Four states now publish how much announced data center demand they actually expect. All four throw out most of it.

Four states now publish how much announced data center demand they actually expect. All four throw out most of it.

Key points

  • Seven filings from four states each remove 50% to 95% of the data center demand put to them.
  • AEP Ohio's inquiries fell from 30,000 MW to 5,642 MW after a tariff made developers pay for reserved capacity.
  • Pennsylvania has more than 100 announced projects and 5 that hold every permit.

Between December 2025 and August 2026, grid operators, utilities and regulators in four states each published their own count of how much of the data center demand put to them they actually expect to serve. Seven such filings exist, none of them using the same method, and each removes between 50 and 95 percent of what it was asked for. The largest single drop belongs to AEP Ohio, which reported inquiries totaling more than 30,000 megawatts in February 2026 and, after a tariff required data centers above 25 megawatts to pay for at least 85 percent of the power they subscribe to whether or not they use it, put expected load at 5,642 megawatts. None of these figures is new. The pattern across them is.

There is a way to test the AI bubble argument that does not rely on anyone's view of stock valuations, and it sits in utility paperwork recording which developers have asked for electricity and how far those requests have progressed. Those filings have been accumulating for most of a year, and read together they say something the individual releases did not.

Take the oldest and plainest of them. AEP Ohio supplied both ends of the comparison on Feb. 13, 2026. More than 50 customers had approached the utility about over 90 sites seeking a combined 30,000-plus megawatts. The amount AEP expected to materialize was 5,642 megawatts.

One thing changed between those figures. On July 9, 2025, the Public Utilities Commission of Ohio approved a data center tariff that applies to customers above 25 megawatts. Under it, a developer pays for a minimum of 85 percent of the energy it subscribes to whether it uses that energy or not, over a 12-year term that includes a four-year ramp, with an exit fee and proof of financial viability up front. Four out of every five megawatts of stated interest did not survive being asked to pay for itself.

The same thing appears elsewhere once utility records replace project announcements. A data center can demand as much electricity as a mid-sized city, and access to that electricity is not automatic. The developer enters a queue while the grid operator studies whether enough generation and transmission capacity exists. Every stage asks for a firmer commitment. A press release is cheap. A grid study requires paperwork. An interconnection agreement creates obligations. Collateral requires cash.

What the other states found

The pattern holds across operators who do not coordinate, use different methods and answer to different regulators.

Who is screeningAsked forSurvivedCut
Pennsylvania DEP100+ projects5 fully permitted95 percent
Georgia Power76,200 MW pipeline12,400 MW committed84 percent
AEP Ohio30,000 MW of inquiries5,642 MW81 percent
PPL, Pennsylvania31,800 MW advanced pipeline11,000 MW signed65 percent
ERCOT, 2030 forecast47,783 MW19,560 MW with agreements59 percent
ERCOT, Batch Zero screen498 GW205 GW study-eligible59 percent
NERC, on Texasrequested loadabout half50 percent

Georgia Power files its version quarterly with the state Public Service Commission, which makes it the most auditable of the set. Its Q1 2026 report put the pipeline at 76,200 megawatts and firm commitments at 12,400 megawatts across 31 customers. PSC staff testimony records 33 projects totaling 11,332 megawatts leaving the pipeline since 2023.

Texas is the largest number and the loudest fight. ERCOT is fielding about 474 gigawatts of connection requests, close to 90 percent of it data centers, against an all-time record demand of 85,508 megawatts across the entire grid. Governor Greg Abbott ordered an audit of the queue on Aug. 3, directing the Public Utility Commission and ERCOT to verify every project "before any additional data centers are approved to move forward." Projects that fail, his letter says, "must be denied connection to the Texas grid."

Pennsylvania published its own count on Aug. 18, inside the executive order in which Governor Josh Shapiro barred the state from permitting any data center that lacks local approval. More than 100 projects appear in public databases. Fifty-eight have engaged with the environmental regulator. Fifteen have filed an application. Five hold every permit needed to build a first phase.

A different claim from the one SemiAnalysis rejected

One widely circulated estimate says about half of the US data center capacity planned for 2026 will be delayed or cancelled. SemiAnalysis directly challenged that figure in June. It argued that the underlying estimate for capacity already under construction was wrong by multiples and that the calculation counted early announcements as real projects. SemiAnalysis changed its own North American year-end 2026 forecast for hyperscaler self-build by about 1 percent over six months.

That criticism lands, and it is worth stating plainly that the numbers above are a different measurement. SemiAnalysis defines a genuine project as one with secured land, a defined power solution and approved permits, and says those are largely on schedule. Every screen in the table is a test of exactly that transition. The question is not how many press releases die, which has always been most of them. It is how much of the announced pipeline ever becomes a project that clears those three bars.

The timing follows from that distinction. Interconnection agreements and permits usually lead concrete work by 18 to 36 months. These figures therefore say little about capacity scheduled for this year. They say more about 2028 through 2030, the years embedded in current valuations.

Companies affected by the screening

PPL (PPL) is the most exposed name in Pennsylvania and reports the clearest version of the funnel: 31,800 megawatts of data center pipeline in advanced stages of planning, more than 11,000 megawatts under signed electric service agreements, and more than 6,500 megawatts actually under construction.

A rule that slows new entrants does not much trouble a project that is already sited and locally supported, which is the part of this most likely to be misread. Talen Energy (TLN) sold its 960-megawatt Cumulus campus next to the Susquehanna nuclear plant to Amazon (AMZN) for $650 million and expanded that into a 1,920-megawatt supply agreement running to 2042. Constellation (CEG) is restarting Three Mile Island Unit 1 as the Crane Clean Energy Center. Both are existing licensed sites rather than greenfield fights. At Homer City, a 3,200-acre former coal plant east of Pittsburgh planned at 4.5 gigawatts and about $10 billion, GE Vernova (GEV) is supplying seven gas turbines and EQT (EQT) has an agreement in principle to supply the gas. Environmental groups are crowdfunding a legal challenge to that site, so it is not untouched.

Then there is the group that moves on these headlines without owning anything in the state involved. When New York paused state permits for data centers on July 14, TeraWulf (WULF) finished as the biggest decliner among the AI infrastructure names we track and Nebius (NBIS) was close behind, despite Nebius having no connection to New York. CoreWeave (CRWV), IREN (IREN), Cipher Mining (CIFR), CleanSpark (CLSK) and Core Scientific (CORZ) tend to trade the same way. That is a sentiment reaction, not a change in any of their businesses, and the two are worth telling apart before treating a move as information.

Evidence on the other side

The same filings also show that substantial capacity is real and already being built. On July 30, AEP (AEP) told investors it had 45 gigawatts of contracted load growth in Texas supported by fully executed letters of agreement. The company had collected almost $2 billion in cash or collateral, covering all required credit support for that amount. Georgia Power says 8,500 megawatts of committed demand is under construction and 10 projects are operating.

Local defeats also do not always stick. Saline Township, Michigan rejected a $16 billion campus four votes to one in September 2025, was sued two days later for exclusionary zoning, settled within weeks for about $14 million in community benefits, and saw ground broken in November. And the political current runs both directions: West Virginia enacted a law in 2025 preempting local ordinances that limit large data center projects.

What would change the reading is the direction of travel rather than any single number. If the share converting from pipeline to signed agreement is stable, this is just the ordinary gap between a sales funnel and a backlog. If it is falling, the out-year capacity everyone is underwriting gets smaller quietly, and long before it shows up in anyone's guidance. We are tracking the seven screens here, and will add readings as new filings land, which for most of these bodies means quarterly at best.

The Ohio number is the one worth keeping in mind, and not because it is the biggest. It moved because of a tariff rather than a protest. Nobody blocked those projects. They were asked to pay for the electricity they had reserved, and more than 24,000 megawatts of it went away.

Sources

This is general market commentary, not investment advice.

Frequently asked questions

Why did AEP Ohio's data center demand fall from 30,000 MW to 5,642 MW?

Because developers were asked to pay for the power they had reserved. On July 9, 2025 the Public Utilities Commission of Ohio approved a tariff covering data centers above 25 megawatts, requiring them to pay for a minimum of 85 percent of the energy they subscribe to whether or not they use it, over a 12-year term with an exit fee and proof of financial viability. AEP Ohio had fielded inquiries totaling more than 30,000 megawatts and afterwards put expected load at 5,642 megawatts.

Is the AI data center boom a bubble?

This reporting does not answer that, and no single figure can. What the filings show is narrower: the demand numbers underpinning AI infrastructure valuations are consistently between two and twenty times larger than the demand that has passed a binding test such as a signed interconnection agreement, a permit or posted collateral. Whether that gap is a bubble or the ordinary difference between a sales pipeline and a signed backlog depends on whether the conversion rate is falling. This is general market commentary and not investment advice.

Which stocks are exposed to data center permitting rules?

PPL is the most directly exposed, reporting 31,800 MW of Pennsylvania data center pipeline against 11,000 MW under signed service agreements and 6,500 MW under construction. Talen Energy (TLN) and Constellation (CEG) own existing licensed nuclear sites serving data centers, and GE Vernova (GEV) and EQT supply turbines and gas to Pennsylvania's largest campus. Separately, Nebius (NBIS), TeraWulf (WULF) and CoreWeave (CRWV) often move on data center policy headlines without owning anything in the state involved.

Does slower permitting mean fewer AI data centers get built?

Not in the near term. Permits and interconnection agreements run 18 to 36 months ahead of construction, so current screening says little about capacity arriving this year and more about 2028 through 2030. Large projects are still being built: AEP reports 45 gigawatts of contracted Texas load backed by nearly $2 billion in posted cash or collateral, and Georgia Power has 8,500 megawatts under construction with 10 projects already online.

More on PPL and CEG

Dennis Singleton
Dennis Singleton

Dennis Singleton has spent years following the markets, but what keeps his attention is how AI is built. He writes about the companies behind the technology, from semiconductor designers and advanced packaging to photonics, memory, networking, and the hardware powering modern AI. His approach starts with filings, earnings, and industry research, then translates the important details into clear, straightforward analysis without unnecessary hype.