Texas Puts AI Data Centers on Hold as Grid Queue Swells

Texas has spent years selling itself as the place where big things get built. Now the state is making some of the biggest proposed electricity users wait. Gov. Greg Abbott ordered a comprehensive audit of data centers seeking connections to the ERCOT grid on Aug. 3, then directed the state environmental regulator on Sept. 21 to halt permits sought by data centers until that review is complete.

The immediate trigger is a planning number that stopped behaving like a plan. Texas officials said ERCOT was considering approximately 474 GW of connection requests, more than five times the grid's record peak demand, with about 90% attributed to data centers. That total measures requests in a queue, not power plants humming or servers running. The policy fight begins with the distance between those two things.

Ryan McEntush, an investing partner on Andreessen Horowitz's American Dynamism team, unpacked that distance in an Oct. 7 analysis titled "Why Texas Is Making Data Centers Wait". His argument lands beyond AI: the queue is exposing how the United States decides which industrial projects are real enough for the grid to build around.

A 474 GW queue is not a 474 GW forecast

The queue grew faster than the process designed to study it. In June, ERCOT said it was tracking more than 438,000 MW of large-load requests, nearly 89% from data centers. By Aug. 3, the governor's office cited approximately 474 GW and said roughly 90% of new power requests came from data centers.

Those figures are real as administrative demand, but they are not proof that 474 GW of facilities will be financed, built, and energized. Developers can evaluate multiple sites before choosing one. Some projects may lack a committed customer, land control, equipment, financing, or a credible power plan. Counting every request as future load gives planners an enormous number with uneven evidence behind it.

That distinction matters because grid studies produce real investment decisions. Utilities and regulators use them to determine where transmission lines, substations, transformers, and generation must be added. If a speculative project drives an upgrade and disappears, somebody else can inherit the bill. If planners discount too much demand, a credible project can wait years for capacity that should have been prepared earlier.

Batch Zero turns project maturity into evidence

The Public Utility Commission of Texas approved ERCOT's Batch Zero process in June. Instead of studying every large request in isolation, ERCOT can evaluate a group of projects against the same network constraints and allocate capacity among them. The process is designed to identify which projects have enough maturity for the grid to rely on their assumptions.

Abbott's Aug. 3 directive widened that test. It ordered regulators to collect information about ownership, public incentives, projected electricity use, on-site generation, water consumption, cooling systems, and community protections. ERCOT's data-center request for information applies to developers pursuing projects of 25 MW or more, with responses due Oct. 12 and a report planned by Dec. 10.

This is where the tech news story becomes an operating story. An interconnection request is no longer enough. A developer has to show that the customer, capital, land, equipment, water, and power strategy belong to one executable project. The queue still records ambition, but Batch Zero asks whether the evidence can survive contact with the transmission system.

Bringing your own power only moves the bottleneck

Many developers are responding to long grid timelines with behind-the-meter generation, batteries, or hybrid campuses that can operate before a full connection arrives. That can shorten time to power, which matters when expensive chips are sitting idle. It does not make the grid irrelevant.

Large campuses still value grid service for redundancy, economics, and the ability to draw from a much larger pool of generation when an on-site plant fails. A data center that promises flexible demand can help, but that promise has to be measurable and available when ERCOT calls for it. A campus that can curtail load for a few hours is different from one that can provide firm flexibility through a prolonged grid emergency.

The hardware constraint also keeps moving. More generation still needs turbines, engines, fuel, inverters, transformers, switchgear, transmission, crews, and permits. McEntush notes that even campuses planning private generation often expect to connect later because grid power is generally cheaper and more resilient than running an island indefinitely.

The real infrastructure test is who pays

Texas began this policy turn before the audit. On June 10, Abbott directed the PUCT and ERCOT to protect residential ratepayers from data-center infrastructure costs. The Sept. 21 permit directive added water, environmental review, and community impact to that economic test.

The state's problem is not simply whether AI data centers consume a lot of electricity. It is whether developers reserve capacity honestly, fund the upgrades tied to their projects, and bear the consequences when a project slips or disappears. A credible project should not wait behind a stack of paper campuses. A household should not finance transmission built for a campus that never arrives.

That makes Texas an early case study for a much larger industrial question. Data centers hit the wall first because their loads are unusually large and their timelines unusually aggressive. Factories, electrified transport, advanced manufacturing, and other large users will eventually run into the same equipment and transmission constraints.

What the December audit will reveal

The audit's most important output will not be a smaller headline number. It will be a more trustworthy queue. Regulators need to identify unique projects, verify their maturity, quantify their real grid dependence, and separate near-term load from optional site positions.

For developers, the commercial lesson is already visible. Power strategy now sits beside capital, customers, and product as a gating function. Deposits, interconnection milestones, flexible-load commitments, on-site generation, and community agreements are becoming part of the proof that a campus can move from proposal to operation.

For infrastructure startups, the opening is equally specific. The market needs equipment and software that can add capacity, measure flexible demand, improve transmission utilization, and keep large electronic loads stable. The tech news will keep circling AI, but the durable opportunity may belong to the companies that make power commitments credible enough for planners to use.

Texas's December report will show how much of the queue can carry that weight. Every project removed from the speculative pile sharpens the signal for the ones still standing, and every verified project makes the next question harder to dodge: how quickly can the physical grid catch up?