Amazon's 7.65 GW gas plant: your cloud is going off-grid

Four days ago we wrote that your AI compute had become a permit. It got more specific. The New York Times reported on August 8 that Amazon is the customer behind a private power plant in Pecos County, Texas that is permitted to emit 33 million tons of CO2 a year — more than any power plant currently operating in the United States. The Amazon GW Ranch project is the clearest signal yet that hyperscalers have stopped waiting in the interconnection queue and started building their own grid.

What actually happened

The plant is GW Ranch, developed by Pacifico Energy. Per the developer's own project page, it sits on 8,000+ acres above the Permian Basin and holds a Texas Commission on Environmental Quality air permit for up to 7.65 GW of gas-fired generation across 35 turbines, plus 1.8 GW of battery storage and up to 750 MWac of solar. Energy Tech reports the TCEQ approval — announced January 28 — is among the largest air permits ever granted in the US.

Pacifico's published schedule: construction from Q1 2026, first power Q1 2027, 1 GW operational in 2028, 5+ GW by 2031. This is behind-the-meter power — a private grid, not an ERCOT interconnection.

For scale on the emissions ceiling: the Times' 33-million-ton figure is roughly double the annual output of Alabama's James H. Miller Jr. plant, one of the largest coal facilities in the country. Amazon's response, quoted by TechCrunch, is that the campus "will be powered by new on-site generation that won't raise electricity costs for Texas families." Amazon's own carbon emissions rose 16% last year, against a 2040 net-zero pledge.

Why an off-grid power plant matters for your business

You do not buy gigawatts. You buy an inference API and a per-seat SaaS bill, and both are priced against capacity somebody promised would exist. Three things change when that capacity moves behind the meter.

The delivery risk changes shape. A grid interconnection slips because a commission is slow. A private power plant slips because of turbine supply, gas offtake, construction labor, and — now — environmental litigation from anyone who reads a 33-million-ton permit. Different failure modes, same effect on your 2028 token price. Nothing here reaches your bill before 2028; treat vendor pricing that assumes it as a forecast, not a floor.

Carbon becomes your problem via procurement, not conscience. If you sell software into companies that report Scope 3 emissions, the carbon intensity of the region your workload runs in eventually lands in someone's vendor questionnaire. Knowing which region your data sits in, and being able to move it, is turning into a sales requirement.

The operator moves don't change, which is the point. Keep AI contracts short. Keep the model layer swappable so a regional constraint is a routing change. Price automations at today's token cost with real headroom. When infrastructure gets this capital-intensive and this political, the only durable hedge is portability you own.

Sources: TechCrunch, Pacifico Energy, Energy Tech.