Invest1 distinct publisher3 min readUpdated
BloombergNEF tracks 99 on-site gas plants totalling 126 gigawatts. Run at normal rates they would add about a fifth to US power sector emissions, with no utility or grid operator vote.
The Investor · Invest desk

Compiled by The InvestorSomething wrong?How this is made
BloombergNEF tracks 99 on-site gas plants totalling 126 gigawatts. Run at normal rates they would add about a fifth to US power sector emissions, with no utility or grid operator vote.
Follow any of these and your For You feed starts watching them — no settings page required.
Bloomberg News, working from BloombergNEF data, counted 99 proposed on-site gas plants for data centers that would emit roughly 318 million metric tons of carbon dioxide a year if run at industry-standard rates [1]. The US electric power industry emitted about 1,485 million tons last year, according to Energy Information Administration data, so one subsegment of data center infrastructure could lift power sector emissions by 20 percent, and by as much as a third if the plants run flat out [2].
The mechanism matters as much as the tonnage. These are behind-the-meter projects, which can be permitted and built without the approval of utilities or the independent system operators responsible for grid reliability [3]. That is the point: the buildout has already strained the grid, drawing reliability warnings and moratoriums on new approvals, and even approved facilities wait years to interconnect [4]. "There is immense, immense pressure on the whole sector to get power, and get it fast," said David Pomerantz of the Energy and Policy Institute, a utility watchdog that promotes renewables, adding that developers are "sort of agnostic if it is clean or dirty" [5].
The estimate rests on 126 gigawatts of planned on-site gas capacity, a utilization range of 60 percent to 100 percent, and the burn rate of a typical single-cycle generator [6]. At the 60 percent end that implies roughly 0.48 tons of CO2 per megawatt-hour [7], and running the same fleet around the clock would put the figure near 530 million tons, about 36 percent of last year's power sector total [8]. Single-cycle units are now among the most commonly planned, and they are dirtier than the combined-cycle plants developers actually want but cannot get because of a yearslong turbine backlog [9]. Supply constraints, not preference, are setting the emissions intensity.
Discount the pipeline. Bloomberg notes that the scramble to serve AI demand has produced phantom projects and long-shot pitches, and not all of these plants will be built [10]. The concentration is still striking: 22 states from Alaska to Georgia, with more than a third of the projects in Texas, where Governor Greg Abbott recently announced a pause in data center approvals [11]. Cleanview identified Amazon as the developer of an 8,000-acre site in Pecos County that will become among the largest single sources of carbon pollution in the US [12], and about 30 miles west Chevron is building Microsoft a gas plant for a 2,000-acre complex [13]. Those two sites alone could produce more than 10 gigawatts, enough to power New York City on a hot summer day [14], with combined emissions as high as 45 million tons of CO2 equivalent according to regulatory filings [15] - roughly 14 percent of the entire tracked total from two addresses [16], and slightly less than half the cumulative emissions of Washington state, where both companies are headquartered [17].
Both Amazon and Microsoft back clean energy projects and say they aim to zero out their emissions contribution, pledges made before the AI boom under pressure from employees and activists [18]. The pipeline also includes projects backed by OpenAI and Anthropic, self-styled AI data center specialists, and investor groups shopping for tenants [19].
Watch three things: how long the Texas pause holds and whether other states copy it, whether the turbine backlog clears enough to shift new builds to combined cycle, and how emissions from a Chevron-owned plant serving a Microsoft load get booked. The last one determines whether the disclosures stay intact while the carbon goes up anyway.
Ranked by verification strength, evidence, and original report placement.
Ninety-nine proposed data center gas plants tracked by BloombergNEF would emit about 318 million metric tons of carbon dioxide annually if run at industry-standard rates, according to a Bloomberg News analysis.
The entire US electric power industry emitted about 1,485 million metric tons of carbon last year according to EIA data, meaning the tracked plants could lift US power sector emissions by 20%, and as much as a third if the new plants run flat out.
Combined annual emissions of the two Texas plants may be as high as 45 million metric tons of carbon dioxide equivalent, according to regulatory filings.
The 45 million tons is slightly less than half the cumulative emissions of Washington state, where both Amazon and Microsoft are headquartered.
Both Amazon and Microsoft are big backers of clean energy projects and have said they aim to zero out their contribution to carbon emissions; those pledges were made before the artificial intelligence boom, when technology companies were under pressure from employees and outside activists to cut emissions.
So-called behind-the-meter projects can be permitted and built without the approval of utilities or the independent system operators charged with ensuring grid reliability.
Evidence-backed comparisons of source perspectives and observed adoption signals. Read the methodology
Which Builder, Operator, and Investor concerns the observed source mix emphasized—not a truth score.
Evidence, demonstrated adoption, hype gap, incentives, and confidence are assessed independently, each on its own current evidence. How these are measured.
Disclosed methodology, single-organization sourcing
The core numbers rest on a named dataset (BloombergNEF), a published baseline (EIA power sector emissions) and regulatory filings for the two Texas sites, and the article discloses its capacity base, utilization range and generator assumption. But the estimate, the underlying dataset and the analysis all come from one organization carried by one publisher, with no project-level permit or financing breakdown and no independent replication in the cluster.
Two anchor projects underway, pipeline mostly proposals
Concrete adoption is limited: one plant under construction by Chevron for Microsoft and one large Amazon site identified by Cleanview, both with regulatory filings. The remaining tracked capacity is proposals across 22 states that the article itself says will not all be built, and Texas - more than a third of projects - has just paused data center approvals.
Headline scenario runs ahead of committed buildout
The framing figures - 318 Mt, a 20% lift, a third at full output - assume the whole 126 GW pipeline is built and run at 60%-100% utilization, while only two projects are documented as advancing and the article concedes some proposals are phantom. The overstatement is modest rather than severe because the assumptions are disclosed, the upper bound is labelled as such, and the two named projects' 45 Mt from regulatory filings shows the mechanism is real.
Affiliated data provider plus disclosed advocacy voices
Bloomberg News analyses a dataset from BloombergNEF, which the article discloses is owned by Bloomberg LP - an in-house data relationship. The two named critical voices are a renewables-promoting utility watchdog director and a former Microsoft employee who organized sustainability advocacy, both disclosed. Company positions appear only as brief spokesperson statements, and the fossil and hyperscaler firms named have obvious commercial stakes in how this pipeline is characterized.
Credible core numbers, single-publisher cluster
Confidence is mid-range: the mechanism (behind-the-meter gas bypassing utility and ISO approval), the two anchor projects and the EIA baseline are well documented, but the fleet-wide emissions total is a scenario from one organization's dataset with no independent check in this cluster, and the Texas pause plus phantom-project caveat add real uncertainty about how much of the pipeline materializes.
invest
The NRSC just told AI firms their data centers are a losing issue1 distinct publisher
invest
The AI moat is now a balance sheet, so price the financing and not the model1 distinct publisher
build
SemiAnalysis to software teams: your token cost starts at the fab, not the price list1 distinct publisher
build
Grok 4.6 lands in Copilot two days after launch, and the model picker becomes a procurement problem1 distinct publisher
Distinct publishers with included, body-backed reporting in this cluster.
1 article · August 18, 2026