Quantifying AI data center nitrogen oxide (NO x) emissions from space

arXiv:2608.22153 · physics.ao-ph · Submitted 2026-08-23 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Physics. The day in general physics: instruments, methods and results across the field.

Kai: I'm Kai.

Mira: And I'm Mira, here with Lev.

Kai: We're looking at the paper: "Quantifying AI data center nitrogen oxide (NO x) emissions from space".

Paper summary: Kai: So we're looking at this paper right now called "Quantifying AI data center nitrogen oxide (NO x) emissions from space." The main thrust is that because of the massive power demand from AI data centers, which are using gas turbines, there's this pollution issue that hasn't been measured with atmospheric observations before.

Mira: Exactly. They demonstrate how you can use space-based detection and quantification of nitrogen oxide emissions specifically from the SpaceXAI Colossus two power plant in Southaven, Mississippi > <ref:2608.22153#pg1>

Lev: It’s important to know that they use observations from the geostationary TEMPO satellite instrument > <ref:2608.22153#pg2>

Kai: And what they found was that after the plant started operating in late two thousand twenty-five they detected a strong increase in local mean NO two column concentrations > <ref:2608.22153#pg1>

Mira: They then used TEMPO to estimate two-week-average NOx source rates from August two thousand twenty-five to August two thousand twenty-six and they calibrated those estimates against continuous emission monitoring system data from US power plants > <ref:2608.22153#pg1>

Lev: This whole study hinges on comparing what the satellite sees with what the ground-based monitoring systems report > <ref:2608.22153#pg1>

Kai: And they show that TEMPO first detected emissions in December two thousand twenty-five but then the rate spiked, averaging eight hundred seventy plus or minus one hundred sixty-five kilograms per hour after February two thousand twenty-six > <ref:2608.22153#pg1>

Mira: That average is about nineteen times higher than what the March two thousand twenty-six permit expected for forty one turbines operating under best available control technology requirements which was about forty seven kilograms per hour > <ref:2608.22153#pg4>

Lev: So, the numbers show a huge gap between what they are allowed to emit and what is actually happening on the ground > <ref:2608.22153#pg4>

Kai: The methodology involves comparing oversampled TEMPO NO two column concentrations near the plant before and after operations began, which revealed this strong increase in local NO two levels > <ref:2608.22153#pg1>

Mira: Then they average those hourly TEMPO observations over time using wind rotation to generate images of what they call two-week-average NO two plumes from the facility > <ref:2608.22153#pg1>

Lev: And for estimating the actual emissions, they infer biweekly NOx emissions by using a cross sectional flux method calibrated against continuous emission monitoring systems data from seven US power plants > <ref:2608.22153#pg1>

Kai: They did calibrate those estimates against CEMS data, but there's a discrepancy where the TEMPO source rate estimates underestimate the CEMS by about a factor of one point seven eight plus or minus zero point one four on average > <ref:2608.22153#pg7>

Mira: That difference might come from errors in the nitrogen oxide to nitrogen dioxide ratio, plume masking, and or the TEMPO NO two column retrievals themselves > <ref:2608.22153#pg7>

Paper summary: Lev: The final source rate estimate is calculated by summing random error terms in quadrature using a formula that involves a correction factor 'a' which is one point seven eight plus or minus zero point one four on average > <ref:2608.22153#pg10>

Kai: So, looking at the magnitude, they found that emissions were first detected at five hundred thirty plus or minus two hundred twenty kilograms per hour in December two thousand twenty-five > <ref:2608.22153#pg8>

Mira: The emissions were markedly higher in February and later months, peaking around one thousand two hundred fifty plus or minus two thirty five kilograms per hour in August > <ref:2608.22153#pg10>

Lev: And the average after February two thousand twenty-six was eight hundred seventy plus or minus one hundred sixty-five kilograms per hour > <ref:2608.22153#pg10>

Kai: This average is about nineteen times higher than what was outlined in the permit for forty one permanent turbines under best available control technology requirements which was forty seven kilograms per hour > <ref:2608.22153#pg4>

Mira: They also inferred an emission intensity of zero point nine three plus or minus zero point eighteen grams per kilogram of kilowatt hour for non-selective catalytic reduction equipped turbines > <ref:2608.22153#pg10>

Lev: This suggests that the actual operational emissions are significantly higher than the permitted levels set by the Mississippi Department of Environmental Quality records > <ref:2608.22153#pg4>

Kai: The implications here is that satellites can detect and quantify NOx emissions from AI data centers powered by non-SCR equipped gas turbines, which allows for independent monitoring of air pollution and checking compliance with environmental regulations > <ref:2608.22153#pg10>

Mira: They also suggest that similar analysis could be done using NO two sensitive instruments that have global coverage in low Earth orbit > <ref:2608.22153#pg10>

Lev: And as AI data centers grow to multigigawatt scales, with facilities approaching five gigawatts already under development and about one hundred forty gigawatts of new turbine capacity announced in the US alone, dedicated gas power generation could become a much bigger source of air pollution > <ref:2608.22153#pg10>

Kai: Satellite observations provide this capability to monitor these emerging sources worldwide > <ref:2608.22153#pg10>

Mira: When we look at the expected emissions under best available control technology requirements, it's based on public records from the Mississippi Department of Environmental Quality > <ref:2608.22153#pg4>

Lev: That total fleet emission under those rules is forty six point nine kilograms per hour or four hundred sixteen kilograms per hour annualized > <ref:2608.22153#pg4>

Kai: So what the paper really highlights is that there's a significant gap between the permitted levels and these observed emissions from the Colossus two plant > <ref:2608.22153#pg4>

Paper summary: Mira: This paper, "Quantifying AI data center nitrogen oxide (NO x) emissions from space," shows that this is something we can independently monitor using satellite data > <ref:2608.22153#pg1>

Lev: And the method they use to compare their satellite measurements with the ground monitoring system is pretty rigorous, but they admit that there's some uncertainty in how much the CEMS data actually matches the satellite readings > <ref:2608.22153#pg7>

Kai: That uncertainty is real because of things like errors in how they calculate the nitrogen oxide to nitrogen dioxide ratio and plume masking > <ref:2608.22153#pg7>

Mira: And the paper also points out that for their two-week-average source rate retrievals, they discard estimates below eighty percentiles because of potential biases in the near field and far field measurements > <ref:2608.22153#pg10>

Lev: That means their final estimate is based on taking the mean of whatever observations are left after applying those filters, which is a specific decision they made about how to handle those data biases > <ref:2608.22153#pg9>

Kai: The limitations they state are that their source rate estimates might be biased low in the near field where ozone titration delays NO two formation, and in the far field due to NOx oxidation to nitrogen trioxide > <ref:2608.22153#pg10>

Mira: And they also note that their results are insensitive to the choice of percentile threshold if you vary it between seventy percentiles and ninety percentiles > <ref:2608.22153#pg10>

Lev: So, for someone building real hardware, this tells us that when you try to run these estimates on real systems, you have to account for the fact that the noise in your measurement record isn't just random error > <ref:2608.22153#pg10>

Kai: This research is a demonstration that satellites can detect and quantify NOx emissions from AI data centers powered by non-SCR equipped gas turbines, which enables independent monitoring of air pollution and compliance with environmental regulations > <ref:2608.22153#pg10>

Mira: It opens the door to using NO two sensitive instruments with global coverage in low Earth orbit for similar atmospheric measurements > <ref:2608.22153#pg10>

Lev: As AI data centers expand to multigigawatt scales, with facilities approaching five gigawatts already under development and about one hundred forty gigawatts of new turbine capacity announced in the US alone, dedicated gas power generation could become a much bigger source of air pollution > <ref:2608.22153#pg10>

Mira: That's what the paper "Quantifying AI data center nitrogen oxide (NO x) emissions from space" shows about the need for better quantification of this pollution > <ref:2608.22153#pg1>

Lev: It’s a reminder that as we build out these massive systems, gas power generation could become an increasingly important source of air pollution > <ref:2608.22153#pg10>

Conclusion: Kai: So we've been looking at how they used TEMPO from space to measure nitrogen oxide emissions from an AI data center power plant called Colossus two > <ref:2608.22153#pg1>

Mira: Right, it’s about taking something that was previously unquantified with atmospheric observations and using a satellite to get hard numbers on NOx pollution from those gas turbines > <ref:2608.22153#pg1>

Kai: The authors are pointing out the difference between what the permit allowed, which is based on control technology, and what they actually measured in August two thousand twenty-six > <ref:2608.22153#pg4>

Lev: It's a reminder that the expected emissions under best available control technology requirements is based on public records from the Mississippi Department of Environmental Quality, which yields a fleet total of forty-six point nine kilograms per hour or four hundred sixteen kilograms per hour annualized > <ref:2608.22153#pg4>

Mira: And they show that the actual observed emissions were about nineteen times higher than what that permit allowed for those turbines > <ref:2608.22153#pg4>

Kai: So what they're showing us here is that satellites can detect and quantify NOx emissions from AI data centers powered by non-SCR equipped gas turbines, enabling independent monitoring of air pollution and compliance with environmental regulations > <ref:2608.22153#pg10>

Lev: This paper also suggests that similar analysis could be done using NO two sensitive instruments with global coverage in low Earth orbit > <ref:2608.22153#pg10>

Mira: And as AI data centers expand to multigigawatt scales, with facilities approaching five gigawatts already under development and about one hundred forty gigawatts of new turbine capacity announced in the US alone, dedicated gas power generation could become an increasingly important source of air pollution > <ref:2608.22153#pg10>

Kai: Satellite observations provide this capability to monitor these emerging sources worldwide > <ref:2608.22153#pg10>

Lev: The core issue they highlight is that as we build out these massive systems, gas power generation could become a much bigger source of air pollution > <ref:2608.22153#pg10>

Mira: It really shows how much the environmental impact of AI infrastructure might be underestimated if you don't have this kind of independent, global atmospheric monitoring > <ref:2608.22153#pg1>

Massachusetts Institute of Technology · Carbon Mapper

physics.ao-ph

Submitted: 2026-08-23

Updated: 2026-11-18

The gist: AI data center power demand is spurring rapid deployment of onand near-site natural gas turbines, and the resulting nitrogen oxide (NOx) pollution is a growing concern but has not previously been…

Key concepts

TEMPO Satellite Instrument
TEMPO is a geostationary satellite instrument used to monitor NO2 column concentrations across large areas of the continent hourly. It was crucial for detecting the initial increase in NOx emissions near the power plant and for generating images of two-week-average plumes.
Continuous Emission Monitoring System (CEMS)
CEMS are ground-based systems used to continuously measure and report pollutant emissions from power plants. The study calibrated its satellite estimates against CEMS data from seven US power plants to improve the accuracy of the source rate calculations.
Source Rate Estimation
This involves calculating how much NOx is being emitted by a specific source, like a power plant. The researchers used a cross-sectional flux (CSF) method, calibrated with CEMS data, combined with satellite NO2 measurements and wind rotation to estimate the biweekly emissions of NOx.
Best Available Control Technology (BACT)
BACT refers to the highest level of pollution control technology required for a power plant. The study compared observed emissions against BACT limits set in permits, showing that the actual emissions were significantly higher than what was permitted.

Terminology

Summary

AI data center power demand is spurring rapid deployment of onand near-site natural gas turbines, and the resulting nitrogen oxide (NOx) pollution is a growing concern but has not previously been quantified with atmospheric observations. p. 0 Here we demonstrate space-based detection and quantification of NOx emissions from the SpaceXAI Colossus 2 power plant in Southaven, Mississippi. p. 0 Using observations from the geostationary TEMPO satellite instrument, we detect a strong increase in local mean NO2 column concentrations after the plant began operations in late 2025. p. 0 We then use TEMPO to estimate two-week-average NOx source rates from August 2025 to August 2026, calibrating against continuous emission monitoring system (CEMS) data from US power plants. p. 0

Detection and Initial Observation

TEMPO first detected NOx emissions in December 2025 at 530±220 kg h−1 <ref:2608.22153#pg8>. We find that emissions increased through August 2026, averaging 870±165 kg h−1 after February 2026, roughly 19±4 times higher than expected from the facility’s March 2026 permit for 41 turbines operating under best available control technology (BACT) requirements (∼47 kg h−1) <ref:2608. p. 022153#pg4>. Emissions at the expected level would be undetectable by our TEMPO analysis <ref:2608. p. 022153#pg4>.

Methodology for Quantification

The study employs satellite instruments like TEMPO, which provides hourly km-scale NO2 monitoring across continental domains. The process involves several steps to estimate source rates, including:

  1. Comparing oversampled TEMPO NO2 column concentrations near the power plant in the months before and after operations began, revealing a strong increase in local NO2 levels <ref:2608. p. 122153#pg5>.

  2. Averaging the hourly TEMPO observations over time with wind rotation to generate images of two-week-average NO2 plumes from the facility <ref:2608. p. 122153#pg9>.

  3. Inferring biweekly NOx emissions using a cross-sectional flux (CSF) method calibrated against continuous emission monitoring systems (CEMS, [6]) data from seven US power plants <ref:2608. p. 122153#pg10>.

Calibration and Source Rate Retrieval

The retrieval of source rates is calibrated against CEMS data, where the TEMPO source-rate estimates underestimate CEMS by a factor of 1.78±0.14 on average (1. p. 044–1.89 per site) <ref:2608. p. 222153#pg7>. Potential contributors to this discrepancy include errors in the NOx-to-NO2 ratio, plume masking, and/or the TEMPO NO2 column retrievals <ref:2608. p. 222153#pg7>. The final source rate estimate Q is given by a formula that sums random error terms in quadrature <ref:2608. p. 822153#pg10>.

Key Findings on Emissions Magnitude

The analysis shows that emissions were first detected at a rate of 530±220 kg NOx h−1 in December 2025 <ref:2608.22153#pg8>. The emissions were markedly higher in February and later months, peaking at ∼1250±235 kg h−1 in August and averaging 870±165 kg h−1 after February 2026 <ref:2608. p. 322153#pg10>. This average is ∼19±4 times higher than outlined in the permit for 41 permanent SCR-equipped turbines under best available control technology (BACT) requirements (47 kg h−1) <ref:2608. p. 322153#pg4>. We infer an emission intensity of 0. p. 393±0.18 g NOx kWh−1 for non-SCR-equipped turbines <ref:2608. p. 322153#pg10>.

Implications for Monitoring

Our study demonstrates that satellites can detect and quantify NOx emissions from AI data centers powered by non-SCR-equipped gas turbines, enabling independent monitoring of air pollution and compliance with environmental regulations <ref:2608.22153#pg10>. Similar analysis could be conducted using NO2-sensitive instruments with global coverage in low-Earth orbit <ref:2608. p. 3 p. 322153#pg10>. Satellite observations provide a capability to monitor these emerging sources worldwide <ref:2608. p. 3 p. 3 p. 322153#pg10>.

Expected Emissions Comparison

The expected NOx emissions of 47 kg h−1 under best available control technology (BACT) requirements is based on public records from the Mississippi Department of Environmental Quality (MDEQ) <ref:2608.22153#pg10>. Summed over all 41 turbines, this yields a fleet total of 46. p. 8 p. 89 kg h−1 (3-hour basis) or 41. p. 8 p. 86 kg h−1 (annualized) <ref:2608. p. 0 p. 022153#pg10>. The comparison highlights the significant discrepancy between permitted levels and observed emissions <ref:2608. p. 022153#pg4>.

Capacity and Future Scope

The figures for deployed turbine capacity are based on public records from the Mississippi Commission on Environmental Quality (MCEQ) <ref:2608.22153#pg11>. Total installed capacity according to MCEQ is 1770. p. 972 MW, with 1540 MW attributed to turbines not equipped with SCR <ref:2608. p. 9 p. 922153#pg10>. As AI data centers expand to multigigawatt scales, dedicated gas power generation could become an increasingly important source of air pollution <ref:2608. p. 322153#pg10>. Satellite observations provide a capability to monitor these emerging sources worldwide <ref:2608.22153#pg10>.

Extended Methods Detail

For wind-rotated averages, we define two-week intervals for each month from December 2024 through August 2026, between the 1st and 15th and between the 15th and 1st of the next month <ref:2608.22153#pg9>. We use wind-rotated TEMPO observations from before the emissions began to define a 2D background field, which we then subtract from post-operations wind-rotated scenes <ref:2608. p. 722153#pg9>. To create plume masks for the wind-rotated two-week-average retrieval fields, we apply a total-variance filter and then mask downwind enhancements that are at least 2–3 standard deviations above the background noise floor <ref:2608. p. 722153#pg9>. We require masks to begin within 10 km of the source location, discard mask regions smaller than 5 pixels, and remove regions outside the downwind section of the scene (±25° wedge from source location) <ref:2608. p. 7 p. 722153#pg9>.

Error Characterization

We characterize errors in our two-week-average source-rate retrievals using the results of the CEMS calibration <ref:2608.22153#pg10>. Error in the best-fit slope a is the standard deviation of per-site slope estimates, and random error in individual estimates is characterized as the standard deviation of fit residuals <ref:2608. p. 822153#pg10>. These two error terms are summed in quadrature, and the final source rate estimate Q is given by a formula that sums random error terms in quadrature <ref:2608. p. 822153#pg10>.

Source Rate Calculation

We multiply these line integrals by the average HRRR PBL wind speed across all TEMPO observations and a NOx-to-NO2 ratio of 1.38±0.10 [11], representing photochemical steady-state conditions, to produce a downwind profile of sourcerate estimates <ref:2608. p. 8 p. 822153#pg10>. These estimates are expected to be biased low in the near-field of the source, where ozone titration by the emitted NO delays NO2 formation, and in the far-field due to NOx oxidation to HNO3 <ref:2608. p. 8 p. 822153#pg10>. To account for this, we discard estimates below the 80th percentile and report the mean of the remaining estimates <ref:2608. p. 8 p. 822153#pg10>. Our results are insensitive to the choice of percentile threshold; varying it between the 70th and 90th percentiles produces a spread of. p. 8 p. 8

Final Source Rate Estimate

For our average after February 2026, we find σs = 165 kg h−1 across 11 samples <ref:2608. p. 822153#pg10>. This represents the standard deviation across samples σs in place of σresid <ref:2608. p. 822153#pg10>. When averaging over multiple two week periods, we use the same quadrature summation in Eq. p. 8 p. 8 1, but use the standard deviation across samples σs in place of σresid <ref:2608. p. 8 p. 822153#pg10>.

Expected NOx Emissions and Reported Turbine Power Output

Our figure for expected NOx emissions of 47 kg h−1 under best available control technology (BACT) requirements is based on public records from the Mississippi Department of Environmental Quality (MDEQ) <ref:2608.22153#pg10>. The MDEQ permit to construct 41 permanent turbines granted on 11 March 2026 [5] lists three turbine blocks (AA-000, AB-000, and AC-000) spanning three different turbine models <ref:2608. p. 822153#pg10>. All are natural gas-fired simple cycle combustion turbines equipped with Selective Catalytic Reduction (SCR) and oxidation catalyst (oxcat) <ref:2608. p. 822153#pg10>. The BACT emission limits for the three blocks are listed at 1. p. 845, 3.05, and 3.74 lb h−1 (0.66, 1.38, 1.70 kg h−1) respectively for AA-000, AB-000, and AC-000 <ref:2608. p. 822153#pg10>. Summed over all 41 turbines, this yields a fleet total of 46.9 kg h−1 (3-hour basis) or 41.6 kg h−1 (annualized) <ref:2608.22153#pg10>.

Our figures for deployed turbine capacity are based on public records from the Mississippi Commission on Environmental Quality (MCEQ) <ref:2608.22153#pg11>. Information on the Colossus 2 turbines reported to MCEQ includes the unit type, emission controls, nameplate capacity, expected local expected power output, date added to the site, and retirement date for each unit [4] <ref:2608. p. 922153#pg11>. In Figure 3, we chart the total nameplate capacity over time based on the date each turbine was added <ref:2608. p. 922153#pg11>. Total installed capacity according to MCEQ [4] is 1770. p. 972 MW, with 1540 MW attributed to turbines not equipped with SCR <ref:2608.22153#pg10>. To compute annualized emissions per kW of capacity post-February 2026, we divide our 870±165 kg h−1 estimate by the time-averaged capacity of installed turbines without SCR over the averaging period (933. p. 95 MW) <ref:2608. p. 022153#pg11>.

The study demonstrates that satellites can detect and quantify NOx emissions from AI data centers powered by non-SCR-equipped gas turbines, enabling independent monitoring of air pollution and compliance with environmental regulations <ref:2608.22153#pg10>. Similar analysis could be conducted using NO2-sensitive instruments with global coverage in low-Earth orbit <ref:2608.22153#pg10>. As AI data centers expand to multigigawatt scales, with facilities approaching 5 GW already under development [14] and ∼140 GW of new turbine capacity announced in the US alone [15], dedicated gas power generation could become an increasingly important source of air pollution <ref:2608. p. 322153#pg10>. Satellite observations provide a capability to monitor these emerging sources worldwide <ref:2608.22153#pg10>.

We use the TEMPO L2 V04 tropospheric NO2 vertical columns for our time averaging analysis, including only observations with the highest per-pixel quality flag (QA value of 0) and effective cloud cover below 0.2 albedo <ref:2608. p. 0 p. 022153#pg9>. For a baseline estimate, we compute non-rotated time averages from April through July for 2024, 2025, and 2026 <ref:2608. p. 7 p. 722153#pg9>. In 2026, this is a period of sustained high emissions from the Colossus 2 power plant with available CEMS data from the nearby TVA Southaven Power Plant <ref:2608. p. 7 p. 722153#pg9>. The vector-averaged PBL wind speeds for those periods are 1. p. 7 p. 79, 2.5, and 2.1 m s−1, with peak enhancements of 4. p. 7 p. 70 × 1015, 3. p. 22×10 15, and 5.3×10 15 molec cm−2, respectively <ref:2608. p. 7 p. 722153#pg9>. Between 2024 and 2025, the decrease in peak enhancement by ∼25% is consistent with the ∼32% increase in mean wind speed, and the peak enhancement falls about 8 km north and 6 km east of the Colossus 2 power plant <ref:2608. p. 7 p. 722153#pg9>. In 2026, the peak enhancement falls in the pixel immediately north of Colossus 2, and the 66% increase in peak enhancement cannot be explained by the ∼15% decrease in wind speed <ref:2608. p. 7 p. 722153#pg9>.

Plume Masking Procedure

To create plume masks for the wind-rotated two-week-average retrieval fields, we apply a total-variance filter and then mask downwind enhancements that are at least 2–3 standard deviations above the background noise floor <ref:2608.22153#pg9>. We require masks to begin within 10 km of the source location, discard mask regions smaller than 5 pixels, and remove regions outside the downwind section of the scene (±25° wedge from source location) <ref:2608.22153#pg9>. We then apply a sequence of erosion and dilation operations to ensure the final mask is relatively smooth and connected <ref:2608. p. 722153#pg9>.

Final Source Rate Calculation

We multiply these line integrals by the average HRRR PBL wind speed across all TEMPO observations and a NOx-to-NO2 ratio of 1.38±0.10 [11], representing photochemical steady-state conditions, to produce a downwind profile of sourcerate estimates <ref:2608.22153#pg10>. These estimates are expected to be biased low in the near-field of the source, where ozone titration by the emitted NO delays NO2 formation, and in the far-field due to NOx oxidation to HNO3 <ref:2608.22153#pg10>. To account for this, we discard estimates below the 80th percentile and report the mean of the remaining estimates <ref:2608.22153#pg10>. Our results are insensitive to the choice of percentile threshold; varying it between the 70th and 90th percentiles produces a spread of.

Q = aQˆ ± r Q σ ˆ a squared + σ squared resid. (1) <ref:2608. p. 022153#pg10>. where Qˆ is the initial CSF estimate before correction by the CEMS calibration slope a = 1. p. 878, σa = 0. p. 814 is the estimated 1σ uncertainty in a, and σresid = 213 kg h−1 is the standard deviation of the fit residuals <ref:2608. p. 822153#pg10>. When averaging over multiple two week periods, we use the same quadrature summation in Eq. 1, but use the standard deviation across samples σs in place of σresid <ref:2608.22153#pg10>. For our average after February 2026, we find σs = 165 kg h−1 across 11 samples <ref:2608.22153#pg10>.

We use the TEMPO L2 V04 tropospheric NO2 vertical columns for our time averaging analysis, including only observations with the highest per-pixel quality flag (QA value of 0) and effective cloud cover below 0.2 albedo <ref:2608.22153#pg9>. For a baseline estimate, we compute non-rotated time averages from April through July for 2024, 2025, and 2026 <ref:2608.22153#pg9>. In 2026, this is a period of sustained high emissions from the Colossus 2 power plant with available CEMS data from the nearby TVA Southaven Power Plant <ref:2608.22153#pg9>. The vector-averaged PBL wind speeds for those periods are 1.9, 2.5, and 2.1 m s−1, with peak enhancements of 4.0 × 10 15, 3.2×10 15 molec cm−2, and 5. p. 23×10 15 molec cm−2, respectively <ref:2608.22153#pg9>. Between 2024 and 2025, the decrease in peak enhancement by ∼25% is consistent with the ∼32% increase in mean wind speed, and the peak enhancement falls about 8 km north and 6 km east of the Colossus 2 power plant <ref:2608.22153#pg9>. In 2026, the peak enhancement falls in the pixel immediately north of Colossus 2, and the 66% increase in peak enhancement cannot be explained by the ∼15% decrease in wind speed <ref:2608.22153#pg9>.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Automated Regulatory Compliance Monitoring: An AI system can use satellite observations to assess regulatory compliance by detecting NOx emissions from data center power generation, allowing for independent monitoring of air pollution and compliance with environmental regulations.

  2. Emission Quantification and Prediction: The system can estimate two-week-average NOx source rates from space-based data, enabling researchers to quantify emissions that would be undetectable by our TEMPO analysis based on permit levels.

  3. Operational Efficiency Assessment: By comparing estimated emissions with expected levels from permits, the AI can infer an emission intensity of 0.93±0.18 g NOx kWh−1 for non-SCR-equipped turbines, providing a metric for assessing the true environmental cost of specific power generation methods used by data centers.

  4. Source Rate Calibration: The system utilizes a method to calibrate this source-rate retrieval against CEMS data at US power plants, which helps improve the accuracy of emission estimates from satellite observations by accounting for known emission patterns.

Related papers