Amazon confirmed it acquired the GW Ranch site in Pecos County, Texas, and will buy power from a 7.65-gigawatt natural-gas plant built to serve a dedicated AI data center campus. The facility holds a Texas permit allowing up to 33 million tons of carbon dioxide a year, more than any operating U.S. power plant currently emits.
The plant starts entirely off the ERCOT grid. It sits inside a much larger, quieter shift: developers have already announced roughly 90 gigawatts of similar behind-the-meter gas capacity for data centers, most of it since early 2025.
A 7.65-Gigawatt Private Power Island in Pecos County
Pacifico Energy is developing the project on more than 8,000 acres in the Permian Basin. Amazon will take the power for its on-site AI campus. The air permit covers 35 gas turbines. First power is targeted for the first quarter of 2027, with the buildout able to reach more than 5 GW by 2031.
Amazon’s statement stressed the design “won’t raise electricity costs for Texas families” and that the company still intends to meet its Climate Pledge net-zero target by 2040. Spokeswoman Margaret Callahan put it directly:
The world looks different now than when we co-founded the climate pledge. Still, our commitment hasn’t changed.
Callahan’s comment followed years of rising company emissions. Amazon’s 2025 sustainability report showed 80.9 million metric tons of CO2 equivalent, a 16 percent jump driven largely by data-center energy use.
| Metric | GW Ranch (permitted) | Largest recent U.S. coal plant (Miller, 2023) |
|---|---|---|
| Capacity | Up to 7.65 GW | ~2.8 GW class |
| Annual CO2 (tons) | Up to 33 million | ~16.6 million |
| Grid connection | None initially (behind-the-meter) | Connected |
| Primary fuel | Natural gas (35 turbines) | Coal |
| Owner/offtaker | Pacifico Energy / Amazon offtake | Alabama Power |
Plants rarely run at their full permitted emissions ceiling, yet the authorized volume already exceeds every coal unit now operating. Simple-cycle or aeroderivative machines chosen for speed also run less efficiently than modern combined-cycle plants, raising the CO2 and NOx intensity per megawatt-hour.
Grid Queues Forced the Off-Meter Turn
Interconnection wait times for large loads have stretched into multiple years in many U.S. markets. Hyperscalers face AI demand that outruns both chip supply and grid upgrades. Amazon CEO Andy Jassy told investors in late July that AWS was growing at its fastest rate in 18 quarters and still could not meet all compute requests.
Behind-the-meter generation solves the timeline problem. Power can be online in months or a couple of years instead of waiting for transmission studies and upgrades. It also keeps the new load off local rate bases, a political advantage after communities protested bill spikes from earlier data-center waves.
The same logic is driving projects across the industry. Microsoft partnered with Chevron on a 2 GW off-grid gas campus only 30 miles west of GW Ranch. Amazon is also in talks on a separate 4.5 GW gas plant at Homer City, Pennsylvania. Meta, Oracle, OpenAI and Anthropic have signed similar leases or offtakes totaling more than 10 GW of private power.
Nearly 90 Gigawatts Already on the Drawing Board
Cleanview’s analysis of 59 behind-the-meter data centers totaling about 90 GW shows how fast the model scaled. Ninety-two percent of that capacity was announced since the start of 2025. Texas leads by a wide margin because of Permian gas, pipelines and a permitting culture that favors speed.
- Equipment reality: Heavy combined-cycle turbines have multi-year backlogs, so developers use mobile generators, aeroderivative units, reciprocating engines and refurbished industrial turbines.
- Revenue math: An AI data center can generate $10-12 million per megawatt per year. A one- or two-year head start can mean tens of billions in revenue, making lower efficiency acceptable.
- Status today: Only about 2 GW is operating (mostly xAI’s Memphis turbines). Roughly 3 GW is expected online by the end of 2026. Permitting and pipeline delays have already slowed several large projects.
- OEM share: Caterpillar leads tracked equipment, followed by others supplying the fast-start machines hyperscalers need.
Five states hold 83 percent of the proposed capacity. The pattern clusters around gas basins and friendly regulators. What began as a niche fix for xAI’s Colossus has become standard practice for every major cloud and AI lab.
How 33 Million Tons Stacks Against Coal Giants
The Texas Commission on Environmental Quality issued the largest TCEQ air permit for 7.65 GW in January 2026. That single authorization tops the annual greenhouse-gas output of Alabama’s James H. Miller Jr. plant, which led the nation at roughly 16.6 million tons in 2023 and has hovered near or above 20 million short tons in other recent years.
Amazon’s own footprint already rose 16 percent in 2025. Adding a plant permitted for 33 million tons, even if actual output runs lower, moves the company further from its 2040 net-zero line at the exact moment AI demand is accelerating. The company notes it has enabled 10 GW of carbon-free energy across 40 Texas projects for existing operations and plans 750 MW of solar plus 1.8 GW of battery storage at GW Ranch itself. Those additions help, yet the core capacity remains gas-fired and dispatchable.
Texas Families Stay Off the Hook, Emissions Do Not
Pacifico’s private-grid power generation campus design deliberately isolates the load. No new ERCOT transmission is required. Consumer bills in Texas do not absorb the capital or fuel costs. The company highlights 1,000-plus construction jobs, local tax revenue, use of non-potable brackish groundwater, and “CCS-ready” emissions controls.
Local economic upside is real in a rural county. The pollution and carbon accounting sit on Amazon’s books and the national inventory. Crowd reaction on X captured the trade-off cleanly: the AI race has become an energy race, and private islands let hyperscalers move at the speed of their revenue models rather than utility queues. Critics call the 2040 pledge decorative; supporters note that grid-connected alternatives would have raised Texas rates and delayed capacity by years.
Similar tensions appear in Amazon’s other infrastructure moves, from the Amazon and Google multicloud AI networking push to Amazon’s large-scale storage expansion in Egypt. Scale wins first; decarbonization pathways are added later.
Solar Batteries and CCS Sit Beside the Turbines
Pacifico’s own materials list the hybrid package: up to 7.65 GW of gas, 1.8 GW of battery storage and 750 MWac of solar. The storage and solar improve ramp rates and can clip peak emissions. The gas turbines still provide the firm capacity that AI training clusters demand 24 hours a day. The site is described as remote, flat and free of wetlands, chosen for construction speed and gas access.
- January 2026: TCEQ issues the 7.65 GW air permit.
- July 2026: Satellite imagery shows land clearing under way; Amazon files construction permits for three data-center buildings.
- August 2026: Amazon publicly confirms site acquisition and power offtake from Pacifico.
- Q1 2027 target: First power delivery.
- 2028-2031: Phased expansion toward 5-plus GW online.
Whether carbon capture is ever installed remains an open capital decision. The permit and the turbines come first.
The Capacity Amazon and Peers Cannot Wait For
AI data centers generate revenue that dwarfs the cost of even inefficient gas generation. Grid planners and renewable developers cannot match that urgency on current interconnection timelines. The result is a parallel private power system growing inside the United States, concentrated in gas-rich states and largely outside traditional utility oversight.
Amazon’s Pecos plant will be the largest single example if built to the permitted scale. It will not be the last. The same model is already spreading to Pennsylvania, Ohio, New Mexico and beyond. Emissions inventories will show the cumulative effect long before the 2040 pledges come due. The turbines are being ordered now.
