For years the hyperscalers told a tidy story about AI’s energy appetite: yes, the models are hungry, but the buildout would be matched by renewables, PPAs, and eventually small modular reactors. Reporting from The New York Times, picked up by The Verge and Tom’s Hardware, punctures that story in the bluntest way possible. Amazon is backing a custom-built natural gas plant in Pecos County, West Texas — roughly 35 turbines, about 7.65GW of capacity — permitted to emit up to 33 million tons of greenhouse gases annually. That figure would put it among the largest single emitters in the country.
The number that matters isn’t 7.65GW
It’s 33 million tons. Capacity is a ceiling; emissions permits are a statement of intent about how hard you plan to run. A plant sized like a small national grid, built to serve one customer’s compute, is not a peaker sitting idle waiting for a heat wave. Data centers want flat, uninterrupted baseload. That’s precisely the demand profile gas serves well and intermittent renewables serve badly without storage nobody has built at this scale.
That’s the real admission here. Amazon didn’t abandon its climate commitments in a memo; it ran into a physics-and-schedule problem. Grid interconnection queues stretch for years. Nuclear is a decade out. Training clusters are being commissioned on an 18-month cadence. When the timeline for compute is that much shorter than the timeline for clean firm power, the gap gets filled with turbines — and building your own plant conveniently sidesteps the interconnection queue entirely.
Why West Texas, and why it won’t stay there
Pecos County is Permian Basin country: cheap stranded gas, permissive permitting, existing pipeline infrastructure, and a state grid operator eager for generation. It is close to an ideal jurisdiction for putting up a very large emitter very quickly. That combination is not unique to Texas, which is why treating this as one bad site misses the point. Every hyperscaler is racing the same clock, and the same arbitrage — build behind the meter, burn what’s local, apologize with offsets later — is available to all of them.
The accounting problem
Expect this to be reported as investment in a plant, not as Amazon’s emissions. That distinction is doing enormous work. If the plant exists to serve a specific data center, the carbon is functionally attributable to the compute it powers, whoever holds the operating license. Corporate net-zero targets built on annualized renewable matching were never designed to survive a dedicated fossil plant standing next to the load.
The honest framing is this: AI’s energy cost has stopped being an abstraction buried in a sustainability PDF and has become a physical object with a permit number. Anyone arguing that efficiency gains will absorb the demand curve now has a 35-turbine counterexample in West Texas.