The ceiling written on the permit
Amazon has acquired an off-grid power plant under construction in Texas to serve one of its data centres, and the permit for the site allows up to 33 million tons of greenhouse gases a year. Permitted emissions on gas plants normally run well above what a plant actually releases; even so, at half that ceiling the facility would still put out more greenhouse gas than 78 average-sized natural gas plants, on the EPA's comparison. The 33 million tons is therefore a legal maximum, and the measured annual figure that would test it does not exist yet.[1]
This column drew the same line on 6 August, reading two energy companies' data-centre power disclosures and finding that the only firm quantity was the contract term, that the capacity and emissions figures stood as permit ceilings, and that grid interconnection was the next decisive step. In Texas, Amazon repeats the pattern and adds a wrinkle: the plant is islanded, so there is no interconnection date to wait for, and the company says only that the design is to transition to grid-connected service as interconnection timelines allow.[1], [2]
What does the simple cycle fail to give back?
A simple-cycle turbine mixes fuel and air in a combustion chamber and lets the waste heat go. A combined-cycle unit adds a steam turbine that captures that heat and turns it into more electricity, which is why an on-grid gas plant emits less for the same energy. Amazon's Texas plant relies solely on simple-cycle turbines, so the efficiency step a combined-cycle design would supply never enters the site at build time, and no later operating choice can add it.[1]
The choice extends well past Amazon. Permit applications show at least three Ohio plants that the oil and gas company Williams is developing for Meta running on simple-cycle units. Crusoe's Stargate campus in Abilene, Texas is being built with 39 simple-cycle turbines to power a Microsoft data centre, and the permit for Google's Goodnight site, also developed by Crusoe, describes 20. xAI's Colossus 1 and Colossus 2 in Memphis run on dozens more. Each of these is permitted to release millions of tons of greenhouse gases, and Williams and Crusoe both say their plants meet local and national rules and carry emissions control equipment.[1]
The queue is on turbines, the bill lands in emissions
Britt Burt of Industrial Info Resources, who tracks the energy market, says he did not expect data-centre developers to lean this hard on simple-cycle units, and calls it not the most efficient way to use a gas turbine. His explanation is a supply one: with a yearslong backlog on turbine orders and associated parts, operators are having to go with what they can get their hands on, and a handful of simple-cycle machines installs faster and cheaper than the extra equipment a combined-cycle plant needs. On that reading the binding constraint has moved onto turbine delivery, and efficiency is what gets traded away to clear it.[1]
Two other quantities sit alongside. A rapidly swinging load favours many small machines: Stephen Lynch, who directs Penn State's Center for Gas Turbine Research, Education, and Outreach, puts it as 30 units that can individually turn on and off against one giant unit that dislikes being cycled, and AI training draws exactly that kind of spiky demand — spiky enough that operators have reported turbines breaking under the stress. Water is the second: simple-cycle units generally need none, while Burt estimates a 250-megawatt combined-cycle plant could use about 75,000 gallons, roughly 284,000 litres, per hour. Amazon's permit application lists that absence of water use as an environmental benefit.[1]
That leaves one observable step. Simple-cycle peaker plants have a legitimate place on the grid, supplying power quickly when demand runs high. What is new is a set of islanded plants built with the same machines and designed to last decades, which amounts to a parallel fossil fleet rather than a peaking reserve. Amazon says the Texas site is designed to transition to grid-connected service as interconnection timelines allow, and Williams has said it may consider its Ohio plants for interconnection later. If a public interconnection filing for the Texas site appears by 31 December 2026, the plant leaves the islanded rung; if none does, the permit ceiling stays the only published number attached to it. Burt supplies the other end of the range: if the AI buildout slows and energy demand dips, these plants become what he calls a stranded asset.[1]