Texas Froze New Data Center Power Hookups. "Ghost Demand" Is the Real Story.

In early September 2026, ERCOT, the grid operator covering most of Texas, did something no major U.S. grid operator had done before: it stopped processing new requests from large power users, including data centers, to connect to the transmission grid. The trigger was a number. Interconnection requests sitting in ERCOT's queue had reached roughly 474 gigawatts, more than five times Texas's own record peak electricity demand. Nationally, it's worse: a review of utility and grid filings found electricity requests from very large power users, overwhelmingly data centers, topping 700 gigawatts across the Midwest, the Mid-Atlantic and the South.

Most coverage ran with some version of "AI is breaking the power grid." That framing is wrong in a specific, useful way. The queue isn't full because data centers actually need 474 gigawatts. It's full of what utilities now call ghost demand: requests that cost almost nothing to file, and that until this year nobody checked before counting.

Ghost demand has a simple cause: filing is free

The industry has a name for what's clogging the queue: ghost demand, sometimes called phantom demand. A developer files an interconnection request for a data center campus at a specific substation. In most queues, filing costs little more than a modest fee and some paperwork. There's no requirement to show a signed tenant, a funding source, or even confirmed land rights at the time of filing. So developers file everywhere.

Three distinct behaviours all inflate the same number. Legitimate developers hedge: they file at three or four substations for one project, intending to build at whichever site clears fastest and cancel the rest, but the queue counts all four as live demand until they withdraw. Landowners file speculatively to make a parcel more attractive to a buyer, with no plan to build anything themselves. And undercapitalised developers file because a queue position has resale value on its own; someone else may eventually pay to take it over. None of this is fraud. All of it inflates the figure regulators use to plan a grid.

The scale of the drift is the tell. Texas's queue went from roughly 48 gigawatts in 2023 to 474 gigawatts now, close to a tenfold jump in three years. Actual U.S. data center electricity draw, across the entire existing fleet, is estimated at 60 to 70 gigawatts at any given moment. The national interconnection queue is now more than ten times the size of what's actually running.

The 474 gigawatt figure, unpacked

No utility has produced a state-specific estimate of how much of the Texas queue is real; that's what ERCOT's audit, due by December 2026, is meant to produce. The closest available data point comes from Exelon, one of the larger utility holding companies managing this kind of load: it expects only 22 percent of its own 65-gigawatt data center pipeline through 2040 to actually get built.

Read that carefully. It doesn't mean 78 percent of data center announcements are lies. It means that in a queue with no cost to enter, a request surviving from filing to energised campus is the exception, not the rule. Every planning document built on the raw queue number was, in effect, planning around demand that was never going to show up.

It isn't only a Texas problem

Texas got the headlines because ERCOT's freeze was the most visible single action, but the same distortion shows up wherever a grid operator lets developers file for free. The Reuters review that put the national figure above 700 gigawatts covered utility territory well beyond ERCOT, including several Midwest and Mid-Atlantic operators managing the same kind of speculative pileup. PJM Interconnection, which serves roughly thirteen states from Illinois to the Mid-Atlantic, has spent the past two years tightening its own queue reform rules for exactly this reason, well before ERCOT's freeze made the problem a national story. What's different about Texas is the scale and the bluntness of the response: an outright pause on new agreements, rather than a slower phase-in of new rules.

That timing gap matters. Utilities that started reforming their queues in 2023 or 2024 are now working from figures that are already somewhat cleaner. ERCOT is starting its cleanup from a number that grew nearly tenfold before anyone applied a filter, which is part of why the freeze, rather than a gradual reform, was the tool it reached for.

This isn't a demand problem. It's a missing cost function

Anyone who has designed a system that allocates a scarce, contested resource has seen this failure shape before. A queue with free entry and no penalty for abandonment will always fill with more requests than the resource can serve, not because demand is infinite, but because there's no cost to speculating. TCP handles this with SYN cookies and congestion windows. Cloud schedulers handle it with quotas and preemption. Even ticketing systems handle it, badly, with bot detection.

Interconnection queues had exactly that shape until this year. Filing was nearly free. There was no meaningful penalty for withdrawing a request after consuming years of a utility's study time and engineering capacity. And utilities themselves had a quiet incentive not to fix it: larger demand forecasts justify larger capital projects, which is how a regulated utility grows its rate base and its return. Nobody in the chain was pricing in the cost of a fake request, so fake requests kept accumulating until a number too large to ignore showed up in Texas.

None of this means AI data center growth is a myth. Hyperscalers are placing real, multi-billion-dollar capital commitments and real transformer and turbine orders, both of which show up as visible, checkable demand well before a facility is finished. The distinction that got lost in the "474 gigawatts" headline is between demand backed by a purchase order and demand backed by a paperwork filing. Both count as gigawatts in a queue. Only one of them counts as electricity a grid actually has to deliver.

What utilities are actually changing

The reforms rolling out through 2026 are, functionally, admission control for a physical grid.

Several large utilities now require a real cash deposit before a project enters the queue at all, money that's forfeited if the project doesn't proceed, which prices in a cost that used to be zero. Others require proof of site control or a confirmed equipment order before a filing counts; transformers and turbines both carry multi-year lead times, which makes a signed purchase order a decent proxy for whether a developer is actually building. ERCOT's own response was blunter still: freeze new interconnection agreements above a size threshold entirely, and spend the rest of the year auditing what's already in the queue.

What it means if you're planning on cheap AI compute later

The immediate effect for anyone buying or renting GPU capacity isn't more power on the grid, it's a more honest planning number. A grid operator that finally knows its real demand can commit multi-year transmission and generation projects against demand that's actually going to show up, instead of demand that was filed and forgotten. Over several years, that should shorten the waits that have become normal for large compute buildouts, because capital stops being spread across four times more "demand" than will ever draw a watt.

In the near term, expect the opposite: fewer new interconnection agreements while queues get cleaned out, more upfront capital required from legitimate developers who used to file for free, and utilities publicly revising their data center demand forecasts downward, not because AI buildout is slowing, but because the number being revised was never real. Anyone reading a utility's public data center demand forecast this year should ask the Exelon question: what fraction of this is actually going to get built.

This isn't the first time a free-to-join queue inflated a forecast everyone believed anyway. Fibre build-out ahead of the dot-com crash followed a similar shape, on a slower clock. Grid interconnection is the same failure playing out on an earnings-call timeline instead of a multi-year one, and this time, the people managing the queue noticed while there was still time to fix the incentive, not just count the wreckage afterward.