Generated by Codex with GPT 5.6 Sol XHigh

Techmeme surfaced Hiroko Tabuchi’s August 8 New York Times report, “New Amazon Data Center Is Set to Have the Most Polluting Power Plant in the U.S.,” which connects Amazon to a private power project of extraordinary scale. Amazon confirmed that it acquired the GW Ranch site in Pecos County, Texas, and plans to build an AI data center campus there. The campus would initially get its electricity from an on-site natural-gas plant rather than the Texas grid.

The headline number is 7.65 gigawatts: the combined nameplate capacity of 35 permitted gas turbines. Texas regulators describe the plant’s nominal delivered output as 5 gigawatts, so 7.65 gigawatts should not be read as continuous production. Even with that qualification, GW Ranch would be comparable to the electricity demand of a major city. It is not merely a backup generator attached to a server farm; it is a power system built around computing demand.

That distinction explains why the project matters. AI companies are no longer just buying electricity from the energy system. To secure enough power quickly, they are beginning to build private energy systems of their own—and making choices about generation, pollution, water, and infrastructure that were once made mainly by utilities and public regulators.

The scarce resource is time to power

Amazon’s decision is best understood as a response to an infrastructure bottleneck. AWS is growing quickly, and Amazon has said demand for computing capacity exceeds the data centers it can supply. New utility connections can take years as grid operators study transmission needs, allocate costs, and wait for new lines and substations. A plant located behind the data center’s meter lets construction proceed without waiting for that process.

The Texas Commission on Environmental Quality’s project summary says GW Ranch will not be capable of buying from or selling electricity to the local utility system. Amazon argues that this separation means Texas households will not subsidize the campus through higher electricity rates. The company says the project will use new on-site generation and is designed to connect to the grid later, once interconnection timelines allow.

That is a real benefit, but it addresses only one kind of cost. Keeping the plant off the grid can separate its electricity bill from other customers; it cannot keep its emissions inside the property line. Nor does private construction remove public questions about air quality, water use, land, pipelines, or the long operating life of new fossil-fuel equipment.

The structure also changes the economics of AI development. A company that controls both the data center and its electricity supply can treat power availability as a competitive advantage. The limiting factor is no longer just chips or models, but how fast the company can assemble turbines, fuel, cooling, buildings, and network connections into one functioning campus. AI infrastructure is becoming vertically integrated all the way down to energy.

A permit ceiling, not a forecast

GW Ranch’s air permit allows roughly 33.2 million tons of carbon-dioxide-equivalent emissions per year. At that level, it would emit more climate pollution than any existing US power plant. The permit also lists thousands of tons of carbon monoxide and nitrogen oxides, along with particulate matter, volatile organic compounds, and other pollutants.

The word “allows” is essential. A permit establishes the maximum legal emissions under the approved configuration; it does not predict how often all 35 turbines will run or what the plant will actually release. Power plants commonly operate below their permit ceilings, and the project may be built in phases. Calling 33 million tons an assured annual result would therefore overstate what the documents show.

The ceiling is still consequential. It reveals the scale regulators have authorized and the upper bound Amazon has preserved for future operation. It also makes the tradeoff legible: the company is buying speed and dedicated capacity with a design that could lock in substantial gas consumption for years. Actual emissions will depend on how much of the plant is built, turbine utilization, efficiency, pollution controls, and the contribution of other power sources.

Developer Pacifico Energy says the broader private grid could also include 750 megawatts of solar generation and 1.8 gigawatts of battery storage. Those resources could reduce some gas use, especially during sunny hours, but they are much smaller than the permitted turbine fleet and do not by themselves turn the campus into a carbon-free system. Amazon also says it has enabled 10 gigawatts of carbon-free energy across 40 projects in Texas and that GW Ranch will use non-potable brackish groundwater rather than water suitable for drinking or irrigation. Those claims address important parts of the project’s footprint, but detailed operating data will be needed to show the resulting mix in practice.

The climate pledge meets the AI buildout

Amazon still promises net-zero carbon emissions across its operations by 2040. Its own 2025 Sustainability Report, however, says absolute emissions rose 16 percent from 2024 to 2025, even as carbon intensity remained 38 percent below 2019. The two measures tell different stories: Amazon is producing fewer emissions per unit of business than it once did, but growth is pushing the total footprint upward.

GW Ranch makes that tension physical. Annual renewable-energy matching can support new wind and solar projects, but it does not mean every data center is powered by carbon-free electricity every hour. A private gas plant beside an AI campus exposes the difference between balancing energy purchases over a year and supplying a specific workload in real time.

The project is also part of a wider shift. The original Cleanview investigation counted nearly 60 behind-the-meter gas projects announced since the start of 2025, totaling about 90 gigawatts, though many may never be completed. A July Texas Tribune analysis found 74 proposed US gas plants of at least 100 megawatts dedicated to data centers, including 32 in Texas. The pattern suggests that off-grid gas is becoming an industry response to slow grid expansion rather than an isolated Amazon exception.

The cleanest way to judge GW Ranch is therefore through milestones rather than slogans. Observers should track how many turbines are actually installed, how heavily they run, whether the promised solar and storage arrive, when the campus connects to the grid, how much water it consumes, and what its measured emissions are. Those facts will determine whether the permit describes an extreme contingency or the normal operating model.

The larger lesson is that the AI race is now an energy race. Building private generation can keep compute expansion moving and shield other customers from some direct grid costs. It can also move consequential energy decisions behind the meter, where speed and corporate demand exert more influence than long-term public planning. Amazon’s commitment may not have changed, as the company says, but the world in which it must meet that commitment has: AI demand is growing faster than clean power and transmission can be delivered. GW Ranch shows what fills the gap when compute cannot wait.