wired.com web signal

Texas Data Center Boom Uses Dry-Cleaner Air Permit Loophole

TL;DR

  • At least 38 Texas data centers have secured 'minor' air permits since 2024 for on-site power sources, avoiding public notice and environmental review, per Floodlight.
  • Nine of those permits together allow more than 130 million tons of greenhouse gases a year, which the reporting equates to about 35 coal-fired power plants.
  • Stargate alone was cleared to run 10 gas turbines and 62 diesel generators emitting 1.6 million tons of greenhouse gases yearly beside a residential compound.

Texas is quietly building the fossil-fuel backbone of the American AI boom, and it is doing it through a permit form designed for auto body shops and dry cleaners. Wired, in partnership with Floodlight, reports that at least 38 data centers across the state have secured 'minor' air permits since 2024 to run on-site power sources like turbines and backup generators, without the public notice or environmental review a normal power plant would trigger.

The specifics get uncomfortable fast. Stargate cleared 10 gas-powered turbines and 62 backup diesel generators through the same 'permits by rule' and standard-permit route. Those units are currently allowed to emit more than 1.6 million tons of greenhouse gases and 1,000 tons of harmful air pollutants every year, per the permits Floodlight reviewed, and the turbines sit directly beside a residential compound. Across just nine of the permits Floodlight examined, the combined annual allowance tops 130 million tons of greenhouse gases, which the reporting equates to roughly 35 coal-fired power plants.

Why this matters if you are not in Texas: the AI compute buildout is being framed at the federal level as a national infrastructure race, and the regulatory path of least resistance runs through a state that has leaned into being exactly that path. If minor permits keep clearing the queue, the emissions math being talked about in AI keynotes is not the math being filed on the permits, and part of the cost of AI capacity is being paid in local air quality by rural neighbors who never saw a hearing notice.

The honest caveat is that these are permitted maximums, not measured smoke; a turbine allowed to emit that much greenhouse gas is not certain to run at that level, and the reporting does not tell you how many of the 38 sites are actually online yet, or whether federal regulators will step in. What the story does make clear is the direction. If you are watching where AI infrastructure risk actually lands, it is not the model card or the training run, it is a permit filing nobody was told to look at.

Shared on Bluesky by 2 AI experts