Amazon confirmed it is financing a 7.65-gigawatt natural gas power plant in Pecos County, Texas (GW Ranch, developed by Pacifico Energy) to power a new off-grid AI data center campus. The facility would use 35 natural gas turbines and be permitted to emit 33 million tons of CO2 annually—more than any single power plant currently operating in the United States. Permits were filed in early August 2026 after satellite imagery connected three data center construction filings to the gas project.
This is Amazon's first major behind-the-meter (off-grid) power commitment, following similar announcements by Microsoft (2GW near Pecos, partnered with Chevron), Google, Meta, and xAI. Since early 2025, at least 59 behind-the-meter gas projects have been announced across the U.S., totaling ~90GW of planned capacity, with Texas hosting the largest share. The strategy reflects hyperscalers' move away from grid dependency due to long interconnection timelines and public utility infrastructure bottlenecks.
Amazon framed the decision as cost-neutral to Texas ratepayers: 'Our commitment hasn't changed' regarding its 2040 net-zero Climate Pledge, and the company said it is exploring solar and battery storage on-site, using non-potable groundwater rather than drinking water. Pacifico announced plans for 750MW solar and 1.8GW battery storage at the same site. However, 33M tons CO2/year authorization dwarfs these offsets substantially.
For architects: this signals a hard constraint shift. Hyperscalers can no longer assume grid capacity for training clusters—instead, they're financing dedicated generation. The emissions profile is worse than grid power in most U.S. regions (wind/solar heavy), but it guarantees uptime and CapEx predictability. Expect more 5–10GW gas projects tied to specific foundation model training timelines in frontier energy markets (Texas, Pennsylvania, Ohio), reducing architectural reliance on spot capacity but increasing cost-per-token at the inference edge.