Key Takeaways
Amazon is backing a gas plant next to a planned data center in Pecos County, Texas. Texas permitted the site to release up to 33 million tons of CO2 a year, which The New York Times reported is more than any other power plant in the country.
That 33 million ton figure is a permit ceiling, not a forecast. But I ran the arithmetic against the plant's reported 7.65 gigawatts, and the ceiling only balances if the turbines run hard at peaker level efficiency. At efficient combined cycle rates the permit is above what the plant could physically emit running flat out all year.
The Verge reported the plant would not connect to the Texas grid, at least initially. That is the part worth your attention. It is a scheduling decision before it is an emissions one.
Berkeley Lab puts the median wait from interconnection request to commercial operation at over five years for projects that came online in 2025, and only 13% of the capacity that applied between 2000 and 2020 ever reached operation. Amazon is not waiting in that line.
If you run agents, the practical read is that capacity is being priced by how fast it can be built, not by what it costs to run. The waste inside your own agent loops is the one variable you actually control.
A story broke on Saturday that got covered as a climate story, and I think that framing buried the more useful fact underneath it.
Amazon is building a data center campus in Pecos County, Texas. To power it, the company is backing a new gas burning plant on the same site. Texas granted that plant a permit to release up to 33 million tons of carbon dioxide a year. The New York Times reported that no other power plant in the United States is permitted to emit more, and both TechCrunch and The Verge picked the story up on August 8.
The headline number is real and it is worth arguing about. But I want to pull on a detail that appeared in exactly one of the two writeups, got no follow up in either, and explains the whole thing better than the emissions number does. If you track this beat, it belongs in the same file as the other infrastructure stories worth watching rather than in the climate pile.
The Verge carried the detail TechCrunch did not: the plant would not be connected to the Texas grid, at least initially.
What did Amazon actually get permission to build in Pecos County?
A gas plant sited next to a planned Amazon data center campus in West Texas, permitted to emit up to 33 million tons of CO2 per year. Amazon has confirmed it bought the site and plans to buy power from the plant. The reporting traces back to The New York Times, and both outlets that covered it on August 8 are working from that.
The Verge added specifics that TechCrunch's brief did not carry. Citing Cleanview, which tracks data centers and the power projects attached to them, the plant is called GW Ranch. It would run 35 natural gas turbines delivering 7.65 gigawatts, and that power would go primarily to the data center rather than onto the Texas grid. At least initially, the plant would not be connected to the grid at all.
Amazon gave The Verge a longer statement on August 9. The company said the campus "is powered by new on site generation that won't raise electricity costs for Texas families and is designed to transition to grid connected service as interconnection timelines allow." It also said it would use water unsuitable for drinking or irrigation, is exploring solar and battery storage on the site, and expects the campus to create thousands of jobs.
Read that middle clause again, because Amazon just told you the actual constraint in its own words. The plant transitions to the grid as interconnection timelines allow. The grid is not refusing Amazon. The grid is slow.
Why does the 33 million ton number not mean what the headline says?
Because it is a permit ceiling, not a projection. A permit sets the legal maximum a facility may emit, and plants routinely emit well under theirs. The Verge said as much. What neither piece did was check whether that ceiling is even reachable, and the answer turns out to be interesting.
I ran it against the reported nameplate. At 7.65 gigawatts running every hour of the year, the plant would generate about 67 million megawatt hours. To emit 33 million tons across that output, it would need to average roughly 0.49 tons of CO2 per megawatt hour. That number sits above what a modern combined cycle gas plant produces and down near simple cycle territory.
| Assumed turbine efficiency | CO2 per MWh | Capacity factor needed to reach 33 Mt | Physically possible? |
|---|---|---|---|
| Combined cycle, efficient | 0.36 t | 137% | No |
| Combined cycle, typical | 0.40 t | 123% | No |
| Simple cycle turbines | 0.55 t | 90% | Yes, running hard |
Those are standard intensity ranges, not GW Ranch's actual specifications, which are not public. Treat the table as a sanity check rather than a measurement. But the shape of it holds: if this plant runs as efficient combined cycle, it cannot reach its own permit ceiling even at 100% uptime. The ceiling is only reachable if the turbines are less efficient and run close to flat out.
So the permit is doing one of two things. Either it carries a lot of deliberate headroom, which is normal and boring. Or the plant genuinely leans toward simple cycle turbines running as baseload. And here is why the second reading matters: simple cycle turbines are what you install when you need power soon. They are faster to build and faster to start than combined cycle. Both readings point at the same motive, which is speed.
There is one more wrinkle that cuts against the usual "permits are overstated" intuition. Peaker plants sit idle most of the year, which is why they underrun their permits so badly. A data center does not sit idle. It is a baseload customer running 24 hours a day. The gap between permit and reality is normally large because demand is lumpy. Here the demand is flat and constant, so expect that gap to be narrower than the reflex suggests.
TechCrunch ran it as a five paragraph In Brief. It is the only one of the two carrying Amazon's own emissions number, up 16% last year.
What is AI data center energy actually constrained by?
Not fuel, and not money. It is constrained by the interconnection queue, which is the study and approval process a power project goes through before it is allowed to connect to the transmission grid. Berkeley Lab tracks this, and the numbers explain Amazon's decision better than anything in the emissions story.
Lawrence Berkeley National Laboratory's Queued Up series covers seven grid operators plus 50 non ISO utilities, roughly 98% of installed US generating capacity. Its 2026 edition, published in June with data through the end of 2025, reports:
More than 2,060 gigawatts of generation and storage were actively seeking a grid connection at the end of 2025, across about 8,200 projects.
For regions with available data, the median time from interconnection request to commercial operation was over five years for projects that came online in 2025.
Of the capacity that applied between 2000 and 2020, only 13% had reached commercial operation by the end of 2025. Fully 75% was withdrawn.
Active gas capacity in the queues jumped to 253 gigawatts in 2025, up 86% in a single year, while solar fell 19%, storage fell 16% and wind fell 19%.
Put those together. If you join the queue today, you are looking at a median five year wait and roughly a one in eight chance of ever getting built. Meanwhile the AI capacity you are building is being planned on an 18 month cycle. The two clocks are not compatible, and no amount of capital fixes it, because the constraint is study throughput and transmission upgrades rather than money.
Berkeley Lab's queue data through 2025. The whole queue shrank 10% last year. Gas went up anyway.
That 86% jump in queued gas is the same story as GW Ranch, just visible at national scale. Gas is winning right now not because it is cheap or clean but because it can be delivered on a schedule that matches the buildout.
Why would a company put a power plant on its own site instead of waiting?
Because paying for a plant on your own site takes less time than joining a queue to reach one somewhere else. That is the entire logic, and once you see it as a scheduling decision the rest of the story reorganizes itself around that.
A plant that serves one customer on the same site skips the part that takes five years. There is no transmission impact study, no cost allocation fight over network upgrades, no position in a queue behind 8,200 other projects. You are not connecting to anything. Amazon's own statement says the site is "designed to transition to grid connected service as interconnection timelines allow," which is a polite way of saying the company will join the grid later, when the paperwork catches up, and is not prepared to wait for it now.
This pattern is spreading. The Verge noted Meta and Google have both moved toward building their own generation, and it has separately reported that Meta faces a Democratic led probe over plans to power a giant data center with gas. The queue data says the same thing from the other direction. When the shared resource gets congested, the largest players stop sharing it.
I want to be careful not to launder that into approval. Building private fossil generation to dodge a slow public process is a real cost pushed onto people who did not choose it, and the Pecos County air does not care that the permit was legal. But if you are trying to predict what happens next, "companies are impatient with a five year queue" forecasts far better than "companies do not care about emissions."
What does this do to Amazon's climate pledge?
It strains it badly. Amazon co founded The Climate Pledge in 2019, committing to net zero carbon by 2040, and the company's emissions have gone the wrong way for several years running. TechCrunch reported Amazon disclosed a 16% rise last year, attributing the trend to AI demand.
Amazon's response to the story was the same in both outlets. A company spokesperson, Margaret Callahan, told The New York Times that "the world looks different now than when we co founded the climate pledge," while adding that "our commitment hasn't changed."
Both of those things can be true at once and that is what makes the statement uncomfortable rather than dishonest. The commitment is unchanged. The arithmetic under it changed completely, and a 2040 target set in 2019 did not price in a demand curve nobody forecast. What Amazon has not done is restate the target to match the new arithmetic, and until it does, the pledge and the permit are just going to keep pointing in opposite directions.
Amazon's sustainability site two days after the Pecos County permit was reported.
What should you do about this if you run AI agents?
Stop treating your token bill as the whole cost and start treating scheduled capacity as the scarce input. The practical version of that is unglamorous: measure the waste inside your own agent loops, because it is the only part of this chain you control, and in my experience it is larger than teams expect.
Here is the connection people miss. When capacity is priced by how fast it can be built rather than what it costs to run, the cost of compute stops falling smoothly with model prices. You can see this already. Opus 5 halved frontier pricing and most teams' agent bills barely moved, because per token price was never what was driving their spend. Volume was. And the Army's unlimited token deal ran dry in weeks for the same reason.
I've shipped 126 production systems, and the single most common source of spend is not the model choice. It is loops that reread context they already have. An agent that resends a 40,000 token document on every step of a 12 step plan is paying for that document 12 times. Retries on tool failures compound it. A planner that starts over from scratch after every observation instead of amending its plan doubles the whole run.
What I tell clients to do first, in order:
Instrument per run token counts before optimizing anything. Most teams cannot say what a single agent run costs. You cannot cut what you have not measured, and the answer is frequently a surprise.
Cache the stable context. Prompt caching buys the biggest saving for the least work, and it takes an afternoon.
Cap retries and make failures cheap. An uncapped retry loop against a flaky tool is how a $3 run becomes a $200 one overnight.
Route by difficulty. Most steps in most agent runs do not need the frontier model. Picking per step rather than per project is where the durable savings are. If you are already on AWS, Bedrock's managed agents make the routing decision easier to instrument than rolling it yourself.
A client of mine got the largest single drop from the first two alone, without changing models or touching a prompt. None of it required a view on Texas air permits. It just required knowing where the tokens went. Rippling built an internal console for exactly this after finding it was on track to spend 40% of its engineering payroll on tokens.
The second invoice for AI is physical, and it is being written right now in Pecos County and a few hundred places like it. You do not get a vote on that one. You do get a vote on how much of your own compute is doing useful work.
Frequently asked questions
Is the Amazon plant definitely going to be the biggest polluter in the US?
No. It is permitted to be. A permit sets a legal maximum, and plants generally emit less than they are allowed. The reporting from The New York Times says no other US power plant holds a higher permit, which is a claim about paperwork rather than about measured emissions. Nothing has been built or measured yet.
Why does it matter that the plant would not connect to the grid?
Because that is what makes it fast. A plant that serves one on site customer skips the transmission studies and cost allocation process that Berkeley Lab measures at a median of over five years. Going around the grid is how you get power on an AI buildout schedule instead of a utility schedule.
Does this raise electricity prices for people living nearby?
Amazon says it does not, on the grounds that the generation is new, on site and not drawing from the grid. That argument is coherent while the plant stays disconnected. It gets more complicated when the site connects to grid service later, which Amazon has said it intends to do.
Will AI data center energy demand keep growing at this rate?
The queue data suggests planners think so. Active gas capacity seeking interconnection rose 86% in 2025 while solar, wind and storage all fell, and developers do not spend money on queue positions casually. That is a forecast about expected demand, not a measurement of it, and forecasts of this kind have been wrong before.
Does any of this change which model or framework I should use?
Not directly, and be suspicious of anyone who says it does. Energy constraints show up in capacity availability and price floors over years, not in your framework choice this quarter. The actionable layer is loop efficiency and spend instrumentation, which pays off regardless of what happens in Texas.
The short version
A permit ceiling made the headline. A five year queue made the decision. Amazon is backing 7.65 gigawatts that deliberately does not touch the grid because the grid takes longer to join than the AI cycle allows, and the emissions permit is downstream of that choice rather than the cause of it.
If you are running agents in production and want a structured read on where your own compute and process waste actually sits, the AI readiness assessment walks through it in about ten minutes. Or see the agent systems I build if you want the version where someone else does the measuring.
Citation Capsule: The Pecos County plant is permitted for up to 33 million tons of CO2 per year, per The Verge (Aug 8, 2026) citing Cleanview, and TechCrunch (Aug 8, 2026), both crediting New York Times reporting. The 35 turbines, the 7.65 GW nameplate and the plant being unconnected to the Texas grid at least initially come from The Verge alone. Interconnection figures, a median of over five years from request to operation and a 13% completion rate for 2000 to 2020 applicants, plus 253 GW of queued gas up 86% in 2025, come from Lawrence Berkeley National Laboratory, Queued Up 2026 Edition (June 2026). Capacity factor figures in the table are my own calculation from the reported 7.65 GW nameplate against standard gas turbine emission intensities.
Top comments (0)