Originally published on The AI Prism
The Flip Nobody Planned
For decades, the gas-power buildout chart had one shape: China up, everyone else behind. Not anymore. Global Energy Monitor’s new analysis finds the US is now building twice as much gas-fired capacity as China — more than any other country on Earth (The Guardian).
The driver isn’t a manufacturing renaissance or a cold snap. It’s AI. Roughly half of the new capacity is tied directly to the datacenters that power the models we train, fine-tune, and serve every day (Global Energy Monitor). GEM now counts 189 GW of US gas capacity across the announced, pre-construction, and construction phases that is explicitly intended to meet datacenter demand (GEM).
Here at The AI Prism, we’ve been watching AI’s electricity appetite rewrite infrastructure economics for a year now. This is the story of how a software boom became a steel-and-turbine boom — and who ends up paying for it.
The US gas buildout is the most physical artifact of the AI boom you can point to. Understanding what’s getting built, whether it will ever run, and what it costs is now core to understanding AI itself.
China Out-Built the US for Decades. In Six Months, That Flipped.
For years, China added gas power faster than the US, period. In the first half of 2026, that reversed: under-construction gas projects in the US jumped 76%, reaching 52 GW, while China’s under-construction fleet sits at 24 GW (GEM).
“Six months ago, China had more gas plants under construction but that has now flipped,” says Jenny Martos, project manager at GEM (The Guardian).
The full pipeline ballooned even harder. Since January, gas capacity in development in the US has grown 50% — from 252 GW to 378 GW — and now accounts for one-third of the global total (GEM).
Add announced and pre-construction projects, and the US is building nearly three times as much as China (The Guardian). The US now accounts for nearly a quarter of all global gas capacity in development, with China, Vietnam, Iraq, and Brazil trailing (The Guardian).
To be fair to China: it isn’t standing still. It installed 22.4 GW of gas last year, its most ever in a single year (The Guardian). The difference is direction of travel — and 2026 US additions are now set to surpass the 100 GW annual record set back in 2002 (The Guardian).
Notice also where the plants sit. A year ago, GEM estimated that a third of the 252 GW then in development would be located on-site at datacenters (The Guardian). The model has shifted from “the grid will provide” to “the server farm brings its own power plant.” The historical order of things didn’t just bend. It inverted.
Texas Is the Epicenter, and the Data Center Is the Customer
Texas accounts for nearly one-third of the entire US pipeline: 122 GW of gas-fired capacity in development, up 51% in six months — a 41.4 GW jump larger than any other country’s entire buildout (GEM).
Of that, 77 GW — roughly two-thirds — is planned to directly power datacenters (GEM). Texas already led every state last year with 57.9 GW of new gas under way, ahead of Louisiana and Pennsylvania (The Guardian).
There is a reason the boom concentrates in Texas. The state has its own grid, fast permitting, and an electricity market that pays builders to show up. It is also the state where datacenter developers and gas developers have figured out how to sign contracts with each other.
Look at the customer of record and the pattern is unmistakable. The 21st-century gas plant isn’t being built for a factory or a subdivision. It’s being built for a server farm that hasn’t finished signing its lease.
AI Broke the Turbine Supply Chain
This boom collided with the physical world in a very specific way. Gas turbines are the most critical and expensive component of a gas plant, and the three largest manufacturers are now reporting rising order backlogs and multi-year lead times (GEM).
The datacenter stampede from Google, OpenAI, and Amazon has left developers waiting — and some have stopped waiting. Elon Musk’s xAI switched to smaller, less efficient turbines that emit more per megawatt (The Guardian).
GEM’s data shows developers increasingly skipping turbines entirely. Engine capacity in development more than doubled in six months, from 31 GW to 67 GW, and engine capacity tied to datacenters more than tripled, to 45 GW — nearly a quarter of all in-development gas for data centers (GEM).
Engines and simple-cycle turbines now make up nearly half of the generating technology behind data-center gas proposals, versus just 17% for projects not tied to datacenters (GEM).
Here is why that detail matters. Reciprocating engines and simple-cycle turbines are cheaper and faster to deploy than combined-cycle plants, but they are typically less efficient and carry higher emissions per unit of electricity generated (GEM). They were built for peak-hours duty, not for running a datacenter around the clock.
The scramble for speed is writing higher emissions and higher fuel costs directly into the design — before a single gigawatt of AI demand is confirmed.
The $647 Billion Question: Will Any of It Get Built?
If every project in the pipeline is completed, the US gas fleet grows by roughly two-thirds at a capital cost of more than $647 billion (GEM).
That “if” is doing heavy lifting. More than three-quarters of the global gas pipeline is still in early-stage development, and roughly 45 GW of announced and pre-construction capacity had its planned start year pushed back in the first half of 2026 alone (GEM).
The signs of a proposal economy are everywhere: two-thirds of global in-development gas — and more than half of data-center-tied projects — has no named turbine manufacturer, and nearly a quarter of data-center projects have no named start year (GEM).
“It is nearly impossible nowadays to guess what is a pie in the sky proposal, and what has a real chance of getting built,” Martos says. “The projects that eventually clear those hurdles are paying top dollar for turbines, locking in emissions, and pushing up electricity prices” (GEM).
We asked what survives when the AI bubble bursts, and the same logic applies to power: announced capacity is cheap, built capacity is real, and the gap between them is where the risk lives. Projects that stall don’t just fail quietly — they strand land, contracts, and investor capital. That uncertainty cuts both ways: for the climate math, and for the companies paying top dollar for turbines today.
Watch the same pattern that defined the GPU boom: hyperscalers announce capacity as a competitive signal, then the construction timeline does the talking. In energy, the lag is longer — a combined-cycle plant takes years to permit and build even when turbines are available. The pipeline you see today is a bet on demand forecasts from 2024, not a response to demand that has actually arrived.
The Emissions Math Is Ugly
Using gas rather than renewables to feed this datacenter glut could raise US power-sector emissions by as much as 20%, according to one estimate (The Guardian). In an economy that has spent two decades flattening its power emissions, that is a reversal, not a blip.
GEM’s January analysis put the lifetime cost in perspective. US gas projects in development would, if all completed, emit 12.1 billion tonnes of CO2 over their lifetimes — double the US’s entire current annual emissions from all sources. Worldwide, the planned gas boom totals 53.2 billion tonnes (The Guardian).
“Building all of this gas for AI locks in decades of pollution,” Martos says (The Guardian).
The uncomfortable part is that these are lifetime numbers. A gas plant ordered in 2026 is still likely to be running in 2056, well past every climate deadline on the books. The AI models these plants serve may be obsolete in five years; the turbines won’t be (The Guardian).
Renewables Were the Available Alternative
None of this was inevitable. GEM’s own analysis argues the demand “could be solved with flexible, clean power” (The Guardian). Gas is being chosen, not forced.
The comparison country is instructive. China — the world’s largest emitter, and the one the Trump administration points at — is adopting clean energy rapidly even as it builds gas (The Guardian). The International Energy Agency now forecasts US spending on coal- and gas-fired plants will outstrip China’s for the first time in decades (The Guardian).
The president’s framing — “their air is dirty, and it drifts over to us” — describes a China that is, on this measure, decarbonizing faster than the US (The Guardian). GEM put the fork in the road more bluntly in January: “As the AI bubble inflates, the US must decide whether it will double down on a fossil future while the rest of the world pivots to renewables” (The Guardian).
Gas plants do have a real role: they are dispatchable, and they can firm up intermittent wind and solar. But this buildout isn’t a reliability hedge. It’s a datacenter-driven sprint, and the turbines are being ordered before the demand curve has finished inflating.
Communities Are Saying No — and Politics Is Catching Up
The backlash is measurable. A Heatmap poll found three-quarters of Americans don’t want to live next to a datacenter — a huge jump in opposition within a year (The Guardian).
Add water to the fight: two-thirds of more than 800 planned datacenters are located in drought-stricken areas (The Guardian). A gas plant can be sited in weeks; a water supply cannot be conjured at all.
New York in July became the first state to enact a temporary ban on new hyperscale datacenter permitting and construction, and dozens of cities and counties have imposed their own restrictions (The Guardian).
The administration has doubled down. Trump has promised to do “whatever it takes” for US AI leadership and to sweep away “foolish rules” that slow the buildout; this month he said “data centers could be bigger than oil” and urged governors to cut taxes to attract them (The Guardian; The Guardian). Environmental reviews have been eliminated to speed construction (The Guardian). With the midterms in November, the politics of datacenter sprawl could bite — and voters in drought states are the ones holding the pencil.
The Cost Lands on Your Bill — and on AI’s
“It is also locking in dependence on a volatile fuel cost, which will get passed down to rate payers,” Martos warns (The Guardian).
The economics compound. Developers are paying top dollar for scarce turbines, less efficient machines burn more gas per megawatt, and domestic gas prices are forecast to surge again next year after a static 2026 (GEM; The Guardian). Every link in that chain is a cost that eventually shows up on a bill.
For the AI industry, electricity is the input nobody can optimize away. Every inefficient turbine and every delayed plant is a cost that ultimately lands on anyone paying for inference — which is everyone building on top of AI models. The training-run economics everyone obsesses over matter less than the price of the electrons the model eats in production.
Because so much of this capacity is being built on-site at datacenters, hyperscalers are signing their own gas contracts and eating the fuel-price risk directly (The Guardian). That means the cost shows up twice: once in their margins, and again in the prices they charge for AI services.
The bet is that AI demand justifies all of it. The risk is that the grid is being rebuilt for a demand curve that hasn’t finished inflating — and that the ratepayers, not the shareholders, absorb the difference.
The Bottom Line
The US gas buildout is the most concrete artifact of the AI boom: twice China’s pace, half of it datacenter-driven, more than $647 billion of capacity that may or may not get built. The turbines, the bills, and the emissions are real. The demand that justifies them is the one thing still in question.
The US is building twice as much gas-fired capacity as China — for one reason: AI. What happens to all that steel and gas when the models stop scaling as fast as the buildout?
References
• AI boom is driving a surprise resurgence of U.S. gas-fired power — Seattle Times, via Hacker News
The post The US Is Building Twice as Much Gas as China. AI Did That. appeared first on The AI Prism.
Cross-posted from theaiprism.com — Cutting Through the AI Noise 🧊
Top comments (0)