---
title: "Hyperscale Data Centers Use 1.8% of U.S. Power, but Owners Report a Fraction of the CO2"
description: "A Harvard-led study ties 403 buildings to 25.7 million tons of CO2, while Google and Amazon book less than a fifth of their grid-based power emissions."
author: "rews special report"
published: 2026-11-16T20:52:10Z
modified: 2026-11-17T13:34:24Z
url: https://rews.cc/a/403-data-centers-the-electricity-of-a-state-harvard-counts-w-292cbc
language: en
type: special report
tags: ["ai", "energy", "infrastructure", "climate", "data-centers", "tech"]
publisher: "Rews (https://rews.cc)"
---

# Hyperscale Data Centers Use 1.8% of U.S. Power, but Owners Report a Fraction of the CO2

*A Harvard-led study ties 403 buildings to 25.7 million tons of CO2, while Google and Amazon book less than a fifth of their grid-based power emissions.*

Special report · By rews special report · November 16, 2026 · https://rews.cc/a/403-data-centers-the-electricity-of-a-state-harvard-counts-w-292cbc

## In brief

- A Harvard-led study estimates 403 U.S. hyperscale data centers used about 82 TWh, 1.8% of U.S. electricity, from May 2024 to April 2025
- The published estimate ties them to 25.7 million metric tons of CO2, down from 37 to 54 million in a June preprint
- Google reported 2.8 million tons of market-based electricity emissions for 2025, against 15.1 million on a grid basis
- Princeton research found annual clean-power matching has little long-run effect on emissions, while hourly matching cuts them
- The GHG Protocol plans its next consultation in Q2 2027 and a final standard in Q4 2028

The 403 hyperscale data centers that a Harvard-led team could identify in the United States used about 82 terawatt-hours of electricity in the 12 months that ended in April 2025, and the power plants behind them emitted an estimated 25.7 million metric tons of carbon dioxide, according to a study [published Monday in Nature Sustainability](https://www.nature.com/articles/s41893-026-01985-7).

That is about 1.8 percent of the nation’s electricity, and a carbon load Harvard compared to Denmark’s. It is also a tonnage the companies running such buildings do not show in the climate reports they publish. Google, Amazon, Microsoft and Meta count their electricity emissions mainly under rules that let them subtract renewable-energy certificates and long-term power contracts from what their servers draw off the grid. Google’s report for 2025 put its electricity emissions at 2.8 million tons. Counted against the grids it actually used, the figure was 15.1 million.

Both numbers are permitted, and the distance between them is the subject of a long fight inside the GHG Protocol, the body that writes the carbon-accounting rules companies use. Peer-reviewed research finds that buying clean power to match a year’s consumption does little to cut emissions on real grids. Google’s own hour-by-hour measure scored its electricity 65 percent carbon-free last year. A proposal to make companies match clean power by the hour and by grid drew little or no support from 82 percent of companies that responded, and final rules are not expected before late 2028. The outside count has soft spots too: Harvard’s estimate is a model with a wide range, its carbon figure fell by more than 40 percent between draft and publication, and it leaves out the gas turbines data centers are now installing on their own land.

![Three server racks stand before a red city silhouette and factories releasing billowing smoke against a yellow background.](https://rews.cc/img/82508b87fbdd3dd6b04d48441db29e6d18b16f49.webp)

## What the study counted

The team was led by Gianluca Guidi, a postdoctoral fellow in biostatistics at the Harvard T.H. Chan School of Public Health. Francesca Dominici, a professor of biostatistics and director of the Harvard Data Science Initiative, was the senior author. They built their list of facilities from satellite imagery, estimated each building’s consumption from May 2024 to April 2025 under several operating scenarios, and got a range of 68 to 99 terawatt-hours around the central 82. Each site went into one of 23 grid regions, and emissions were assigned from Environmental Protection Agency data on the plants in each, [according to Harvard’s announcement](https://hsph.harvard.edu/news/hyperscale-data-centers-energy-consumption-carbon-emissions-newly-estimated-by-researchers/). The estimate ignores any operator’s clean-energy purchases and does not break out use by company.

Scott K. Johnson of [Ars Technica](https://arstechnica.com/science/2026/11/hyperscale-data-centers-use-2-of-us-electricity/) rounded the share to 2 percent and set it beside federal figures. A Lawrence Berkeley National Laboratory report put all U.S. data-center use, crypto mining excluded, at 76 terawatt-hours in 2018 and 192 in 2024. By that yardstick these 403 buildings are a third to a half of the national total.

About 54 percent of the electricity attributed to them came from fossil plants, 38 percent from gas and 15 percent from coal, with 25 percent from renewables and 21 percent from nuclear. That left the mix slightly cleaner than the country’s: about 314 grams of carbon dioxide per kilowatt-hour against 348 nationally, [Implicator.ai reported](https://www.implicator.ai/harvard-study-hyperscale-data-centers-fossil-fuels/) from the paper. Virginia, Oregon, Ohio and Iowa took more than half the electricity. Virginia alone had 142 of the sites.

The average hides a wide spread. The grid behind the Oregon sites, heavy with hydropower, emitted about 119 grams per kilowatt-hour. The grid behind the Iowa sites emitted about 445. Oregon ranked second in electricity use and sixth in emissions, [GeekWire reported](https://www.geekwire.com/2026/study-403-u-s-hyperscale-data-centers-rival-denmarks-emissions-oregon-ranked-no-2-in-power-use/). Dr. Guidi told [Time](https://time.com/article/2026/11/16/ai-data-centers-electricity-carbon-emissions-study) that the same electricity yields about seven times as much carbon dioxide in Wyoming as in Oregon. “So these environmental impacts are not necessarily inevitable,” he said. “Geography matters.”

Dr. Dominici gave [The Associated Press](https://www.nbcnews.com/tech/tech-news/harvard-study-estimates-magnitude-us-data-center-pollution-climate-cop-rcna609567) a comparison of her own.

> It is the equivalent to the CO2 emission of an entire state like Michigan or a small European country
>
> — Francesca Dominici, Harvard professor of biostatistics and senior author of the study, [nbcnews.com](https://www.nbcnews.com/tech/tech-news/harvard-study-estimates-magnitude-us-data-center-pollution-climate-cop-rcna609567)

Fengqi You, an energy systems scientist at Cornell who was not involved, said the study exists because the companies disclose so little. “We cannot manage environmental impacts effectively without reliable measurements and disclosure,” he told the AP.

## A number that moved

The headline figure was not always 25.7 million tons. A preprint of the paper, [posted to arXiv on June 3](https://arxiv.org/abs/2606.05420), estimated 37 to 54 million metric tons for the same 403 buildings and the same 68 to 99 terawatt-hours. It put their power at about 545 grams of carbon dioxide per kilowatt-hour, “about 48% above” a national average of 370. The journal version has them slightly cleaner than average. Harvard’s announcement does not say what changed.

The group’s earlier work pointed the same way as the draft. A [2024 preprint](https://arxiv.org/abs/2411.09786) by Dr. Guidi, Dr. Dominici and colleagues covered 2,132 data centers of every size between September 2023 and August 2024. It found they used more than 4 percent of U.S. electricity, 56 percent of it from fossil fuels, and attributed more than 105 million tons of carbon dioxide equivalent to them, at 548 grams per kilowatt-hour.

![Red-handled scissors cut through a black footprint connected to an electrical cord and plug against a bright blue background.](https://rews.cc/img/8f43a019efe48a67262f5de8cb22c657359fb513.webp)

Read together, the versions show which number is sturdier. The electricity range held from draft to journal. The carbon figure depends on how each region’s plants are matched to its load and on which year of plant records is used (the published paper used EPA data for 2023), and it moved by more than 40 percent. Both versions rest on grid averages, the location-based logic the companies themselves publish next to their headline numbers.

## Two sets of books

Under the GHG Protocol’s 2015 guidance, a company can report electricity emissions two ways. The location-based method multiplies its consumption by the average emissions of the grids it draws from. The market-based method assigns the emissions of whatever power it holds contracts or certificates for, so electricity matched by a renewable-energy certificate or a power purchase agreement counts as zero, even when the match is made over a whole year and from a distant grid. The protocol’s [proposed revision](https://ghgprotocol.org/blog/upcoming-scope-2-public-consultation-hourly-matching-and-deliverability) would require matching by hour and by deliverable region instead.

![A smoking factory forms a giant electrical plug connected by a teal river to a small hydroelectric dam against a yellow background.](https://rews.cc/img/ba3fcae4acd69e046aa17bc47a8525932a4206d3.webp)

Google’s environmental report, released June 30, shows what the choice is worth. Kate Brandt, the chief sustainability officer, wrote in [a blog post](https://blog.google/company-news/outreach-and-initiatives/sustainability/2026-environmental-report/) that Google “successfully reduced our operational emissions by 2%” as electricity demand rose 37 percent. That 2 percent is the market-based figure. On the location-based method Google’s electricity emissions rose 37 percent, to 15.1 million tons, while the market-based number fell 3 percent, to 2.8 million, according to [an analysis by Mehar U Nisa](https://thedecadeahead.substack.com/p/google-says-100-renewable-what-does) in her Substack newsletter, The Decade Ahead.

Google, which aims to run on carbon-free power every hour by 2030, also publishes an hourly score, and it is plainer than the annual one. The company matched 100 percent of its annual electricity with renewable purchases for a ninth straight year. Hour by hour, its use was 65 percent carbon-free in 2025, down from 66. In the PJM grid that serves Northern Virginia, the score was 57 percent. The newsletter calculated that 19 percent of Google’s market-based cuts came from certificates bought on the open market, and that of more than 12 gigawatts of clean power contracted in 2025, about 2 gigawatts began operating that year. “Check the calculation before you repeat the headline,” it advised.

Amazon said in its [2025 sustainability report](https://sustainability.aboutamazon.com/2025-amazon-sustainability-report.pdf) that all the electricity it used was matched with renewable energy for the third year running. Its formula adds the renewable power already on the grids where it operates to the output of projects it backed, and divides by its global use. It does not disclose total consumption. For 2024, its location-based electricity emissions were 17.76 million metric tons and its market-based figure 2.80 million, according to [Ernst & Young’s assurance statement](https://sustainability.aboutamazon.com/2024-ghg-verification-scope-1-2.pdf). For both Amazon and Google the market-based number is less than a fifth of the grid-based one. Amazon said emissions from purchased electricity rose 34 percent in 2025, driven partly by data centers.

![A large gauge is obscured by black smoke rising from a red machine in front of a factory on a teal background.](https://rews.cc/img/fee2454939b5446db3057dca539620821e390218.webp)

Meta’s [2025 sustainability report](https://www.smartenergydecisions.com/wp-content/uploads/2025/10/Meta_2025-Sustainability-Report_-compressed.pdf) put the effect in a sentence: total 2024 emissions “with contractual instruments applied (8.2 M MT CO2e) were 48% less than our location-based emissions (15.6 M MT CO2e).” Shareholders led by As You Sow said in [a filing](https://www.sec.gov/Archives/edgar/data/0001326801/000121465926005915/o511264px14a6g.htm) before Meta’s May 27 annual meeting that its location-based electricity emissions rose nearly 150 percent from 2019 to 2024 and that its renewable claim “relies heavily on RECs.”

Microsoft shows the certificates’ weight from the other side. In February 2025 it stopped buying unbundled certificates, the kind sold apart from any power. In July, Brad Smith, the vice chair and president, and Melanie Nakagawa, the chief sustainability officer, [wrote](https://blogs.microsoft.com/on-the-issues/2026/07/09/responsibly-building-the-ai-future/) that total emissions rose 25 percent, driven by data-center expansion and by “pausing our use of non-additional, unbundled renewable energy certificates.” Electricity went from nearly 2 percent of the reported footprint to 13 percent. Microsoft still says it matched 100 percent of its annual electricity with renewables, and [Stand.earth](https://stand.earth/insights/microsofts-un-sustainability-report-2026-analysis/), an environmental group, wrote that the report leaves out its gas plant projects.

None of this is new. A [Guardian analysis by Isabel O’Brien](https://www.aol.com/finance/data-center-emissions-probably-662-150002099.html) in September 2024 took certificates and offsets out of the reports of Google, Microsoft, Meta and Apple and estimated that emissions from their own data centers in 2020 to 2022 were probably 662 percent, or 7.62 times, higher than reported. For 2022, Meta’s official data-center figure of 273 metric tons became more than 3.8 million. Microsoft’s 280,782 became 6.1 million.

Investors now say so openly. Natasha Landell-Mills of Sarasin & Partners, an asset manager, was quoted in an Ethical Corporation article [distributed by Reuters in May](https://www.kitco.com/news/off-the-wire/2026-05-26/big-tech-splits-over-proposed-shake-clean-energy-accounting).

> Accounting for carbon emissions is too often a game of smoke-and-mirrors
>
> — Natasha Landell-Mills, Sarasin & Partners, [kitco.com](https://www.kitco.com/news/off-the-wire/2026-05-26/big-tech-splits-over-proposed-shake-clean-energy-accounting)

## The long view

Worry about data-center power predates the cloud. In August 2007 the Environmental Protection Agency told Congress, in [a report prepared by Lawrence Berkeley](https://www.energystar.gov/ia/partners/prod_development/downloads/EPA_Datacenter_Report_Congress_Final1.pdf), that servers and data centers used about 61 billion kilowatt-hours in 2006, 1.5 percent of U.S. electricity, at a cost of about $4.5 billion. It projected 100 billion by 2011.

The forecast missed. Berkeley Lab’s next full count, [published in June 2016](https://newscenter.lbl.gov/2016/06/27/data-centers-continue-proliferate-energy-use-plateaus/), estimated 70 billion kilowatt-hours in 2014, about 1.8 percent of the national total. Growth had slowed to about 4 percent over 2010 to 2014, from 24 percent in the five years before and nearly 90 percent in 2000 to 2005. The lab credited the 2008-09 recession, virtualization and other efficiency work.

Then the curve turned. The lab’s [December 2024 report](https://eta-publications.lbl.gov/sites/default/files/2024-12/lbnl-2024-united-states-data-center-energy-usage-report_1.pdf), required by the Energy Act of 2020, revised 2014 down to 58 terawatt-hours and found use had reached 176 by 2023, 4.4 percent of U.S. electricity. Annual growth went from about 7 percent in 2014 to 2018 to 18 percent in 2018 to 2023. Servers with A.I. accelerators went from under 2 terawatt-hours in 2017 to more than 40 in 2023. The lab projected 325 to 580 terawatt-hours by 2028, or 6.7 to 12 percent of the country’s power, [the Energy Department said](https://www.energy.gov/articles/doe-releases-new-report-evaluating-increase-electricity-demand-data-centers).

Against that record the Harvard number is simple to place. In 2014, by the lab’s estimate at the time, every data center in the country together used about 1.8 percent of its electricity. A decade later, 403 buildings use the same share.

![Cream balance scales with a lightning bolt support weigh many small cream buildings against three large red buildings on a blue background.](https://rews.cc/img/c929219257ad505ba6eb41772e555074c95d26fd.webp)

## What a certificate buys

Whether a certificate cuts emissions can be tested, and has been. In [a January 2024 paper in Joule](https://techxplore.com/news/2024-01-power-strategies-emissions.html), Qingyu Xu and Jesse Jenkins of Princeton’s ZERO Lab modeled corporate clean-power buying in the western United States. Matching a year’s consumption with clean generation, the most common strategy, had almost no effect on long-run emissions. Matching every hour was the only strategy that consistently lowered emissions across the system. Their explanation: wind and solar are now cheap enough that most projects would find buyers and get built anyway.

Hourly matching costs more, over $20 a megawatt-hour in some regions, while annual matching costs almost nothing extra. “There are beginning to be actual legal and financial ramifications for greenhouse gas accounting,” Dr. Jenkins said when the paper came out. “It’s not just about making marketing claims anymore.”

Meta paid for its own test. WattTime, which estimates the grid emissions caused or avoided by demand and supply, examined Meta’s 2023 data-center load and clean-power buying under three accounting methods in [a case study published May 30, 2025](https://watttime.org/news-and-insights/case-study-carbon-accounting-approaches-and-an-analysis-of-metas-2023-data-center-electricity-consumption-and-clean-energy-procurement/). The shareholders’ filing quotes its finding that “the emissions avoided by \[Meta’s\] clean energy investments were, on average, lower than the emissions induced by data center load.” WattTime drew a conclusion that favors “carbon matching,” the approach Meta and Amazon back, over hourly rules.

Dr. Jenkins and Wilson Ricks, then a Princeton doctoral candidate, walked through the modeling behind the hourly-matching findings in a public webinar.

[![System-Level Impacts of Voluntary Carbon Free Electricity Procurement Strategies](https://i.ytimg.com/vi/tg_suAJoeVo/hqdefault.jpg)](https://www.youtube.com/watch?v=tg_suAJoeVo)

Princeton’s Jesse Jenkins and Wilson Ricks present research on the system-level effects of voluntary carbon-free electricity procurement strategies. Video: Energy Systems Integration Group · YouTube

## The case for the books

The companies that lean hardest on certificates have an argument. Amazon and Meta are founding members of the Emissions First Partnership, formed in late 2022 with Salesforce, Intel, General Motors and others. In [a letter to the GHG Protocol](https://www.emissionsfirst.com/letter-to-the-ghg-protocol) dated July 11, 2025, the partnership said hourly matching “does not measure impact” and wrote: “In many cases, hourly matching may increase emissions despite enabling companies to claim a MBM reduction.” Its example was wind in low-carbon West Texas matched to load in high-carbon East Texas, which it said drives net emissions up.

In April, 48 companies with more than $4.7 trillion in combined annual revenue, among them Amazon, Apple, FedEx and General Motors, told the protocol they were “extremely concerned” that mandatory hourly and deliverability rules would discourage voluntary clean-power buying and raise prices, [ESG Today reported](https://www.esgtoday.com/apple-amazon-schneider-electric-warn-ghg-protocol-that-proposed-scope-2-accounting-changes-will-slow-corporate-energy-transition/). The change, they wrote, “fundamentally threatens participation in this voluntary market.”

Corporate money runs through the research on both sides. InfluenceMap, which tracks corporate lobbying, found in [an Aug. 28 analysis](https://influencemap.org/insight/Corporate-Engagement-on-the-GHG-Protocol-Scope-2-Consultation-39494) that REsurety, an analytics firm whose work has been funded by Meta and Microsoft, and WattTime, funded by Meta and Apple, opposed mandatory hourly matching, while Google-funded academics at Princeton and TU Berlin supported it. Google was the only one of the four big cloud companies to take a public position, backing the changes. Lee Taylor of REsurety stated the objection in the Reuters-distributed piece.

> There is a very real cost of hourly and location matching, but the benefits are not clear.
>
> — Lee Taylor, REsurety, [kitco.com](https://www.kitco.com/news/off-the-wire/2026-05-26/big-tech-splits-over-proposed-shake-clean-energy-accounting)

Killian Daly of EnergyTag, a group that promotes hourly certificates, said in the same article that the proposal concerns counting, not buying. “What the protocol is proposing is hourly and locational accounting, which is very different from making 24/7 compulsory,” he said.

Weighed against the record, the companies’ best point holds in part. Their contracts did pay for real plants. Google says it has signed agreements for nearly 35 gigawatts of clean power since 2010. But where those plants sit on the same grids as the data centers, their output is already in the regional averages Harvard used, shared among every customer. The location-based method does not ignore them. It declines to hand all of their output to the buyer, including the hours when a solar farm is dark and the servers run on gas.

The West Texas example is sound physics and an argument for crediting clean power where it displaces the most carbon. It is not a defense of the annual certificate system in use now, which lets a year of wind certificates cancel a year of grid power regardless of hour. Both studies that tested that system cut against it: Princeton found annual matching barely moves long-run emissions, and the study Meta itself commissioned found that annual matching understated Meta’s actual carbon impact.

The outside count has its own weak points, and they cut both ways. It is a model built from satellite images, estimated capacities and assumed utilization, not meters, which is why the electricity range runs from 68 to 99 terawatt-hours. It uses 2023 plant data for a period running into 2025, cannot separate A.I. from other computing, misses some facilities and leaves out on-site generation. And with no breakdown by operator, it cannot be checked against any one company’s report. Dr. You’s point applies here: only the operators hold metered data for each building, and among the four only Meta publishes emissions by data center.

## Readers’ objections

Readers fought more over the comparisons than the accounting. On Hacker News, in a thread on Wired’s coverage, headlined “Just 400 US Data Centers Use More Electricity Than Michigan,” one reader set Harvard’s wording beside Wired’s.

> The study found that these HDCs, depending on how hard they were running, together used between 68 and 99 terawatt-hours of electricity ... Wired is playing in your faces.
>
> — ibn\_khaldun on [Hacker News](https://news.ycombinator.com/item?id=50522882)

The reader’s numbers check out. Michigan’s retail electricity sales in 2024 were 98,991,014 megawatt-hours, about 99 terawatt-hours, [according to the Energy Information Administration](https://eia.gov/electricity/state/michigan). The central estimate of 82 is about 83 percent of that, so the buildings out-consume Michigan only at the top of the range. Another reader replied that the abstract itself gives 82 terawatt-hours and 1.8 percent, which is accurate and does not support “more than Michigan” either.

A second reader, who said he had read Harvard’s summary and the preprint, took issue with the word “just.”

> One thing to note is that the framing “Just 400 US Data Centers” makes it seem like 400 out of the gazillion data centers use this much electricity, but the original research seems to be that they found exactly 403 “Hyperscale Data Centers” and evaluated them.
>
> — andrewla on [Hacker News](https://news.ycombinator.com/item?id=50521656)

That matches the paper, which covers every hyperscale site the team could document, not a top slice. Others answered that the point changes little, since the 403 are the biggest buildings by definition and, on Ars Technica’s arithmetic, a third to a half of all U.S. data-center power. Both hold. The state comparisons are a framing choice, and 1.8 percent is the figure the paper gives.

## Behind the meter

A growing source of data-center carbon sits outside the study entirely. During the study period, xAI’s data center in Tennessee ran a fleet of gas turbines on site, power generated “behind the meter” and invisible to grid-based accounting. Cleanview, a market intelligence company, [now tracks](https://cleanview.co/) 116 plants that data centers are building or planning for themselves, totaling 126,553 megawatts from announced to operating. By midyear about 2 gigawatts was running, across four projects, [its report found](https://cleanview.co/reports/behind-the-meter-data-centers).

The turbines and engines used on site are less efficient than utility plants and emit 60 to 80 percent more greenhouse gas for the same electricity, Ars Technica reported. Michael Thomas, Cleanview’s founder, wrote on X that hooking a hyperscale campus to the grid can take as long as seven years in Virginia, while building its own power in a state with looser rules can take under two.

Mr. Thomas has followed Amazon, which reports all of its electricity as renewable-matched, into gas. The company confirmed to him that it was behind a site in Pecos County, Texas, planned around a 7.65-gigawatt gas plant. He later posted a fuller count.

> Last month, our team at Cleanview broke the news that Amazon is building the biggest natural gas plant in America.
>
> Since then, we’ve identified 8 gas plants linked to Amazon data centers with a combined capacity of 17.5 GW.
>
> Amazon, like other hyperscalers, have tried to avoid…
>
> — **Michael Thomas** @curious\_founder on X · [September 24, 2026](https://x.com/curious_founder/status/2103160926786773460)

One Hacker News reader in the Wired thread said the power source, not the load, is the trouble.

> Datacenter power consumption wouldn’t be an issue if that power came from local renewable sources. Unfortunately, the gas-fired turbines that are being built now feel like quite a regression.
>
> — renehsz on [Hacker News](https://news.ycombinator.com/item?id=50521482)

Another replied that gas wins because it is on hand while grid upgrades carry a backlog of a decade or more, which squares with Mr. Thomas’s account of interconnection delays. Towns asked to host the buildings have pushed back either way, [at the ballot box](https://rews.cc/a/voters-turned-on-the-data-centers-and-the-tech-lobby-says-th-43cc3a) and against [small projects in Salem and Concord](https://rews.cc/a/data-center-backlash-reaches-small-projects-in-salem-and-con-9128f6).

## What comes next

The paper came out the day Turkey, host of the COP31 climate talks in Antalya, and the International Telecommunication Union launched the Antalya Artificial Intelligence Pledge. Signers commit to running data centers on clean power with efficient energy and water systems, disclosing power, water and emissions, and assessing and publishing their environmental impacts. It drew 80 endorsements at launch, [Agence France-Presse reported](https://www.france24.com/en/live-news/20261116-un-climate-chief-warns-of-ai-double-edged-sword-at-cop31), and the United States, China and the European Union [did not endorse it](https://www.japantimes.co.jp/environment/2026/11/17/sustainability/un-climate-chief-ai-cop31/).

Simon Stiell, the United Nations climate chief, spoke at the launch.

> Claims about future emissions savings cannot excuse rising pollution now, making climate disasters even worse.
>
> — Simon Stiell, executive secretary of U.N. Climate Change, [nbcnews.com](https://www.nbcnews.com/tech/tech-news/harvard-study-estimates-magnitude-us-data-center-pollution-climate-cop-rcna609567)

Caleb Max, president of the National Artificial Intelligence Association, told the AP that “the fears may be overblown.” Philip Newell, a spokesman for Climate Action Against Disinformation, said of the industry, “They’re selling us hope while hiding the harms from us.”

The Harvard team plans to measure health effects next, from the fine particulate matter, nitrogen oxides and sulfur dioxide that fossil plants emit. “Quantifying health impacts is an important next step,” Dr. Dominici said. Dr. Guidi said the footprint may already be much larger than the study’s snapshot.

The accounting rules will take longer. In October the GHG Protocol’s Independent Standards Board proposed two market-based methods, one requiring hourly matching and deliverability and one keeping the 2015 rules until a cutoff it has not set, [ESG Today reported](https://www.esgtoday.com/ghg-protocol-to-push-back-stricter-scope-2-reporting-rules-after-corporate-objections/). Only 12 percent of company respondents had supported hourly matching. The next public consultation, on a joint standard with the International Organization for Standardization, is planned for the second quarter of 2027, with the final standard due in the fourth quarter of 2028, [according to CarbonCredits.com](https://carboncredits.com/ghg-protocol-scope-2-rules-corporate-pushback/).

## Join the discussion

- [Just 400 US Data Centers Use More Electricity Than Michigan](https://news.ycombinator.com/item?id=50520697) — Hacker News · 92 points · 94 comments

## See also

- [Google 2026 Environmental Report](https://sustainability.google/files/google-2026-environmental-report.pdf) — sustainability.google · Full location-based and market-based tables and hourly carbon-free scores by grid region
- [GHG Protocol Scope 2 public consultation summary of feedback](https://ghgprotocol.org/sites/default/files/2026-08/S2-PublicConsultationSummaryofFeedback-2026.08.25.pdf) — ghgprotocol.org · How nearly 1,100 respondents answered the hourly matching and deliverability proposals
- [Microsoft 2026 Environmental Data Fact Sheet](https://cdn-dynmedia-1.microsoft.com/is/content/microsoftcorp/microsoft/msc/documents/presentations/CSR/2026-Microsoft-Environmental-Data-Fact-Sheet-PDF.pdf) — cdn-dynmedia-1.microsoft.com · Microsoft's Scope 2 figures under both methods back to its 2020 baseline
- [Meta 2025 Environmental Data Index](https://sustainability.atmeta.com/wp-content/uploads/2025/10/Meta_2025-Environmental-Data-Index.pdf) — sustainability.atmeta.com · Meta's data, including emissions broken out by individual data center
- [WattTime case study of Meta's 2023 data center emissions](https://watttime.org/wp-content/uploads/2025/06/WattTime-Meta-Emissions-Accounting-Case-Study-vFinal-202505b.pdf) — watttime.org · The full Meta-commissioned comparison of annual, hourly and carbon matching
- [Princeton Andlinger Center: system-level impacts of 24/7 carbon-free electricity](https://acee.princeton.edu/24-7/) — acee.princeton.edu · The ZERO Lab research program behind the hourly matching findings
- [NewClimate Institute briefing on 24/7 matching and the Emissions First proposal](https://newclimate.org/news/briefing-247-renewable-electricity-matching-is-a-far-more-credible-approach-for-the-ghg) — newclimate.org · A detailed critique of impact-based accounting from a climate policy group
- [Congressional Research Service: Data Centers and Their Energy Consumption, FAQ](https://www.congress.gov/crs-product/R48646) — congress.gov · A plain-language federal primer on data-center power estimates
