269.92 BGE and Dominion zones cleared higher: 466.35 and 444.26
2026/27 329.17 Cleared at the approved price cap
2027/28 333.44 Cleared at the approved price cap
2028/29 325.00 Cleared at the approved price cap

Sources: PJM's historical auction presentation, 2026/27 results, 2027/28 results, 2028/29 results.

The main clearing price for the current 2026/27 delivery year is about 11.4 times the 2024/25 price. A household's whole bill does not rise by that multiple: capacity is one component, and retail contracts determine how and when wholesale changes reach customers.

The July 2026 auction for 2028/29 procured 138,318 MW of accredited capacity and demand response. Including separately supplied resources, PJM reported a 6,831 MW shortfall against its reliability requirement. The cleared auction quantity multiplied by its price represented $16.4 billion. PJM cautioned that this is not the total cost paid by customers, because self-supply and bilateral contracts change exposure to the clearing price. PJM's July 14, 2026 results

A shortfall against a planning standard is also different from a prediction that power will necessarily fail. It means the auction did not procure the target level of resources, including the required safety margin.

How much does the market monitor attribute to data centres?

Monitoring Analytics, PJM's independent market monitor, reran auction calculations with data-centre demand removed. Its October 2025 analysis attributed $7.27 billion of additional 2026/27 capacity-market revenue to the inclusion of existing and forecast data-centre load. Its March 2026 annual report put the corresponding 2027/28 figure at $6.50 billion. 2026/27 sensitivity analysis, pp. 1–3, 2025 State of the Market, Volume 1

By August 31, 2026, the monitor's estimate across four auctions, covering 2025/26 through 2028/29, reached $29,368,774,465. Monitoring Analytics, Powering Reliability Through Market Design, p. 2

That is a counterfactual estimate. It compares the auctions with the monitor's recalculations; it is not an invoice item headed “AI”, an estimate of unpaid company bills, or a sum already collected from households. Some of the delivery periods are still in the future. Removing existing data centres is also a broader experiment than cancelling only speculative new projects.

The mechanism is nevertheless clear. Additional demand can change the price paid for capacity across a market. A new customer pays its own capacity charges, but that does not reimburse existing customers for the higher price they may also face.

If a scarce service becomes more expensive after a large buyer arrives, the other buyers can lose even when the newcomer pays in full.

What had already changed by September 2026

Future auctions are only part of the record. The monitor's October 28, 2026 presentation compares wholesale costs during the first nine months of 2025 and 2026:

PJM wholesale component Jan–Sep 2025, $/MWh Jan–Sep 2026, $/MWh Change
Energy 48.47 70.89 +46.3%
Capacity 10.72 19.42 +81.1%
Transmission 18.32 19.16 +4.6%
Total, including other components 79.19 111.61 +40.9%

Source: Monitoring Analytics, October 28 presentation, slides 8–12. Components shown do not include every item in the total.

The same presentation estimates that data-centre load's effect on capacity alone added $9.95/MWh to the 2026 wholesale total: $111.61 rather than $101.66/MWh. That is 9.8% above its counterfactual, not a finding that data centres caused the whole 40.9% year-on-year increase. It excludes their effects on energy and transmission.

For scale, $9.95/MWh equals 0.995 cents/kWh. Applied mechanically to 1,000 kWh, it is $9.95. This is a unit conversion, not a prediction of a particular household's monthly surcharge. Retail exposure, customer allocation, hedging and timing still have to be established.

These distinctions protect the argument from an easy exaggeration. Electricity had become more expensive for several reasons. The monitor identifies a particular contribution under a stated method. It does not assign every change in gas prices, weather, generator availability or market rules to AI.

The transmission line that becomes everybody's expense

A line built to connect one customer presents a different question from a market-wide price increase. Here, the issue is who is charged for a particular investment.

In a September 2025 policy brief, the Union of Concerned Scientists, an advocacy organisation, examined utility transmission plans in seven PJM states. It identified 130 projects approved in 2024, with listed costs totalling $4,356.3 million, associated with connecting data centres. Its table gives this breakdown:

State Projects Listed cost, $ million
Illinois 3 239.0
Maryland 4 107.5
New Jersey 8 14.5
Ohio 37 1,299.7
Pennsylvania 16 491.8
Virginia 60 1,988.0
West Virginia 2 215.8
Total 130 4,356.3

UCS says six projects were directly customer-funded, while costs for the remaining 124 entered the shared pool. This is UCS's analysis of planning records, not a commission finding that households had already paid $4.36 billion. These are project costs, recovered over time; they must not be added to annual capacity charges as though both measured the same thing. UCS, Connection Costs, Table 1 and accompanying analysis

The jurisdictional structure helps explain why the money is difficult to follow. FERC oversees interstate transmission rates and terms. State commissions decide how regulated utilities allocate their assigned transmission costs among retail classes. FERC's public-participation office describes the sequence explicitly: transmission costs reach utilities, utilities bring them into state proceedings, and state commissions determine their treatment in retail bills. FERC's October 2026 explanation

A project can therefore have a recognisable customer at the planning stage and become one entry in a much larger transmission revenue requirement at the billing stage. To establish who subsidises whom, the investigation has to follow the project through both stages, including any customer contributions and revenue credits.

Shared funding is sometimes justified. A larger line might improve reliability for many users, replace equipment that already needed replacement, or serve future customers. Assigning the whole cost to the first newcomer could overcharge it.

But that justification requires evidence. What would have been built without the customer? What extra capacity was added for it? Who benefits from that addition? How much has the customer already contributed? A utility's description of a project as improving reliability cannot settle all four questions.

Virginia: paying an allocated share can still leave others worse off

Virginia's Joint Legislative Audit and Review Commission, or JLARC, made two findings in its December 2024 investigation that are often separated in public discussion.

First, its commissioned cost-of-service study found that data centres were paying their allocated costs under the existing rate structures. Second, their growth was likely to increase system costs and other customers' bills. It estimated that a typical Dominion residential customer's generation- and transmission-related costs could rise by $14–$37 a month by 2040, in constant dollars. This was a long-term scenario estimate, not an approved surcharge or a measurement of today's bill. JLARC, Data Centers in Virginia

Both findings can be true. A customer can pay the share assigned by a tariff while its arrival changes the cost of supplying everyone. “Pays its allocated share” describes an accounting rule. “Leaves existing customers no worse off” requires a comparison with the system that would exist without the new demand.

Virginia subsequently changed the rules. The State Corporation Commission's February 2026 fact sheet describes a separate GS-5 class and several safeguards:

Dominion protection Scope stated by the SCC
Separate large-load rate class Includes new and existing customers meeting the thresholds
Minimum service obligation At least 14 years for new large-load customers contracting on or after January 1, 2027
Minimum transmission and distribution charges At least 85%, regardless of usage; customers beginning service before January 1, 2016 are exempt
Credit support Customers without sufficient credit may have to guarantee up to 60% of minimum contract charges; applies to new contracts from January 2027

Source: SCC fact sheet, case PUR-2025-00058.

Those dates matter. A requirement adopted in 2025 does not necessarily apply to every contract signed before 2027. Nor does an 85% transmission-and-distribution floor mean an 85% floor on every cost in an electricity bill.

Ninety-four cents, and the allocation rule behind it

In its 2026 Rider T1 proceeding, the SCC approved changes that brought the allocation of transmission costs closer to forecast growth in large-load customers and incorporated their minimum transmission charges. Its September explanation says the revised method reduced a projected increase for a 1,000-kWh residential customer from $2.90 to $0.94 a month.

That is a 67.5% reduction in that projected transmission-rider increase, not a 67.5% cut in the whole electricity bill. The commission also required development of a policy for assigning certain direct-connection costs to the large customers that necessitate them. Detailed implementation went into a supplemental proceeding. SCC, cases PUR-2026-00056 and PUR-2026-00131

The change illustrates a practical problem with using historical demand to divide costs during a surge. If the allocation calculation looks backwards while the construction programme looks forwards, established customers can carry a larger share of expenditure associated with newcomers. Changing the calculation can move money without changing a single transmission tower.

Regulatory detail is doing the work here. The protection is the formula, its effective date and the costs it covers.

Louisiana: the guarantee has to reach the right creditor

In Louisiana, the question extends beyond a dedicated line. Entergy is building generation for a very large customer.

On August 20, 2025, Entergy announced regulatory approval for three combined-cycle gas facilities associated with Meta's Richland Parish development, plus transmission investments. Two plants were expected in late 2028 and the third by the end of 2029. Entergy said Meta would pay its share and other customers would be protected. That is the utility's stated position. Entergy's approval announcement

Two months later, Meta announced that funds managed by Blue Owl Capital would own 80% of a joint venture developing and owning the Hyperion campus; Meta would retain 20%. The disclosed development cost was approximately $27 billion. Meta would lease the completed facilities on an initial four-year term, with extension options. It also disclosed a capped residual-value guarantee covering the first sixteen years of operations under specified conditions. Meta's October 21 announcement, updated November 14, 2025

A four-year campus lease is not automatically a four-year electricity contract. A residual-value guarantee supporting campus financing is not automatically a guarantee of every obligation to the utility. The beneficiary, limits and enforcement terms determine what each promise protects.

That distinction became a documented dispute. On January 14, 2026, the Alliance for Affordable Energy and UCS filed a motion in Louisiana docket U-37425, through Earthjustice. They asked the commission to investigate whether the ownership and financing arrangements weakened the approved electricity protections. They questioned the position of the contracting entity, Laidley LLC, the parent guarantee and the priority of utility claims relative to investor payments. The motion described an original electricity-service term of fifteen years. The filed motion and supporting material

These are the intervenors' allegations and questions. The filing does not establish that Meta escaped its obligations, that a guarantee is unenforceable, or that regulators were misled. This article has not established a final ruling resolving those claims.

flowchart TD
    accTitle: Distinct contracts around Louisiana's Hyperion project
    accDescr: Meta disclosed an 80 percent Blue Owl funds and 20 percent Meta campus joint venture, a lease and a capped residual value guarantee. Intervenors separately questioned protections under Laidley's utility service agreement. Dotted links identify disputed questions, not established failures.
    Owl["Blue Owl-managed funds"] -->|"80 percent interest"| Campus["Campus joint venture"]
    Meta["Meta"] -->|"20 percent interest"| Campus
    Meta -->|"Lease and capped residual-value guarantee"| Campus
    Laidley["Laidley LLC"] -->|"Electricity service agreement"| Utility["Entergy Louisiana"]
    Utility -->|"Generation and transmission construction"| Assets["Power infrastructure"]
    Motion["January 2026 intervenor motion"] -.->|"Questions guarantee and claim priority"| Laidley
    Motion -.->|"Requests investigation"| Regulator["Louisiana Public Service Commission"]

The diagram separates the disclosed campus arrangements from the utility contract under dispute. It does not establish a default or reproduce the full corporate structure.

The wider Louisiana pipeline has continued growing. Entergy's second-quarter 2026 SEC filing records a March application for seven additional combined-cycle resources, totalling 5,278 MW, at an estimated $12.9 billion, alongside battery projects. The filing connects the application with service to another Meta subsidiary, Evest LLC. It records an application, not a completed fleet or unconditional final approval. Entergy's Form 10-Q

The underlying test is straightforward even when the structure is complicated: which solvent entity must pay the utility if the computing business changes its mind?

The existence of a wealthy parent does not answer that question. Neither does an investor guarantee unless the relevant utility obligation is actually covered. The enforceable chain of responsibility has to survive a sale, lease termination, restructuring and, potentially, bankruptcy.

Why a utility can welcome a risky forecast

Utilities and their customers do not necessarily face the same incentives.

Under cost-of-service regulation, an approved investment can enter the rate base: the asset value on which a utility is allowed an opportunity to earn a return. Bills also recover depreciation and operating costs. That arrangement finances infrastructure that private firms might otherwise be unwilling to build. It can also make a larger approved construction programme attractive to shareholders.

The October 2026 revision of a working paper by Robyn Meeks, Jacquelyn Pless, Zhiyuan Qi and Zhenxuan Wang, published through MIT's Center for Energy and Environmental Policy Research, finds that data-centre entry increased infrastructure investment and authorised rate bases. Its retail-price effects were larger among investor-owned utilities and smaller where utilisation improved. Who Pays for Growth?

This does not mean a utility earns a guaranteed profit on everything it proposes. Regulators can reject projects, reduce allowed returns or disallow costs they find imprudent. Shareholders and lenders can bear losses. The point is that the investment decision and the allocation decision must be examined together.

Consider a deliberately simplified example:

Illustrative project assumption Amount
New infrastructure $1 billion
Assumed recovery period 30 years
Straight-line depreciation, before financing and operating costs $33.3 million a year
Customer's initial contractual term 15 years
Undepreciated value at year 15 under those assumptions $500 million

No particular project's exposure is estimated here. Actual depreciation, contributions, contract payments, salvage value, replacement customers and regulatory treatment all change the result. But the example shows why a fifteen-year commitment and a thirty-year asset require a plan for the remaining years.

A minimum bill protects only the obligations it actually covers. An exit fee protects only to the extent that it captures the remaining exposure and can be collected. Credit support protects only up to its amount, for its term, from the entity providing it.

If replacement demand is the answer, it needs its own evidence. A specialised substation beside an abandoned campus may be less useful than a transmission upgrade serving several growing towns. “The assets will still be there” is not the same as “someone else will pay enough to use them.”

The serious case that growth can lower bills

There is a credible economic argument on the other side. An electricity system has substantial fixed costs. A new customer using existing capacity can spread those costs across more sales. If its payments exceed the additional costs it creates, existing customers can benefit.

Suppose a system needs to recover $100 million of annual fixed costs from one billion kWh of sales. That component costs 10 cents/kWh. Add 250 million kWh of sales without adding fixed costs, and the average falls to 8 cents. If serving the same customer instead raises fixed costs to $140 million, the average becomes 11.2 cents.

These are illustrative figures, excluding energy costs and differences between customer classes. They show why the sign of the effect cannot be settled by counting megawatts alone.

The empirical literature also deserves more care than a single convenient headline:

Research What it examines Reported result and limit
Watten, Bistline and Blanford, June 2026 US retail rates, 2015–24, using an instrumental-variables design Finds modest rate reductions; preferred residential estimate implies a 3.5% reduction for a doubling of data-centre capacity. Warns future supply constraints could reverse the effect. Paper
Meeks, Pless, Qi and Wang, revised October 2026 Data-centre entry, 2010–24, linked to utilities and regulatory outcomes Finds average retail rates up 2.7%, residential rates up 2.1%, and larger effects at investor-owned utilities. Effects depend on ownership and utilisation. Paper
Kay, Reaser and Taylor, revised October 2026 An hourly dispatch model of 23 trading US electricity regions Estimates wholesale prices 6.9% higher nationally and 15% higher in northern Virginia by 2025; a model-based wholesale result, not an observed household surcharge. Dallas Fed working paper
Elhai and Zhou, October 31, 2026 version Dispatch and data-centre location choices Models price and emissions effects that depend strongly on location; finds developers' preference for flatter supply curves has limited price increases. Current working paper

These papers do not provide interchangeable estimates. They use different samples, identification assumptions, geographic boundaries and price measures. Their disagreement cannot be resolved by averaging the percentages.

The latest Elhai–Zhou version is particularly relevant to claims circulating from earlier drafts. It puts the modelled 2025 wholesale-price effect at about $1.60/MWh nationally. Its appendix finds that standard retail-price approaches produce fragile results; it explicitly concludes that neither of the two reduced-form designs it examines is credible. It also distinguishes generation costs from retail infrastructure allocation. An earlier headline about falling retail prices should not be presented as the conclusion of this revised paper. October 31 paper, introduction and Appendix A

All four are research estimates or working papers, not adjudications of a specific utility's rates. They establish useful questions for a local investigation: Is there spare capacity? Does the customer improve utilisation? What new investment is required? Does the rate structure actually distribute the benefit to existing households?

For a proposed campus, the relevant comparison is the utility's forecast with and without that project, under the contract being offered. A national historical average cannot substitute for it.

A private power deal still has to fit the grid

Buying power directly can help finance supply. It does not, by itself, make the customer's effect disappear.

An agreement that supports a new generator can add electricity. An agreement reserving output from an existing generator can change who has contractual access to electricity already being produced. Those transactions may both be commercially reasonable, but they answer different supply questions.

A second question concerns timing and location. Annual renewable-energy purchases cannot establish, on their own, that a particular facility can run through every constrained hour without additional grid support. The physical network must still carry power, balance the system and provide whatever standby service the customer retains.

FERC's December 2025 co-location order recognised the need for more precise arrangements. It found PJM's existing tariff lacked clarity and consistency and directed development of transmission options for large loads beside generators, including services reflecting limits on their grid withdrawals. FERC's order fact sheet

The fair charge depends on the service actually reserved. A facility that can reliably disconnect when its generator fails should not automatically be billed as though it has unconditional access to the whole grid. A facility retaining substantial backup rights should not be treated as having no dependence on it.

The contract must say what happens on the difficult day: when the plant trips, renewable output falls, fuel supply fails or the wider grid is short of power. That is when an apparently private arrangement can require public infrastructure to perform.

Flexibility can avoid some construction

There is another way to reduce costs: make part of the new demand conditional on available capacity.

A February 2025 Duke University study examined how much large flexible demand existing systems might accommodate. It estimated about 98 GW under a scenario allowing curtailment equivalent to 0.5% of the added load's maximum potential annual energy consumption. That is a modelling result across the systems studied, not a finding that 98 GW can connect anywhere tomorrow. Duke's Rethinking Load Growth, executive summary and methodology

The distinction between an energy curtailment rate and an outage duration matters. In that scenario, reductions were required during 177 hours a year on average, often for only part of the added load. Curtailment events averaged 2.1 hours. A 0.5% energy reduction therefore does not mean a universal forty-four-hour shutdown schedule. Local transmission, reliability and operational constraints still require study.

There are commercial examples. In August 2025, Google announced demand-response agreements with Indiana Michigan Power and TVA, extending its flexibility work to machine-learning workloads. It also acknowledged that flexibility was at an early stage, available only in certain locations and limited by services that require high reliability. Google's announcement

The useful commitment is measurable: how many megawatts can the operator reduce, on how much notice, for how long, and with what penalty for failing to deliver?

Moving a training run to another time or location may avoid part of a peak. Merely promising to be flexible cannot justify treating the same demand as absent from a reliability calculation. Nor should a discounted customer retain the right to demand full service precisely when it agreed to reduce consumption.

The pledge and the enforceable obligation

By March 2026, the issue had reached the White House. Its Ratepayer Protection Pledge called for companies to procure additional supply, fund delivery upgrades and negotiate separate rates payable even when the electricity is not used. The published text describes those rate negotiations as voluntary. The March 4 pledge

Those are relevant commitments. Their practical value depends on their appearance in enforceable utility arrangements, with defined obligations, financial security and consequences for non-performance.

The House-passed text of H.R. 9340, the Ratepayer Protection Act, illustrates another distinction. It proposes a federal standard for recovering incremental upgrade costs from qualifying computing loads of 100 MW or more, including financial assurances. But its implementation provisions direct state authorities and nonregulated utilities to consider the standard and make a determination. That is different from automatically imposing one identical tariff everywhere. This article treats the House text as proposed legislation, not enacted protection. Official House-passed bill text

Regulatory work was still underway as this article was completed. FERC's October 19 explanation of its June show-cause proceedings emphasises that those orders made preliminary determinations, not final new transmission rates. Responses were due in November. FERC's explanation and docket list

PJM also filed an amended reliability-backstop proposal on October 29, 2026. It seeks procurement beginning in March 2027, with commitments of up to fifteen years. PJM says its revised allocation would use large-load growth between the actual 2026 summer peak and the 2028/29 forecast, and it proposes revised collateral and continuing obligations for departing load-serving entities. These were proposed revisions awaiting regulatory treatment, not an auction already completed. PJM's October 29 announcement

The distinction between a large-load customer and its load-serving entity matters here. Assigning a wholesale charge to the organisation serving a data centre is one step. Ensuring that the data centre, rather than that organisation's other retail customers, ultimately bears it requires the retail arrangements to complete the job.

What a protection worth the name must cover

The evidence points to a set of contract questions that can be asked before the first concrete is poured.

First, separate a site enquiry from an investment-grade commitment. Require evidence of site control, paid studies, credible delivery dates and financial support before treating the full requested load as certain. Update the forecast as evidence changes. Ohio shows that this can materially alter the planning number.

Second, identify the incremental assets. Publish the lines, substations and generation investments that would not be needed without the project. Show which costs are directly assigned, which are shared and why. Credit the customer for contributions already paid and credit the wider system for benefits it actually receives.

Third, match the payment obligation to the exposure. Minimum demand charges, ramp schedules and exit payments should cover the costs being incurred, rather than a convenient percentage with no demonstrated relationship to them. A network minimum is not a complete generation guarantee.

Fourth, test the remaining years. If the contract ends before the assets are recovered, identify the remaining balance and credible replacement use. Make the assumptions visible. Do the same for early departure, delayed construction and a campus filled much more slowly than forecast.

Fifth, follow the guarantee through the corporate structure. Identify the legal customer, the guarantor, the cap, expiry, collateral and consequences of a change in ownership. Review the protection after financing changes. A guarantee protecting a landlord or bondholder cannot simply be counted again as protection for electricity customers.

Sixth, recognise wholesale effects separately. Directly funding a connection does not settle the cost of higher capacity or energy prices elsewhere. State what the tariff can prevent, and what remains exposed to regional market conditions.

Seventh, buy verified flexibility when it is cheaper than construction. Give the operator a clear service level and test its performance. A contract for interruptible service should give the grid operator a usable right to interrupt it.

Finally, publish a comparison that households can understand: expected bills with the project, expected bills without it, and the outcome if the project arrives late or leaves early. Include fuel and financing sensitivities. State who absorbs the difference under each scenario.

flowchart TD
    accTitle: Testing who bears a data-centre project's infrastructure risk
    accDescr: A review proceeds from a binding commitment through incremental cost allocation, credit support and departure scenarios. Uncovered amounts remain an exposure requiring an explicit regulatory decision.
    Commitment["Binding load commitment"] --> Costs["Identify incremental assets and market effects"]
    Costs --> Payments["Map costs to charges and contributions"]
    Payments --> Security["Check guarantor, collateral, caps and expiry"]
    Security --> Scenarios["Test delay, lower usage, sale and early departure"]
    Scenarios --> Covered["Amounts recoverable from the customer"]
    Scenarios --> Residual["Remaining exposure"]
    Residual --> Decision["Explicit decision on shareholders, other customers or redesign"]

This is an analytical checklist, not a description of one state's legal requirements.

The customer left on the bill

Data centres can provide tax revenue, construction work, valuable services and investment in electricity supply. JLARC's review recognises those benefits. It also shows why a host locality's economic-development calculation and an electricity customer's cost calculation need not be the same: benefits can be concentrated in one place while utility costs extend across a much larger service territory. JLARC's findings

That makes the terms of the bargain more important, not less. A county receiving revenue cannot, by itself, settle what a household in another county should pay for the connection. A forecast of national technological progress cannot establish the value of a particular power plant. Those decisions require costs, contracts and a comparison with the alternatives.

The records already show that the outcome can change. Ohio made developers commit and obtained a smaller queue. Virginia changed an allocation formula and reduced a projected residential increase. The Louisiana dispute asks whether financial protections survive a change in the structure behind the customer. Each turns on something specific enough to inspect.

In The Capex Confession, the risk sits inside the accounts of the companies building AI. Here it can cross into a regulated system whose customers need electricity regardless of whether the computing investment succeeds.

The developer can reconsider a site. The tenant can negotiate a lease. Investors can price their exposure. A household generally cannot choose a different transmission network when the bill arrives.

Before approving the construction programme, the commission needs an answer to a plain question: if the promised customer does not stay, who is legally obliged to pay what remains?

Records, calculations and unresolved claims

Research completed November 4, 2026. The investigation uses public documents and published research; it does not claim interviews or access to confidential service agreements. Material claims are linked where they appear. Dollar figures are US dollars unless stated otherwise.

Record Why it matters Evidence status
Ohio, 24-0508-EL-ATA Minimum commitments and protection against unused infrastructure Approved tariff as described by regulator and consumer advocate; the advocate also records an industry appeal
AEP Ohio, February 2026 update Requests, paid studies and signed contracts Utility-reported pipeline; future contracts do not establish operating demand
PJM auction records Capacity quantities, clearing prices and delivery years Published auction outcomes; customer exposure varies
Market monitor, October 28, 2026 Wholesale costs through September and the estimated capacity effect Reported market costs plus a separately identified counterfactual calculation
UCS transmission brief Project-level connection costs and allocation criticism Advocacy organisation's analysis; not an adjudicated household-loss total
Virginia, PUR-2025-00058 GS-5 class, minimum payments and credit support Commission's account of adopted provisions, including effective dates and exemptions
Virginia, PUR-2026-00056 and PUR-2026-00131 Transmission allocation and direct-connection policy Adopted framework; detailed implementation handled separately
Louisiana, U-37425 motion Questions about guarantees after the Hyperion financing changes Contested allegations and requests for investigation; no final resolution established here
PJM, October 29 backstop revision Proposed long-term procurement, allocation and credit requirements Proposal, not a completed procurement

The 19% queue conversion, 11.4-times capacity-price comparison, 1 GW energy example, $9.95 unit conversion, and the two simplified financing examples are calculations in this article. Their inputs and assumptions are stated in the text. They are not additional findings by the cited institutions.

The research table compares working papers on their own terms. In particular, the Elhai–Zhou PDF was checked in its October 31, 2026 version; a live working-paper URL may subsequently change. Forecast demand, auction commitments, project capital costs, annual revenues and household bills are different measures. This article does not combine them into a single subsidy total.

Related