In February 2026, AEP Ohio, an American electric utility, published figures showing how far prospective data-center customers had progressed toward securing power.
Developers had initially expressed interest in more than 30,000 megawatts (MW). Of that amount, requests representing roughly 13,000 MW had advanced to paid engineering studies of whether they could connect to the grid. New binding service agreements covered 5,642 MW. A megawatt measures how much power can be supplied or consumed at a given moment. 1
Does that mean all the requests that never became contracts were fictitious?
No. The utility did not publish a project-by-project account showing how much was duplicated or why developers had stopped short of signing contracts. Another 12,219 MW was covered by separate agreements signed before the new tariff. The difference between an early inquiry and a signed agreement is not a measured amount of fake demand.
Yet utilities must make investment decisions before all those plans become real. They need to build substations and reinforce transmission lines where data centers may arrive. If projects are canceled, some of the facilities built for them may go unused. If utilities wait too long, genuine customers may be left without enough power.
What can a utility rely on when deciding what to build?
Ghost demand
Data-center developers need to know how much power a site can receive, and when, before they choose land or decide how much to build.
A company might inquire about several locations at once. It can request a substantial amount of power before it has secured tenants or financing. Once it settles on one site, the other requests may be abandoned. If the project is canceled, none of that anticipated power is consumed.
If utilities count all those requests as future consumption, their forecasts can overstate what will actually be needed. This uncertain prospective load is often called “ghost demand.” 2
Not every such request is dishonest. Comparing sites is part of normal project development, and a sincere proposal can collapse because of financing or permits. That is different from seeking a place in a connection queue without a credible plan to build.
Utilities therefore need to check whether the developer controls the site, can finance the project and will bear some of the cost if it walks away.
The price of reserving power
Ohio introduced a separate tariff to make large data-center developers commit before utilities build expensive infrastructure on their behalf.
Under the rules approved in 2025, qualifying AEP Ohio customers must show that they own the site or have exclusive rights to use it. They also have to agree on construction-cost responsibilities and sign long-term service contracts. If a project is canceled or service is substantially delayed, the developer may have to reimburse specified construction expenses already incurred. 3
Even after service begins, using less power than contracted does not necessarily mean paying a correspondingly small bill. Minimum billing provisions help the utility recover the cost of infrastructure installed to serve that customer.
The tariff also addresses what happens after a data center begins operating.
Suppose a data center relies on its own generators and has arranged for its grid-supplied capacity to be reduced accordingly. If the generators fail, the facility cannot simply replace their entire output with electricity from the grid if that would exceed its contracted limit. AEP Ohio requires customers using this particular arrangement to be capable of promptly reducing their own load. 3
The contract is therefore more than evidence that a project may get built. For some customers, it also sets enforceable limits on what they can draw from the grid after opening.
The costs are contested. Ohio's consumer advocate argues that the provisions help protect existing customers from the expense of projects that never materialize. Data-center industry groups have challenged some of those conditions in court. 4
What $2 billion cannot buy
In July 2026, AEP Texas, part of the same corporate group, reported that it had obtained roughly $2 billion in financial assurances associated with about 40 gigawatts (GW) of prospective large-load projects. One GW equals 1,000 MW. Those assurances included not just cash but also corporate guarantees and letters of credit. 5
Financial backing gives developers a reason to stand behind their requests. It does not mean the grid can deliver 40 GW.
Transmission lines and substations have physical limits. Serving several enormous customers in one area may require new infrastructure. ERCOT, the operator of most of Texas's grid, must study competing requests before available connection capacity can be determined. AEP itself said the amount, timing and cost of connections remained subject to further decisions. 5
Texas rules effective October 8, 2026 likewise separate the applicant's financial commitment from the grid's ability to serve it. Certain large customers must provide financial security before detailed connection studies. If no capacity can be assigned, they can withdraw under specified conditions and recover the balance of their security after relevant costs.
An allocation does not mean the customer can postpone using electricity indefinitely. If the planned start of electricity use slips by at least 24 months, ERCOT may consider reallocating unused capacity in later connection studies. That is not automatic confiscation of a customer's entitlement. 6
AEP Texas obtained its July assurances before this October rule took effect; the two records should not be presented as consecutive steps under one contract.
A developer may be ready to build, but the grid still needs a separate assessment of how much electricity it can supply and when.
When 3,800 MW disappeared
Even an operating data center can cause an unexpected change in demand on the grid.
On July 22, 2026, a transmission fault in northern Virginia prompted nearby data centers to switch from grid electricity to backup generation. According to PJM, the organization that manages the grid and wholesale electricity markets across parts of the eastern and midwestern United States, about 3,800 MW of actual electricity demand suddenly disappeared from the grid. 7
The facilities had not necessarily shut down. They had changed where their electricity came from. But the generators supplying the grid could not instantly reduce output to match the drop in consumption.
When generation and demand move out of balance, voltage and frequency can swing. Protective equipment may disconnect other facilities, compounding the disturbance. If operators cannot arrest a series of disconnections, a local incident could spread into a larger outage.
That danger was visible in Virginia. After the first facilities transferred to backup power, the resulting voltage rise was followed by another roughly 1,100 MW of data-center load switching away from the grid. That additional amount is included in the roughly 3,800 MW total. 7
PJM operators rapidly reduced generation and adjusted voltage controls. The grid-control indicator used to track the imbalance between generation and demand returned to its normal range in about nine minutes. There was no widespread blackout, and PJM reported no material impact on grid reliability.
Successful intervention did not make the pattern harmless. One group of customers leaving the grid had prompted others to follow. PJM subsequently examined stronger technical requirements for large loads.
Ghost demand can lead a utility to plan for electricity that will never be used. Northern Virginia exposed the reverse danger: customers already consuming power can suddenly withdraw that demand from the grid. The first problem can strand investment; the second can threaten real-time grid stability.
When to stay connected—and when to reduce demand
The Virginia facilities switched to backup generation to protect their own equipment. Yet simultaneous protective actions by many customers can make the wider grid less stable. Grid operators consequently seek rules requiring large loads to remain connected through certain brief voltage and frequency disturbances. ERCOT implemented relevant standards in August 2026. 8
A genuine shortage of electricity calls for a different response. To protect hospitals, homes and other essential users, operators may need certain large customers to reduce consumption. Who can issue that order, and which customers go first, must be settled in advance.
PJM proposed registering large customers and introducing priority involuntary curtailment for some new large loads. On October 9, 2026, the US Federal Energy Regulatory Commission (FERC) accepted elements of the definition and registration system for loads of at least 50 MW, but rejected the core priority-curtailment proposal. Among the concerns were whether new customers had been singled out on sufficient grounds and whether the rules reached into states' authority over retail service and disconnection. FERC also required the accepted language to be separated from rejected provisions in a revised filing. 9
There was a further question: could forced curtailment conceal the need for additional generating capacity?
Electricity markets arrange capacity in advance to meet future demand. They can also pay participating customers to voluntarily reduce usage when called upon. But if a market simply removes demand that it plans to curtail involuntarily from its capacity calculations, it may appear to have enough power on paper while still depending on forced customer disconnections in an emergency.
FERC Commissioner David Rosner raised precisely this concern in a separate concurrence. Under the rejected design, shortages could be understated, weakening the market prices that encourage adequate generating capacity to be built or retained. 10
The plan was not implemented, so this is not evidence that prices were actually distorted. The concern nevertheless matters: removing demand from a calculation does not make customers' need for electricity disappear.
Large loads may need to ride through short disturbances without disconnecting. During a genuine supply shortage, they may instead need to cut consumption under agreed rules. Both duties matter.
Who pays?
Suppose a utility upgrades substations and lines for a data center, only for its developer to cancel the project.
The utility may already have paid the contractors. Whether it can recover that spending from the developer depends on the agreement. Ohio's tariff places specified construction costs on developers whose projects are canceled or substantially delayed. 3
Any costs the utility cannot recover from the developer may become losses to the utility. If regulators later allow those costs to be recovered through electricity rates, some of the burden can reach existing households and businesses. Ohio's consumer advocate argues that long-term commitments and minimum bills can limit this risk. 4
Texas's financial-security rules also reduce some risks when applicants withdraw. They do not, by themselves, settle who ultimately pays for every transmission investment. Regulators reserved questions about cost recovery and the rate treatment of forfeited security for separate proceedings. 6
The public data does not show how much, if anything, existing customers in these cases have ultimately paid because of canceled data-center projects. But contract terms and rate decisions determine where that financial risk lands.
What makes a power request credible is not a large number on an application, but a plan establishing when and where electricity will be used—and who is responsible when the promise fails.
Notes and sources
1. AEP Ohio, February 13, 2026: Initial interest exceeding 30 GW, 13,022.7 MW in paid studies, 5,642 MW in new agreements and 12,219 MW in separate earlier agreements are different stages and time frames.
2. Wall Street Journal, October 14, 2026 and Reuters, September 1, 2026: Prior reporting on ghost demand, duplicated requests and financial assurances; these are not claims of an exclusive discovery.
3. AEP Ohio's data-center tariff and PUCO approval notice: Site rights, construction costs, minimum billed demand and immediate load-reduction controls for customers using the specified self-generation arrangement. The latter obligation does not apply to all data centers.
4. Office of the Ohio Consumers' Counsel proceeding: Consumer-cost protection position and industry litigation; any actual savings remain unquantified.
5. AEP Q2 2026 SEC Form 10-Q: About 40 GW of prospective Texas demand and roughly $2 billion of mixed cash, guaranties and letters of credit; neither physical capacity allocation nor service dates were thereby established.
6. Texas Register, 16 TAC §25.194, effective October 8, 2026: Financial security, conditional refunds and possible later reallocation if service is delayed at least 24 months. Broader transmission-investment cost recovery is a separate issue.
7. PJM's official August 11, 2026 incident review: About 3,800 MW of operating load transferred off-grid on July 22, including the later roughly 1,099 MW; the nine-minute interval concerns recovery of a grid-control indicator, not restoration after a blackout.
8. ERCOT NOGRR282: Texas standards on how large computing loads must withstand specified voltage and frequency disturbances. These are not PJM's incident rules.
9. FERC ER26-3515-000 official docket record, Reuters, October 12, 2026 and Utility Dive, October 12, 2026: Acceptance of certain large-load definition/registration provisions and rejection of key priority-curtailment provisions. Editorial source exception: the FERC order's original DOCX text was not read directly; the ruling is corroborated through the official docket and independent reporting, not represented as a clause-by-clause original-text check.
10. FERC Commissioner David Rosner's separate concurrence: Potential competition and capacity-market price signals when involuntary curtailment is treated differently from voluntary demand response; not a finding of observed price distortion or the Commission majority's position.