The paper’s headline figures are already circulating. Gianluca Guidi, Francesca Dominici, Falco Bargagli-Stoffi and colleagues at Harvard’s T.H. Chan School of Public Health compiled facility records for 403 US hyperscale data centres operating between May 2024 and April 2025. They estimate that these facilities used about 82 TWh of electricity, roughly 1.8% of US consumption, and were associated with about 25.7 million tonnes of CO2. The Verge led with the fossil share: about 54% of the attributed generation came from gas and coal.
I found the more interesting story in the peer review file Nature Sustainability published with the article. The electricity estimate barely moved between the last two rounds of review. The carbon estimate fell by 42%, and one of the paper’s interpretive claims flipped. The reviewer who caught it had run the authors’ code and reported that it worked. As he put it: “it output the numbers they said it did.”
Same electricity, two carbon totals
The paper turns electricity into CO2 using EPA’s eGRID2023 plant data. It assigns each data centre to its balancing authority, the regional grid operator, and multiplies the facility’s estimated demand by that authority’s carbon intensity. The methods describe a generation-weighted attribution. Each plant gets a share of the load in proportion to its share of the authority’s generation, and that construction includes nuclear, hydro, wind and solar.
The code did something different. In the third set of reports, Alex Newkirk, the reviewer the journal credits by name, wrote that the notebook “defines the carbon intensity of a balancing authority as emissions intensity of only the combustion powerplants, effectively ignoring all non-fossil plants in each BA.” Combustion-only intensity was then applied to all of the data centres’ electricity. The result was an HDC-weighted carbon intensity of about 545 gCO2/kWh, and a central estimate of 44.6 million tonnes.
His first clue was arithmetic, not code. The manuscript said 38.2% of attributed generation was gas and 15.2% was coal. Using eGRID fleet averages of 616 g/kWh for gas and 986 g/kWh for coal, he got 385 g/kWh, “pretty far from 545.” When he recomputed the intensity on the total-output basis the SI described, he got 307 g/kWh against a national average of 348. He added that this “reduces the headline carbon estimates by roughly half, and notably reverses the core interpretive claim: HDC supplied by less carbon intensive power than the grid as a whole.”
The authors agreed. Their response says “the previous public implementation was inconsistent with the total-output attribution method described in the manuscript,” and that they had corrected the implementation and every affected result. They also dropped the earlier statement that hyperscale electricity was more carbon-intensive than the national grid average, because it “followed from the combustion-output denominator.”
The comparison that flipped
The old “dirtier than the grid” claim came from mixing two bases. In the same set of reports, a different reviewer summarised the paper’s comparison as about 545 gCO2/kWh for hyperscale electricity “versus roughly 370 gCO₂/kWh for the contemporaneous US grid average.” The 545 was a combustion-only number. The national average it was compared with included nuclear and renewables.
The published supplementary table puts both bases side by side. On the combustion-output basis, hyperscale electricity comes in at 543 g/kWh and the US average at 567. On the total-output basis, the figures are 314 and 348. Whichever basis you pick, as long as you use it for both, these data centres sit on somewhat cleaner grids than the US average, about 10% cleaner on total output. They looked dirtier only when the comparison used one basis for the data centres and the other for the country.
This doesn’t make the data centres clean. On the paper’s accounting, 25.7 million tonnes is still a lot of CO2 for 403 buildings, and more than half of the attributed generation is still fossil. The corrected intensity is also an annual average over a whole balancing authority. The authors now say plainly that it is “a location-based attributional average, not a marginal emissions factor.”
Average factors and the textbook size of this error
Emissions accountants know this error well, and the literature gives a fair idea of how large it usually is. Priya Donti, Zico Kolter and Inês Azevedo (2019) tested how common assumption choices change emissions and damage estimates in PJM, the grid region that holds 189 of this paper’s 403 facilities. One of the choices was whether to leave non-emitting generation out of an average factor. For PJM summer load they found that “fossil-only average emissions factors overestimate damages by 63% compared to average factors incorporating nonemitting generation.” Here, the combustion-only intensity of 543 is 73% above the total-output 314. So the error in this paper is the same kind as the one Donti’s team measured, and about the same size.
Noah Ryan, Jeremiah Johnson and Gregory Keoleian (2016) did a broader version for electric-vehicle charging. They reviewed 32 methods and models, then computed CO2 factors with 10 of them at nine charging sites. For a given site, the results differed by up to 68% from the mean factor, and by up to 63% among the average factors alone. Their conclusion was that method selection matters, and that the field needs a consensus on which method fits which load and question.
The corrected number still isn’t the answer to the question most people reading the headline are asking, which is what the next data centre will cause. Kyle Siler-Evans, Inês Azevedo and Granger Morgan (2012) computed the first systematic marginal emissions factors for the US grid. They found that average factors “may grossly misestimate” the avoided emissions from an intervention, and that the same efficiency measure avoids roughly 70% more CO2 in the Midwest than in the West. A new data centre that adds load at night in a gas-heavy authority, or brings its own turbines, does not draw the annual average mix. The paper doesn’t claim it does. Its 25.7 is an inventory figure. It answers who should be allocated what share of last year’s generation, and it can’t tell you what new load adds.
Running the code is the weaker check
The usual reproducibility tests ask whether the authors’ code, run on their data, reproduces the published numbers. Even that bar is often missed. Victoria Stodden, Jennifer Seiler and Zhaokun Ma (2018) sampled 204 computational papers from Science after its data-and-code policy took effect. They obtained artefacts for 44% and reproduced the findings for 26%. Christophe Pérignon and colleagues (2024) ran the code of 168 finance research teams on the same raw data and got exactly the same results only 52% of the time.
This paper’s second-round code would have passed that test. It ran, and it printed 545 and 44.6. What failed it was a second check that reproducibility audits mostly skip: does the code compute what the methods section says it computes? Newkirk got there twice, once with a back-of-envelope fuel-mix calculation that didn’t match, and once by reading a notebook cell against SI section 4.2. That’s the check I trust most, because it asks whether the number is roughly what the inputs imply.
The other number rests on a reviewer’s toy model
The electricity estimate deserves the same scrutiny, and the review file shows why. The data comes from Baxtel, a commercial vendor, and reports each facility’s electrical capacity in megawatts, not metered consumption. Annual energy is capacity times hours times a facility-load coefficient u. Earlier versions used u = 0.663 and reported 93.66 TWh. In the second set of reports, Newkirk, who says he co-authored the 2024 Lawrence Berkeley National Laboratory report the paper cited, objected that the coefficient came from a server-level model and was being applied to total facility capacity, which is sized for worst-case cooling plus contingency. “Even if it happens to land near the right answer for some facilities,” he wrote, “it would be a coincidence of offsetting errors.”
To show the sensitivity, he ran a toy power-flow model of a single site in Loudoun County, Virginia. With conventional servers it produced 57.6% utilisation on a measured-maximum design basis and 48.0% on a rated-maximum basis. The published central and low scenarios are u = 0.58 and u = 0.48. In the next round he asked that the text describe those values accurately, as a sensitivity sweep at one location, “best read as upper bounds on per-facility utilization for a lean, vertically integrated operator,” not as fleet averages. The authors did revise the description. They also cite an independent check based on PUE that gives about 0.59 across the top eight balancing authorities. The authors’ response still states the main issue: none of the four scenarios has been “validated against metered load for all 403 facilities.”
Two more caveats. First, the 403 facilities are 59.7% of the 675 Baxtel lists, and the authors say the excluded records are too incomplete to tell whether they skew small, new or regional. The totals apply to the sample, not to all US hyperscale capacity. Second, estimates built from capacity and assumed load have a long record of spreading wide. Arman Shehabi and colleagues (2018) put all US data centres at a nearly steady 70 billion kWh a year in the years before 2018. Their alternative 2020 projections differed by about 135 billion kWh depending only on how fast efficiency measures were adopted. A range of 68 to 99 TWh for 403 buildings fits that pattern. It’s a scenario band, and I’d read it that way.
What I take from it
I came away thinking better of the paper, not worse. The authors published the code, the reviewer ran it and read it, the journal published the review file, and the corrected number is in print. The process worked, but only because one reviewer did more than confirm the code ran.
Three practical points follow. If you’re quoting this paper, the carbon figure is 25.7 million tonnes on an attributional total-output basis, and the honest summary is that these data centres sit on slightly cleaner-than-average grids that are still more than half fossil. If you’re comparing carbon intensities, use the same denominator on both sides. If you’re reviewing a quantitative paper with a code release, don’t stop at confirming the numbers reproduce. Pick one headline figure, rebuild it from the stated inputs on paper, and see whether you land near it. Newkirk did that and got 385 when the paper said 545.
Guidi told The Verge that “the same computing infrastructure can have dramatically different consequences depending on the electricity system supporting it.” The same point applies to the accounting. With identical electricity in both rounds, a different denominator produced a different carbon total and reversed the comparison with the grid.
References
- Guidi, G., Dominici, F., Squartini, T., Sprinkle, C., Gilmour, J., Butler, K., Bell, E., Delaney, S. and Bargagli-Stoffi, F. J. (2026). Assessing the carbon emissions and energy consumption of US hyperscale data centres. Nature Sustainability, published 16 November 2026. Includes the Supplementary Information and Peer Review File.
- Guidi, G. et al. (2026). hyperscale-emissions: code and reproducibility materials. GitHub repository, README and
results/tables/, accessed 16 November 2026. - Calma, J. (2026). US data centers now burn more electricity than a midsized country. The Verge, 16 November 2026.
- Donti, P. L., Kolter, J. Z. and Azevedo, I. L. (2019). How Much Are We Saving after All? Characterizing the Effects of Commonly Varying Assumptions on Emissions and Damage Estimates in PJM. Environmental Science & Technology, 53(16), 9905-9914, August 2019.
- Ryan, N. A., Johnson, J. X. and Keoleian, G. A. (2016). Comparative Assessment of Models and Methods To Calculate Grid Electricity Emissions. Environmental Science & Technology, 50(17), 8937-8953, September 2016.
- Siler-Evans, K., Azevedo, I. L. and Morgan, M. G. (2012). Marginal Emissions Factors for the U.S. Electricity System. Environmental Science & Technology, 46(9), 4742-4748, May 2012.
- Stodden, V., Seiler, J. and Ma, Z. (2018). An empirical analysis of journal policy effectiveness for computational reproducibility. Proceedings of the National Academy of Sciences, 115(11), 2584-2589, 13 March 2018.
- Pérignon, C., Akmansoy, O., Hurlin, C., Dreber, A., Holzmeister, F., Huber, J. et al. (2024). Computational Reproducibility in Finance: Evidence from 1,000 Tests. The Review of Financial Studies, 37(11), 3558-3593, November 2024.
- Shehabi, A., Smith, S. J., Masanet, E. and Koomey, J. (2018). Data center growth in the United States: decoupling the demand for services from electricity use. Environmental Research Letters, 13(12), 124030, December 2018.