A 35-turbine, 7.65 GW gas plant permitted to emit 33 million tons of CO₂ a year will power an Amazon data center in Pecos County, Texas. The permit is an authorization, not a projection. The structural story is the move from buying grid power to generating it behind the meter, which answers the ratepayer objection completely and the emissions objection not at all.
The New York Times reported on 8 August that a natural gas plant being built to power an Amazon data centre in Pecos County, in rural West Texas, holds a state permit allowing it to release 33 million tons of carbon dioxide a year. If it operated at that ceiling, it would be the single largest emitting power plant in the United States — by a wide margin. The dirtiest plant currently running, the coal-fired James H. Miller Jr. facility in Quinton, Alabama, emits roughly 16 million tons.
The plant is designed around 35 gas turbines producing a combined 7.65 GW. It will not connect to the local utility grid. The emissions permit was issued by the state of Texas in January, to an entity called GW Ranch — Amazon is the offtaker and backer, not the permit holder, which is a distinction that will matter when the objections start.
Before anything else: a permit to emit 33 million tons is an authorisation, not a forecast. It is the maximum the operator is legally allowed to release under the terms of the permit, typically calculated from nameplate capacity running at a high assumed utilisation. Actual emissions will depend on how many turbines get built, how often they run, at what load, and on the plant's realised heat rate.
This distinction is load-bearing and most aggregated coverage of this story will collapse it. Headlines saying Amazon's data centre will emit 33 million tons of CO2 are not reporting the permit — they are reporting a worst case as though it were a projection.
The permit is not a lie and it is not a plan. It is the size of the box the operator has asked permission to fill.
That said, the ceiling is not arbitrary either. Operators do not request permits far above what they intend to use, because permitting is expensive, contested and slow, and because a permit application that overstates capacity invites scrutiny. A 33-million-ton authorisation is evidence about the scale of what is being contemplated. It is the strongest available public evidence about that scale, in fact, which is why it is worth reporting even though it is not a measurement.
The honest framing: the plant is permitted for more annual CO2 than any generating facility in the country, and roughly double the largest one currently operating. What it actually emits is unknown and will remain unknown until it runs.
The number is the headline. The structural shift underneath it is the story, and it generalises well beyond Amazon.
Until recently, a hyperscaler building a data centre was an electricity customer. It negotiated with a utility, waited in an interconnection queue, and inherited the carbon intensity of whatever grid it plugged into. Its emissions were an average of somebody else's fuel mix, and its leverage was the leverage of a large ratepayer.
A hyperscaler that builds behind-the-meter generation is a different entity. It sets its own fuel mix. It skips the interconnection queue, which in ERCOT and elsewhere has become the binding constraint on how fast anything gets built. It controls its own commissioning schedule rather than a utility's. And — this is the part that deserves attention — it changes which political fight it is in.
The dominant public objection to data centres over the past two years has been electricity prices. Residential ratepayers in several markets have watched bills rise and been told, correctly or not, that data centre load is a cause. That argument is fought in public utility commission rate cases, which are adversarial, on the record, and open to intervenors including consumer advocates, municipalities and environmental groups.
A behind-the-meter plant moves the fight to air permitting. Air permitting proceedings are real and can be contested — but they have far fewer participants, narrower standing, more technical thresholds, and in Texas they are administered by a state agency with a track record of approving them. Self-generation converts a public utility fight into a private permitting fight, and the second kind has a much smaller gallery.
Amazon's statement on the project is one line, and it is precisely engineered. The data centre, the company said, will be powered by new on-site generation that won't raise electricity costs for Texas families.
That is very likely true. A plant that never touches the utility grid does not appear in a rate base, does not require transmission upgrades socialised across ratepayers, and does not compete for grid supply in a way that would move wholesale prices for households. On the electricity-price objection, the statement is a complete and accurate answer.
It is also entirely silent on emissions. Not evasive — silent. And that silence is structural rather than rhetorical, because on-site generation solves the price problem by creating the emissions problem. The same design choice that keeps the plant out of the rate base is the choice that makes it a dedicated fossil generator rather than a claim on a grid that is, in aggregate, decarbonising.
Read that way, the statement is not spin. It is an answer to a specific question, offered in place of a different one.
Amazon co-founded The Climate Pledge and committed to net-zero carbon by 2040. Its 2025 sustainability reporting shows emissions of roughly 81 million metric tons of CO2 equivalent, up 16% year over year and up 58% from 2019 — the year the pledge was made. Amazon attributed the increase substantially to data centres: indirect emissions from purchased electricity rose 34%, and Scope 3 emissions, which include data centre construction and now account for around 76% of the company's footprint, grew 20% in a year.
An Amazon spokesperson offered both halves of the position: the world looks different now than when we co-founded the climate pledge, and our commitment hasn't changed. Those two sentences are in tension and the company knows it. They are also both defensible — the AI buildout genuinely was not foreseeable in 2019, and a pledge that is harder to keep is not thereby abandoned. Readers can weigh them without help.
What is worth insisting on is that this is a pattern, not an Amazon anomaly. SpaceX confirmed on 7 August that its Terafab chip facilities will run on natural gas plants rather than Tesla solar. New York State halted construction of all new data centres in July while it works out how to handle the load. Across the industry, the same calculation is being made: the constraint is time-to-power, gas turbines are the fastest firm capacity available, and climate commitments were written against a demand curve that no longer applies.
The generalisable question is what a corporate net-zero pledge is worth when the denominator triples. One reading is that it is worth very little — a promise contingent on convenience. The other, which deserves equal airing, is that a company growing this fast while holding an absolute 2040 target has set itself a harder problem than one that quietly revised the date, and that the honest test is not this year's number but whether the capital going into nuclear, geothermal and long-duration storage arrives in time to matter.
Two things to hold on to. First, do not repeat 33 million tons as a prediction — it is a permit ceiling, and the difference between a ceiling and an outcome is where all the uncertainty lives. Second, the more durable development is that the largest computing companies in the world have started building their own power plants. That decision determines not only what gets burned, but who is in the room when it is decided.
No. A permit sets the legal maximum, typically calculated from full nameplate capacity at high assumed utilisation. Actual emissions depend on how many of the 35 turbines are built, how often they run and at what load. The figure is the strongest public evidence of the project's intended scale, but it is an authorisation, not a forecast.
Power, permits, water and land — DrafterDaily tracks the infrastructure decisions behind the models, and what they actually cost.
Subscribe freeCVE-2026-0768 is an unauthenticated root RCE in Langflow. It was disclosed in January, the fix has shipped through seven releases, and attackers are hitting it in September — because low-code AI middleware became critical infrastructure without acquiring a patch owner.
Opposition to data centres is not a referendum on AI. It is a cost-allocation dispute — and PJM's capacity auction is the mechanism that turned an abstract argument into a number on 67 million households' bills.
OpenAI's first published benchmarks for its custom inference chip are real, from a public benchmark, and normalised on nameplate TDP rather than measured power. OpenAI disclosed that itself, in a sentence nobody is quoting.