Skip to content

Why AIEnergy started

AIEnergy started to correct the erroneous “Data centers are driving up electric rates” stories. AI was just an excuse for the companies involved to raise rates and use obsolete technologies. In very few places did AI raise overall demand more than 2 or 3% per year, less than renewables + batteries + demand response are able to serve. NYT & WSJ believed the hype.

Everyone in energy knows the tech is ready to handle the demand and actually lower prices, but the system is effed up. Good reporting in the energy press and from researchers makes that clear.

I learned from some very smart people and will write as long as I have something to contribute. I’m veteran tech reporter Dave Burstein (daveb@dslprime.com), and I love to hear from you. It’s a great service when you catch my mistakes.

The Off-Grid Gas Buildout: Who’s Building It, Who’s Breathing It, and Why Oversight Is Thinning

This is a Claude backgrounder. I’m still checking before I write my article. Dave

Zeke Hausfather’s Climate Brink analysis, summarized separately here, put a number on how much dirtier off-grid data-center gas is than the grid: 63 to 80% worse per megawatt-hour. That number is the headline. Three other things are happening alongside it worth tracking on their own: how much of the buildout is actually gas, who’s breathing the consequences, and how much easier federal regulators are making all of it.

How much of the buildout is actually doing this

Cleanview, a commercial data-center tracking firm, published the most detailed public inventory of the trend in February 2026 and has since expanded it: as of its most recent count, 59 behind-the-meter data center projects representing roughly 90 GW of planned self-generation capacity are in the pipeline — about 30% of all planned US data-center capacity. Roughly 75% of the identified equipment is natural gas. Cleanview’s own reporting cautions that only about 30% of that planned capacity has specific equipment identified yet, and that a meaningful share of announced projects are “phantom projects” — developers pursuing several parallel sites with the intention of building only one.

Specific, independently reported examples of what’s actually going in: Meta’s El Paso, Texas facility is drawing on more than 800 mobile mini-turbines; Crusoe’s Abilene, Texas “Stargate” campus uses aeroderivative turbines adapted from jet engines; Caterpillar reciprocating engines supply a West Virginia data center; Amazon is financing an off-grid gas plant in Pecos County, Texas large enough, by permit, to become the single largest emissions source in the country; and a 9.2 GW gas-fired project in Ohio, backed by Microsoft and SoftBank, is planned to connect to the grid as well as serve a data center directly.

A separate advocacy-group count — the Environmental Integrity Project, in a July 2026 report that prompted independent Reuters reporting on top of it — identified 74 gas plants proposed specifically to power data centers directly, estimating they would represent 143 GW of capacity and 662 million tons of greenhouse gas emissions annually. EIP described that figure as comparable to the total annual emissions of France or Australia; by our own check against each country’s official inventory, 662 million tons actually exceeds Australia’s most recent full-year total (about 435–467 Mt CO2e) by 40–50%, and exceeds France’s (about 272–390 Mt CO2e, depending on accounting scope) by 70–140% — so if anything, the comparison undersells the number’s scale. Reuters’ own review of permits and hearings corroborated the report’s broader claim that off-grid, “behind-the-meter” projects are winning fast approval partly because developers argue they’re exempt from the standard review process required for plants that connect to the public grid.

It’s not just a climate number — it’s a local air-quality one

A separate, Texas-specific modeling exercise released October 20 by the Houston Advanced Research Center (HARC) found that on-site gas engines emit roughly four times as much nitrogen oxide (NOx) as equivalent grid power, and that statewide NOx emissions could rise more than 136% by 2031 under a buildout scenario weighted toward on-site combustion — enough, researchers said, to erase two decades of the state’s air-quality progress and approach early-2000s pollution levels in some categories. The findings were reported first by the Texas Tribune, which discloses that HARC has been a financial supporter of the Tribune within the past five years.

In Virginia, a narrower but more granular study illustrates the same dynamic at a single-facility scale. EmPower Analytics Group, commissioned by the land-conservation nonprofit Piedmont Environmental Council, used EPA’s COBRA health-impact model to estimate that the permitted (maximum allowable, not necessarily actual) emissions from one Loudoun County data center’s eight gas turbines could cause $53 to $99 million a year in regional health damages, including an estimated 3.4 to 6.5 additional premature deaths annually. (A separate, actual Virginia Commonwealth University study — sometimes conflated with this one in secondary reporting, including a Stand.earth press release that misattributes the $53–99M figure to VCU — looked at a different question: diesel-backup-generator emissions trends across 138 Northern Virginia data centers, finding sharp increases, including a 196% rise in carbon monoxide, from 2018 to 2023.)

A regulatory tailwind, not just a market trend

None of this is happening in a vacuum. Over the past year, EPA under Administrator Lee Zeldin has moved on several fronts that specifically affect the equipment driving these emissions:

  • Methane rule rollback. EPA announced in October 2026 that it would propose rescinding the 2024 Biden-era methane rule for oil and gas production.
  • Weaker NOx limits for new gas turbines. A finalized rule revises nitrogen-oxide limits for new gas-burning turbines and, environmental groups say, explicitly allows certain temporary turbines of the kind used at data centers to emit more NOx than other sources would be allowed to. The Environmental Defense Fund and allied groups sued over the rule in March 2026.
  • Less public notice on permits. EPA has proposed eliminating the federal requirement that states publicize and take public comment on air-pollution permits for “minor source” facilities — the permitting category that covers most data-center generators, including, in Virginia alone, more than 10,000 permitted generators according to a Sierra Club investigation.
  • A new, not-yet-published proposal specifically on data-center turbines. The Environmental Defense Fund flagged on October 20, 2026 that EPA is preparing a further proposal specifically weakening pollution limits for the turbines that power data centers. No text has been published yet; this is a developing story to watch, not a confirmed rule.
  • EPA has separately determined that power plants built exclusively to serve a single data center, without ever connecting to the broader grid, fall outside the Clean Air Act’s acid rain program.

Zeldin has framed the deregulatory push around winning “the race against China to be the AI capital of the world” and has said EPA will not set nationwide environmental standards for data centers, leaving the issue to individual states — several of which (Texas among them) have separately moved to pause or audit new data-center grid connections even as EPA loosens federal backstops on the off-grid alternative.

What’s still contested

The scale claims here — how many gigawatts are really being built, how many tons that adds up to — rest on a commercial tracker’s paywalled report and an advocacy group’s count of proposed plants, both internally consistent but neither independently audited. And EPA’s most consequential pending move on this specific issue hasn’t been published yet.


Verification status — Fidelity: Partial/Secondhand throughout this piece specifically, since Cleanview’s report is paywalled, the HARC study is read via news coverage rather than the primary PDF, and most EPA rollback items rely on independent news and legal coverage rather than primary filings (the exception is EDF’s active lawsuit, a primary legal record). Reliability: the regulatory-rollback claims are Tier 1 for the fact that EPA took these actions, resting on EPA’s own announcements plus independent corroboration. The scale and harm estimates (EIP’s 662 Mt/yr, Cleanview’s ~90 GW, HARC’s 136% NOx projection, EmPower/PEC’s health-cost figure) are Tier 2–3, single-origin, and in two cases [F]-flagged for advocacy or commissioned funding. One misattribution was caught and corrected here: a widely circulated $53–99M Virginia health-cost figure is sometimes attributed to “a VCU study,” but actually comes from EmPower Analytics Group, commissioned by the Piedmont Environmental Council — a distinct, separate VCU study exists and found something different. EPA’s pending “coming soon” turbine proposal is unconfirmed and should be treated as developing. The underlying emissions-gap numbers (63–80% dirtier than grid) are not repeated here — see the companion sources piece on Zeke Hausfather’s analysis for those, verified separately at Tier 1.