Product1 distinct publisher3 min readUpdated
Amazon's Texas plant is permitted for 33 million tons a year and will run only simple cycle turbines. Across the sector, developers are buying speed and skipping the steam loop.
The Product Desk · Product desk

Compiled by The Product DeskSomething wrong?How this is made
Amazon has acquired a large off-grid gas plant under construction in Texas to serve one of its data centers, a facility that WIRED reports could become one of the biggest single sources of greenhouse gas emissions in the United States [1]. The plant is permitted for up to 33 million tons of greenhouse gases a year [2], but the more portable fact for anyone planning capacity is the equipment list: it will rely solely on simple cycle turbines [3]. The distinction is unglamorous and expensive. A simple cycle turbine mixes fuel and air in a combustion chamber and lets off a significant amount of waste heat [3]. A combined cycle unit adds a steam turbine that captures that heat and converts it into more electricity, so it emits less for the same energy [4]. Combined cycle is what on-grid power plants generally use [4]. Data center developers are picking the other one. This is not one company's choice. At least three plants that Williams is developing in Ohio for Meta will use simple cycle units, according to permit applications [5]. Crusoe is building a plant with 39 simple cycle turbines on its Stargate campus in Abilene, Texas, to power a Microsoft data center [6], and the permit application for Google's Goodnight data center at another Crusoe-developed Texas site describes 20 [7] - 59 turbines across the two campuses [8]. xAI's Colossus 1 and Colossus 2 in Memphis are powered by dozens of simple cycle turbines [9]. All of these projects are permitted to release millions of tons of greenhouse gases [10]. Williams and Crusoe told WIRED their gas plants comply with local and national regulations and are fitted with emissions control technology [11]. The reason is schedule, not physics. Britt Burt, a senior vice president at Industrial Info Resources, told WIRED he did not expect developers to lean this heavily on simple cycle: "It's not the most efficient way to use a gas turbine" [12][13]. Installing a handful of simple cycle units is quicker and cheaper than waiting on the extra equipment a combined cycle plant needs, against a yearslong backlog for turbine orders and parts [14]. In Burt's words, "they're having to go with what they can get their hands on" [15]. There is a genuine operating argument alongside the procurement one. Fewer moving parts means simple cycle units start and stop faster, which suits AI training loads that demand large spikes of power [16]. Stephen Lynch, a Penn State mechanical engineering professor who directs the school's Center for Gas Turbine Research, Education, and Outreach, put it this way: "If you have a highly fluctuating load, it's better to have 30 things that can individually turn on and off than to have one giant thing that doesn't like to be turned on and off" [17]. Those spikes already carry a maintenance cost, with many operators reporting turbines breaking from the stress of cycling [18]. Simple cycle also generally needs no water [19], where Burt estimates a 250 MW combined cycle plant could use roughly 75,000 gallons an hour [20]; Amazon's permit application specifically cites simple cycle's lack of water requirement [21]. Then the bill arrives. Permitted emissions on gas plants usually run well above actual emissions [22], but even at half its permit the Amazon plant would out-emit 78 average-sized US natural gas plants, according to the EPA [23] - which implies an average plant under about 212,000 tons a year [24]. Amazon spokesperson Heather Callahan said the Texas plant provides "on-site generation that won't raise electricity costs for Texas families," that the site will include solar and battery storage, and that the company remains committed to net-zero emissions by 2040 [25].
Follow any of these and your For You feed starts watching them — no settings page required.
Ranked by verification strength, evidence, and original report placement.
Amazon has acquired a massive off-grid power plant being built in Texas for one of its data centers that could be one of the biggest sources of greenhouse gas emissions in the US.
Amazon's Texas power plant is permitted to emit up to 33 million tons of greenhouse gases a year.
The project will rely solely on simple cycle turbines, which mix fuel and air in a combustion chamber to create energy, letting off a significant amount of waste heat in the process.
Combined cycle units, generally used at on-grid power plants, add a steam turbine component that captures waste heat and turns it into more energy, so they release fewer emissions for the same amount of energy.
At least three power plants being developed in Ohio by oil and gas company Williams for Meta will also rely on simple cycle units, according to permit applications.
In Abilene, Texas, Crusoe is building a power plant on its Stargate campus to power a Microsoft data center with 39 simple-cycle turbines.
Evidence-backed comparisons of source perspectives and observed adoption signals. Read the methodology
Which Builder, Operator, and Investor concerns the observed source mix emphasized—not a truth score.
Evidence, demonstrated adoption, hype gap, incentives, and confidence are assessed independently, each on its own current evidence. How these are measured.
Documented but single-outlet
The reporting rests on named primary artifacts (permit applications for Amazon, Williams/Meta, Google Goodnight), two named experts, and on-record statements from Amazon, Williams, and Crusoe. However, only one publisher is in the cluster, no permit documents or EPA datasets are reproduced, the per-site permitted tonnages outside Amazon are given only as 'millions of tons', and the turbine-failure claim carries no named source or count.
Widespread across major AI buildouts
Simple cycle behind-the-meter generation is documented at five separate buildouts spanning Amazon, Meta (via Williams), Microsoft and Google (via Crusoe), and xAI, with concrete turbine counts at two Texas campuses. That breadth across independent hyperscalers and developers indicates the spec choice is standard practice rather than an isolated case, though it is all evidenced through one report.
Mildly overstated by permit-ceiling framing
The headline magnitude leans on permitted rather than actual emissions, and the article itself concedes permits usually far exceed realized output; the forward-looking 'parallel fossil fuel grid' and stranded-asset framings are projections rather than measurements. Overstatement is modest because the piece discloses the permit caveat, runs the half-permit sensitivity against EPA averages, and grounds the mechanism in verifiable turbine counts and expert explanation.
Multiple interested parties on record
Every stakeholder quoted has a stake: Amazon's spokesperson frames the plant around consumer bills, co-located solar and storage, and a 2040 net-zero pledge; Williams and Crusoe assert compliance and emissions controls; the analyst works for a commercial energy market intelligence firm selling insight into this buildout; and developers' own incentive to move fast is the article's stated cause of the spec choice. WIRED's framing interest in AI environmental impact is also visible. Nothing in the cluster discloses financial relationships beyond these roles.
Moderate, constrained by single sourcing
Specific, falsifiable details (33 million tons permitted, 39 and 20 turbine counts, 78-plant EPA comparison, 75,000 gallons per hour) and named experts support the core mechanism, and multiple independent deployments point the same way. Confidence is held mid-range because the cluster contains one publisher, no primary documents are reproduced, actual emissions are unknown, and the reliability and stranded-asset elements are the weakest links.
product
Hyperscalers' gas pivot trades a hedged power bet for an unhedged fuel bill1 distinct publisher
leadership
31 years at Microsoft, then a layoff: what tech workers' regrets say about retention1 distinct publisher
build
SemiAnalysis to software teams: your token cost starts at the fab, not the price list1 distinct publisher
leadership
Arm starts selling silicon, and its licensees inherit a supplier-risk problem1 distinct publisher
Distinct publishers with included, body-backed reporting in this cluster.
1 article · August 14, 2026