Amazon has disclosed that it is considering moving its self-generation system at GW Ranch, an AI campus of 8,000 acres (about 32 square kilometres) in Texas, to an arrangement connected to the grid that could also supply other customers. In a response to Data Center Knowledge, an Amazon spokesperson said the company is "actively considering" moving to front-of-the-meter grid service, and that its energy infrastructure could eventually have opportunities to contribute supply to other customers.
Where behind-the-meter generation serves a company’s own facilities, front-of-the-meter generation connects to the grid and can supply other customers. Amazon has not disclosed the timing of a transition, whether it has submitted a large-load interconnection request to ERCOT (Electric Reliability Council of Texas), or the grid capacity required by the campus.
Generation scale: 5,000MW and 35 units in TCEQ filing; 7.65GW in Pacifico disclosure
Pacifico Energy will provide GW Ranch’s generation, and Amazon has confirmed both acquisition of the site and power purchases from the company. A Texas Commission on Environmental Quality (TCEQ) preliminary determination describes a planned 5,000MW natural-gas plant comprising 35 simple-cycle turbines, designed to supply the AI data center on the site. It assumes no buying or selling of power with the local utility distribution network.
Pacifico says it plans initial output in the first quarter of 2027, 1GW in operation in 2028, and more than 5GW by 2031. Multiple US outlets, including the New York Times, The Verge, and Futurism, have highlighted the TCEQ document’s maximum permitted annual emissions of about 33 million tonnes of CO2 equivalent. That figure is a theoretical maximum if all 35 turbines run for all 8,760 hours of a year, not a forecast of actual annual emissions.
Why self-generation at this scale is unusual
Joshua Rhodes, who researches power markets and infrastructure at the University of Texas, says behind-the-meter generation at this scale remains unusual because it is difficult to execute. Many companies have considered it, he says, but abandoned the effort after discovering how difficult it is.
Turbine procurement is itself a constraint. Rhodes says turbines are currently scarce, and if 5GW of capacity goes to this behind-the-meter data center, it will not be available for other uses. This points to competition between AI-data-center generation procurement and other power projects. Pacifico says the private-grid approach is intended to avoid protracted conventional interconnection procedures and deliver power directly to hyperscale data centers.
What the transition entails
A shift from behind-the-meter self-consumption to front-of-the-meter grid supply is under consideration. Timing, ERCOT application status, and supply scale are undisclosed.
Mismatch in generation scale
TCEQ’s preliminary determination lists 5,000MW and 35 units. Pacifico cites up to 7.65GW plus 1.8GW of batteries and 750MW of solar. Their relationship is not explained.
Competition for turbines
Experts say securing scarce gas turbines at this scale could constrain supply for other generation projects.
Water-use plan
The initial phase is designed to use no cooling water. Later phases would mainly use brackish groundwater and are intended not to compete with drinking-water supply.
Relationship to ERCOT’s large-load review process
ERCOT operates "Batch Zero," a grid-feasibility review process for large-load applications of 75MW or more. It groups projects to assess reliability, allocate grid capacity, and identify needed transmission upgrades. ERCOT’s framework also includes Withdrawal-Limited Private Use Network (WLPUN), commonly called bring-your-own-generation, for large loads with co-located self-generation seeking grid connection. The unresolved issue for GW Ranch is how this framework would apply if Amazon moves from its initial private power system to grid connection while retaining some or all of its on-site generation. Rhodes says the point remains unclear, but should become clearer as ERCOT reviews Batch Zero and subsequent large-load batches.
Rhodes also does not expect hyperscalers to retain power generation as a permanent core business. He says hyperscalers probably do not want to enter the power business because margins are too low, and that it is currently a choice forced on some companies by necessity.
Implications for procurement and design
Amazon has not disclosed investment scale or the campus’s planned IT load, though it expects to create thousands of jobs. The case suggests that hyperscaler self-generation may be changing from a purely dedicated asset to one partly incorporated into the grid. A phased approach—securing generating capacity behind the meter first, then moving front of the meter as grid development progresses—may offer a reference for other AI data-center projects where transmission-upgrade lead times are a bottleneck. Tight turbine supply also affects order conditions for power-equipment suppliers. Whether further reports reveal a formal ERCOT application and the intended supply scale is the key point to watch.