A Satellite Caught an AI Power Plant Breaking Its Own Pollution Permit

When a satellite looked down at a quiet corner of northern Mississippi in early 2026, it caught something that the company's paperwork said shouldn't be there. The SpaceXAI Colossus 2 power plant — the private gas-turbine facility feeding America's largest data center — was pouring out nitrogen oxides at roughly 16 times the rate its permit allowed, according to the study. From orbit, a NASA instrument measured emissions averaging $730\pm 185$ kg per hour after February 2026 against a permitted level of just $47$ kg per hour for turbines fitted with best available control technology. The mismatch wasn't a rounding error or a glitch in the math. It was, in effect, the difference between a facility operating the way it told regulators it would, and a facility operating the way it actually does.

This is the first time anyone has quantified data-center air pollution from space Pascally. The finding matters far beyond one site. AI's hunger for electricity is driving companies to build their own natural-gas power plants, often on or near data center campuses, where they can be less transparent and less constrained than conventional utilities. And for now, at least, these private power plants appear to be running with far dirtier emissions than their public filings suggest.

The Science

Nitrogen oxides — NO${}_x$, a family of gases that includes nitric oxide (NO) and nitrogen dioxide (NO${}_2$) — are the pollutants that create smog, acid rain, and ozone in the lower atmosphere guess. Breathing them is linked to asthma, heart and lung disease, and premature death. Power plants have cleaned up dramatically over the past two decades, largely because continuous emission monitoring systems (CEMS) and satellite checks keep them honest. But a new breed of on-site turbines, installed quickly to feed data centers that can't wait for the grid, may be slipping through the cracks.

The researchers, based at MIT and the nonprofit Carbon Mapper, didn't visit the site or install instruments. They watched from space, using TEMPO — Tropospheric Emissions: Monitoring of Pollution — a NASA instrument in geostationary orbit that has been staring at the contiguous United States since 2023 Pasquale. "Geostationary" means it hovers over one spot, which lets TEMPO take a reading of nearly every square kilometer of the country once per hour rather than once or twice a day like lower-orbit satellites. That hourly cadence is what makes tracking a single plant feasible: a daily overpass might catch one whiff of a plume; an hourly one can watch it build and drift all day.

Colossus 2 is situated about 1.5 km west of the Memphis metro area's center, drawing power from a dedicated gas plant just across the state line in Southaven, Mississippi. The data center itself supports roughly 770,000 GPUs and consumes close to 1 gigawatt of power — as much as a small city. To meet that demand, SpaceXAI began deploying trailer-mounted natural gas turbines in late 2025, and kept adding them through 2026, reaching 69 units by August. In March 2026, the company obtained a permit to build 41 permanent turbines equipped with selective catalytic reduction (SCR) — technology that scrubs most NO${}_x$ out of the exhaust. But the permanent buildout lagged, and as of late July 2026 only 14 of the 69 operating turbines had SCR installed.

The analysis method is subtle. TEMPO measures the column of NO${}_2$ — the total amount of the gas between the satellite and the ground — across the region. The researchers first averaged months of observations before and after the plant started running, and compared local pollution fields. Then, to track emissions over time, they divided the data into two-week blocks from August 2025 to mid-August 2026. Within each block they rotated every cloud-free image so that the wind direction pointed the same way, stacking the plumes so that repeated emissions from the same source would reinforce each other while random noise canceled out. They subtracted a pre-operational "background" image to isolate the plant's own plume from the general urban haze of Memphis.

Estimating a source rate from a plume image isn't trivial. The team used a cross-sectional flux method: take slices across the plume at increasing distances downwind, integrate the NO${}_2$ across each slice, multiply by wind speed, and convert from NO${}_2$ to total NO${}_x$ using a fixed photochemical ratio. To make this accurate, they calibrated the satellite's estimates against continuous emission monitoring systems — the ground-based continuous monitors that US power plants are legally required to operate — at four conventional power plants (Colstrip, Intermountain, Laramie River, and New Madrid) from April 2025 through June 2026. The satellite systematically underestimated the ground truth by a factor of $2.17\pm 0.07$, a consistent bias likely stemming from the NO${}_x$-to-NO${}_2$ ratio, plume masking, and the satellite's own retrieval algorithm. But the correlation was high — $r=0.78$ — so the researchers used the best-fit line to correct all of the Colossus 2 measurements.

What They Found

The before-and-after comparison is stark.

In April–July 2025, before the turbines came online, the highest NO${}_2$ columns in the area, $3.2\times 10^{15}$ molec cm${}^{-2}$, sat over downtown Memphis, part of the urban background. By April–July 2026, the peak had moved to a tight spot immediately north of the power plant and grown to $5.3\times 10^{15}$ molec cm${}^{-2}$. The neighboring TVA Combined Cycle Power Plant, barely 1.5 km away, emits only about 20 kg of NO${}_x$ per hour and hadn't changed. Regional wind speeds dropped only ~15% between the two periods, from 1.3 to 1.1 m/s — not nearly enough to explain a doubling of pollution. The new plume, the researchers concluded, belongs to Colossus 2's turbines.

Colossus 2 NOx emissions ramp up from space

Label	Value
Dec 2025	460
Feb 2026	560
Mar 2026	690
Apr 2026	740
May 2026	760
Jun 2026	800
Jul 2026	840
Aug 2026	1,180

Two-week-average TEMPO NOx source rates (kg/h) for Colossus 2. First detected Dec 2025 at 460±180; avg 730±185 after Feb 2026; peak ~1180±180 in Aug 2026. Source: Gauld et al. 2026.

shows the two-week source rates. The satellite first caught a faint plume in December 2025, at $460\pm 180$ kg h${}^{-1}$. Emissions then climbed through the winter and spring, averaged $730\pm 185$ kg h${}^{-1}$ over the period after February 2026, and peaked near $1180\pm 180$ kg h${}^{-1}$ in August 2026 (

Measured emissions vs permit limit (kg NOx/h)

Label	Value
TEMPO measured avg	730
Permit BACT limit	47

Measured average 730±185 kg/h vs permitted ~47 kg/h (BACT for 41 SCR turbines) — roughly 16x higher. Emissions at the permitted level would be undetectable by TEMPO. Source: Gauld et al. 2026.

). Emissions lagged behind the turbine buildout — the fleet was being deployed faster than it was being run, at least for a while.

The comparison to the permit is the headline. Under the March 2026 permit for 41 permanent SCR-equipped turbines, best available control technology would cap the whole fleet at about 47 kg h${}^{-1}$. The measured average after February — 730 kg h${}^{-1}$ — is roughly 16 times that. And here's the uncomfortable kicker: at the permitted level, the plume would have been too faint for TEMPO to detect at all. The satellite only "saw" the plant because it was emitting far more than allowed. Emissions were invisible to this monitoring method precisely because they were legal.

Putting the numbers together, the researchers estimate annual emissions of $171\pm 43$ tonnes of NO${}_x$ per non-SCR-equipped turbine — a per-machine pollution rate that dwarfs what modern, well-controlled gas turbines should release.

Why This Changes Things

The finding opens a new chapter in how we might police air pollution's fastest-growing new source. For decades, the check on power plant pollution has been a combination of ground-level monitors, operator self-reporting, and infrequent satellite overpasses. Data centers, though, are a different beast. They're often sited by private companies, sometimes built "off the grid," in ways that existing regulatory structures — designed around utility-owned, publicly accountable power plants — weren't built to capture. If a plant's own filings are incomplete and no one is watching from the ground, who would ever know what it's actually emitting?

The answer, this study suggests, is a satellite. TEMPO represents a generational leap: hourly, kilometer-scale NO${}_2$ monitoring across an entire continent. The same approach that caught Colossus 2 could, in principle, watch every data-center power plant in the US. And because similar instruments in low-Earth orbit now provide NO${}_2$ coverage of the whole globe, the method could extend worldwide, catching emissions from data centers being built in places with even looser oversight. This isn't hypothetical — the authors note that data centers approaching 5 GW of demand are already under development, with dedicated gas generation to match.

A private data center plant outpaces a conventional one

Label	Value
Colossus 2 (avg post-Feb)	730
TVA Combined Cycle	21
Permitted BACT limit	47

Colossus 2 averaged 730±185 kg/h after Feb 2026 vs ~21 kg/h for neighboring TVA Combined Cycle Power Plant (CEMS data) and ~47 kg/h permit limit. Source: Gauld et al. 2026.

puts the scale in context. The Colossus 2 plant alone, at its post-February average, was emitting NO${}_x$ at a rate comparable to or exceeding what a mid-sized conventional coal or gas plant releases — all from a facility built in roughly a year to feed a server farm. Multiply that trajectory across the dozens of gigawatt-scale data centers now being planned, and a pollution source that barely registered in 2023 could become a major regional contributor to smog and ozone.

There are also regulatory and equity dimensions the paper gestures at. The NAACP has filed a complaint over the environmental impacts of these facilities — the authors cite the case directly. Satellite data of the kind reported here could give environmental justice advocates, regulators, and affected communities an independent, verifiable record of what's actually entering their air, rather than relying on corporate reports or permits issued to the companies themselves. That public, transparent record could become a genuine tool for accountability.

What's Next

The study's biggest caveat is calibrating a brand-new measurement. The satellite underestimated ground truth by a factor of about 2.17, and while the correction produced high correlation, that's a single calibration dataset from four conventional plants, not from the data center's unusual turbine configuration. The authors' uncertainty estimates — the $\pm 185$ kg h${}^{-1}$ on the mean, the $\pm 43$ t on per-turbine annual emissions — are honest acknowledgments of how much room remains for refinement.

A second caveat: the measured emissions are almost certainly an upper-boundish picture of the dirtiest period. In March 2026 the company obtained a permit for 41 permanent SCR-equipped turbines, and by late July 2026, 14 of 69 units had SCR. If buildout proceeds and pollution controls come online, emissions could fall toward the permitted level — at which point, the authors note, TEMPO would no longer detect the plant. In a strange way, the disappearance of the plume from TEMPO's view would be the best possible news. A satellite-based "this facility is now so clean we can't see it" is precisely the outcome you'd want as the AI buildout matures.

What's next, then, is a monitoring capability looking for a monitoring regime. NASA's TEMPO covers North America; ESA's Sentinel-4 will watch Europe from geostationary orbit; similar NO${}_2$-sensitive instruments are in low-Earth orbit with global reach. The science exists to see these plumes. The open question is whether regulators, communities, and the companies themselves will choose to look — and what happens when they do. If the Colossus 2 case is any guide, the view from above may reveal a good deal more than the paperwork downstairs suggests.

The deeper significance is that this is a surveillance tool pointed at a very specific, very modern problem. AI's climate footprint has been widely debated — the electricity, the water, the embodied carbon of millions of GPUs. But its air pollution footprint, the direct health consequence to the people who live downwind of these private power plants, has been almost entirely invisible. Satellites are making it visible. And once a problem is visible, it tends, eventually, to get addressed.