Skip to main content

AI and the Earth

Where the Cloud Lands

The cloud is made of concrete, copper and power plants. Every answer a model gives is drawn from a grid somewhere, and that grid burns something. A new peer-reviewed study maps the carbon behind 403 of the largest data centres in the United States, region by region.1

403
hyperscale data centres analysed across the United States
81.8 TWh
of electricity a year, central estimate
25.7 Mt
of CO₂ a year from the power they draw
314 g
of CO₂ per kilowatt-hour, on average

The map

Same servers, different skies

A data centre has no chimney. Its emissions leave from the power plants of the grid it plugs into. Guidi and colleagues estimated how much electricity each of 403 hyperscale facilities draws, then traced that load to the grid region that supplies it, using the US Environmental Protection Agency's record of what every region burns.12

The result is a geography of carbon. Virginia alone holds 142 of the 403 facilities, all of them on the PJM grid, which draws about 60% of its power from gas and coal. That one region carries about four tonnes in ten of the national total. Washington State's facilities, fed by Columbia River dams, emit almost nothing.3

The map shows the United States only, because the study does. Two reasons make it the place to start. The country hosts 39% of the 9,347 data centres identified worldwide, 3,619 of them,4 and its environmental agency publishes a plant-by-plant record of what every regional grid generates and emits, which is what lets a kilowatt-hour be traced back to a chimney.2 Mapping other countries the same way would need the same kind of public grid record.

Hyperscale data centres in the United States, by grid region

Carbon intensity of the regional grid, g CO₂ per kWh
Discs sit at the centre of each grid region, not on individual buildings: the study's facility list is licensed from Baxtel and cannot be published. Regions are EPA balancing authorities, the grid operators that match supply and demand. Central scenario, an assumed facility load of 58% of rated capacity. Sources: Guidi et al. 202613, EPA eGRID 20232. White points, when shown: frontier AI data centres tracked by Epoch AI5, a separate dataset that includes sites still under construction.
Back to top

The paper in four figures

What the study found

Four questions organise the paper. Where do the emissions fall? Why does the same kilowatt-hour weigh so differently from one region to the next? What do the plants behind the servers burn? And how much does the answer depend on how hard the servers run?1

A note for anyone who read the first coverage. An earlier version of the study put the total near 53 million tonnes. The published paper divides each region's CO₂ by all the electricity it generates, nuclear and renewables included, rather than by fossil generation alone. On that basis the total is 25.7 million tonnes, and the electricity behind the servers is slightly cleaner than the US average of 348 g/kWh, not dirtier. The fossil-only figure is kept as a diagnostic, and you can switch to it in the last chart below.1

Back to top

Beyond the United States

When a country becomes a server farm

Most countries give their data centres one or two percent of their electricity. Ireland gives them nearly a quarter, up from one twentieth in 2015, and the Netherlands has tripled its share since 2017.67 The United States, home to more data centres than any other country, is now close to five percent.8

Data centres' share of electricity use, 2015–2025

Ireland: data centres' share of all metered electricity, each year the sum of four quarters. Netherlands: electricity delivered to data centres through the public grid, as a share of all electricity used; 2024 is provisional. United States, China, Europe and the world: the International Energy Agency's estimate of data-centre use over total demand from Ember, published for 2020 and 2023 to 2025 only, so those lines run straight across 2021 and 2022 and the dots mark the published years. National statistics and the agency's estimates are measured differently: compare the trends more than the decimals. Sources: CSO Ireland6, CBS Statistics Netherlands7, Our World in Data based on the IEA and Ember8.
Back to top

Whose water?

Servers where water runs short

A data centre draws on the same rivers and aquifers as the towns, farms and factories around it. Rosa and colleagues placed all 9,347 known data centres on a map of water gaps, the months when people already use more water than their place renews.49 A quarter sit where that happens at least one month a year. In a world 3 °C warmer, a third would.

Data centres by months of water gap a year, by country

25% of the world's 9,347 data centres, 2,315 of them, face at least one month of water gap a year in the climate of 2001–2010.

The 15 countries with the most data centres. Each bar splits them by months of water gap; the number on the right is the share facing at least one month.

One disc per country, sized by its number of data centres and placed at the country's centre, never at a facility: the study's site list is proprietary. In each disc the cyan outer ring is data centres with no month of water gap; gold, orange and red count those facing 1 to 4, 5 to 8 and 9 to 12 months, the longest at the core. Areas are true to the counts. A water gap is a month when people, farms, industry and power plants consume more than the renewable water left once rivers keep what their ecosystems need. Today means today's data centres in the 2001–2010 climate; +3 °C is the average of five climate models, rounded to whole data centres. The figure shows exposure to water stress, not the water data centres use. Sources: Rosa et al. 2026410; water gap glow, Rosa and Sangiorgio 20259.
Back to top

Beyond carbon

Water, minerals, land

Water

1 in 4

Of the 9,347 data centres mapped worldwide, 2,315 already sit where water demand exceeds supply for at least one month a year. Under future hydrology, one in three.4

39%

The United States hosts 3,619 data centres, 39% of the world total. 21% of them are in water-scarce regions, rising to 35% in a world 3 °C warmer.4

731–1,125 million m³

Water that AI servers in the United States would consume each year between 2024 and 2030, counting cooling and the power plants behind them.11

313–765 billion L

Water footprint of AI systems worldwide in 2025, in the same range as all the bottled water people drink in a year (446 billion litres).12

Minerals and waste

>10×

Growth in lithium demand for clean energy on a path that holds the Paris goals, and more than seven times for rare earths. That demand comes from the energy transition itself. Data centres will draw on the same minerals.13

1.2–5 Mt

Electronic waste that generative AI could pile up in total between 2020 and 2030. A 2026 recalibration, which caps the estimate by how many AI servers can actually be made, finds 131 to 225 thousand tonnes a year by 2030. The estimates are still far from settled.1415

Land and communities

160+

New data centres sited in water-stressed US regions in the three years before 2025, a 70% increase over the preceding period.16

$64 billion

US data-centre projects blocked or delayed by local opposition between May 2024 and March 2025. In July 2026 New York declared the first statewide moratorium on large data centres.1716

What choices can change

−49% to +90%

Where AI servers are built swings their carbon footprint: from half the emissions in the best US locations to nearly double in the worst.11

1 week in 4

In a replay of 2024 data for six Google sites, drought rules overruled the lowest-carbon location in up to a quarter of weeks. The cleanest grid is not always the wettest basin.18

3 tests

The Earth alignment principle asks that AI respect planetary boundaries, spread its benefits without concentrating power, and strengthen social cohesion, with mandatory environmental reporting.19

Back to top

The wider footprint

The planetary cost, worldwide

Carbon is one line in a longer bill. Data centres draw electricity, evaporate water to stay cool, and are built from minerals mined on other continents. The four biggest cloud companies all reported rising emissions in their latest reports. Google's are 82% above 2019 and Microsoft's 62% above 2020, while both keep their 2030 climate goals. Amazon's rose 16% in 2025 alone.20212223 Six views of the global picture, from public reports and peer-reviewed estimates.

AI's footprint, worldwide

Hover any bar or line for details. Energy projections vary widely by source and method. Per-model emissions come from developers where reported, otherwise from published estimates. Energy: IEA 202524, Carbon Brief 202525, EPRI 202426, Xiao et al. 202511. Carbon: de Vries-Gao 202612, Meta27, Gemma Team 202428, Chowdhery et al. 202229, Patterson et al. 2021 and 20223031, Luccioni et al. 202332. Water: Li et al. 202533, Google 202620, Microsoft 202434, TSMC35. Companies, latest report, scopes 1 to 3, market-based (Meta's latest covers 2024): Google 202620, Microsoft 202621, Amazon 202622, Meta 202523. Grid strain: company announcements36, NERC37. Materials: USGS 202438, Global E-waste Monitor 202439.
Back to top

Sources

References

  1. Guidi, G., Dominici, F., Squartini, T., Sprinkle, C., Gilmour, J., Butler, K., Bell, E., Delaney, S. & Bargagli-Stoffi, F. J. (2026). Assessing the carbon emissions and energy consumption of US hyperscale data centres. Nature Sustainability. doi.org/10.1038/s41893-026-01985-7
  2. US Environmental Protection Agency (2025). Emissions & Generation Resource Integrated Database (eGRID), 2023 data, revision 2. Balancing-authority generation, fuel mix and CO₂ output rates. www.epa.gov/egrid
  3. Guidi, G. (2026). hyperscale-emissions, v0.3.0: code and public aggregate tables (state and balancing-authority level, central scenario). github.com/gianguidi/hyperscale-emissions
  4. Rosa, L., Sangiorgio, M., Scanlon, B. R., Mazzotti, M., Berry, J. & Michalak, A. M. (2026). One quarter of global data centers are located in water-stressed areas. Environmental Science & Technology 60, 27780–27791. doi.org/10.1021/acs.est.6c08088
  5. Epoch AI (2026). Data on AI Data Centers (Frontier Data Centers), retrieved 10 November 2026. Licensed CC BY 4.0. Site locations, power today and planned peak power. epoch.ai/data/ai-data-centers
  6. Central Statistics Office Ireland (2026). MEC02, Metered electricity consumption: data centres and all metered customers, quarterly, 2015–2025. CC BY 4.0, retrieved 18 November 2026. data.cso.ie/table/MEC02
  7. CBS Statistics Netherlands (2025). Levering van elektriciteit via het openbaar net aan datacenters, 2017–2024 (electricity delivered to data centres through the public grid, as a share of total electricity use; 2024 provisional). CC BY 4.0, retrieved 18 November 2026. www.cbs.nl/-/media/_excel/2025/51/datacenters_elektriciteitsleveringen_2017_2024.xlsx
  8. Our World in Data (2026). Share of total electricity demand used by data centers, 2020–2025: data-centre consumption from the IEA, Energy and AI (2025, 2026), over total demand from Ember (2026), processed by Our World in Data. IEA data © OECD/IEA; Ember CC BY 4.0. Retrieved 18 November 2026. ourworldindata.org/grapher/data-centers-share-electricity-demand
  9. Rosa, L. & Sangiorgio, M. (2025). Global water gaps under future warming levels. Nature Communications 16, 1192. Gridded water gaps at 0.5°, baseline and +3 °C five-model mean, Zenodo (doi:10.5281/zenodo.13220763), CC BY 4.0, retrieved 18 November 2026. doi.org/10.1038/s41467-025-56517-2
  10. Rosa, L., Sangiorgio, M., Scanlon, B. R., Mazzotti, M., Berry, J. & Michalak, A. M. (2026). Supplementary dataset for "One quarter of global data centers are located in water-stressed areas": SummaryByCountry.xlsx, data centres by country and months of water gap a year, 2001–2010 climate and +3 °C. Zenodo, CC BY 4.0, retrieved 18 November 2026. doi.org/10.5281/zenodo.22912911
  11. Xiao, T., Fuso Nerini, F., Matthews, H. D., Tavoni, M. & You, F. (2025). Environmental impact and net-zero pathways for sustainable artificial intelligence servers in the USA. Nature Sustainability 8, 1541. doi.org/10.1038/s41893-025-01681-y
  12. de Vries-Gao, A. (2026). The carbon and water footprints of data centers and what this could mean for artificial intelligence. Patterns 7, 101430. doi.org/10.1016/j.patter.2025.101430
  13. UNEP (2024). Navigating New Horizons: a global foresight report on planetary health and human wellbeing. Nairobi, pp. 24 and 31 (citing IEA 2024). doi.org/10.59117/20.500.11822/45890
  14. Wang, P., Zhang, L.-Y., Tzachor, A. & Chen, W.-Q. (2024). E-waste challenges of generative artificial intelligence. Nature Computational Science 4, 818–823. doi.org/10.1038/s43588-024-00712-6
  15. de Vries-Gao, A. (2026). Recalibrating global artificial intelligence e-waste estimates. Resources, Conservation and Recycling 229, 108872. doi.org/10.1016/j.resconrec.2026.108872
  16. Sultana, F. (2026). The Heavy Cloud: hydrosocial dispossession and the water coloniality of AI infrastructure. Geo: Geography and Environment 13, e70113. doi.org/10.1002/geo2.70113
  17. Data Center Watch (2025). $64 billion of data center projects have been blocked or delayed amid local opposition. Research period May 2024 to March 2025. www.datacenterwatch.org/report
  18. Kollia, M. (2026). Governing AI data-centre resources before consumption: the ZERO framework. Carbon Neutral Systems 2, 15. doi.org/10.1007/s44438-026-00037-z
  19. Gaffney, O., Luers, A., Carrero-Martinez, F. et al. (2025). The Earth alignment principle for artificial intelligence. Nature Sustainability. doi.org/10.1038/s41893-025-01536-6
  20. Google (2026). 2026 Environmental Report, covering 2025. Greenhouse gas emissions, scopes 1 to 3 (market-based), 18,847,100 tCO₂e, against 10,354,500 in 2019 (environmental metrics data table, p. 90); ambition-based 14,473,100 tCO₂e, 81% above 2019 (p. 74). Water use by data centres and offices (data-centre consumption 10,523 million gallons). sustainability.google/google-2026-environmental-report/
  21. Microsoft (2026). 2026 Environmental Sustainability Report and Environmental Data Fact Sheet, covering fiscal year 2025 (July 2024 to June 2025). Scopes 1 to 3, market-based: 21,121,000 tCO₂e, against 13,061,000 in fiscal 2020 (Data Fact Sheet, Table 1A, p. 3); 20,290,000 tCO₂e with management's criteria, 25% above fiscal 2024 (Table 1B, p. 4; report pp. 6 and 30). cdn-dynmedia-1.microsoft.com/is/content/microsoftcorp/microsoft/msc/documents/presentations/CSR/2026-Microsoft-Environmental-Data-Fact-Sheet-PDF.pdf
  22. Amazon (2026). 2025 Amazon Sustainability Report. Carbon footprint 80.85 MtCO₂e in 2025, market-based, up 16% from 69.55 MtCO₂e in 2024 (goals table p. 5; data table p. 46). sustainability.aboutamazon.com/2025-report
  23. Meta (2025). 2025 Sustainability Report and Environmental Data Index, covering 2024. Total emissions, market-based, 8,200,595 tCO₂e, against 5,830,244 in 2021; location-based 15,627,509 tCO₂e (Data Index, section 1.1). sustainability.atmeta.com/wp-content/uploads/2025/10/Meta_2025-Environmental-Data-Index.pdf
  24. International Energy Agency (2025). Energy and AI. Paris: IEA. Data-centre electricity 2024 and the 2030 base case. www.iea.org/reports/energy-and-ai
  25. Gabbatiss, J. (2025). AI: five charts that put data-centre energy use and emissions into context. Carbon Brief, 15 September 2025 (AI share of data-centre power, citing a report prepared for the IEA). www.carbonbrief.org/ai-five-charts-that-put-data-centre-energy-use-and-emissions-into-context/
  26. EPRI (2024). Powering Intelligence: analyzing artificial intelligence and data center energy consumption. Palo Alto: Electric Power Research Institute. US data-centre load 2023 and four 2030 scenarios. www.epri.com/research/products/3002028905
  27. Meta (2023, 2024). Llama 2 (Touvron et al. 2023, arXiv 2307.09288, Table 2) and Llama 3 model card: reported training emissions. github.com/meta-llama/llama3/blob/main/MODEL_CARD.md
  28. Gemma Team, Google DeepMind (2024). Gemma 2: improving open language models at a practical size. arXiv 2408.00118 (pretraining carbon footprint). arxiv.org/abs/2408.00118
  29. Chowdhery, A. et al. (2022). PaLM: scaling language modeling with Pathways. arXiv 2204.02311 (PaLM 540B training, 271.43 tCO₂e). arxiv.org/abs/2204.02311
  30. Patterson, D., Gonzalez, J., Le, Q., Liang, C., Munguia, L.-M., Rothchild, D., So, D., Texier, M. & Dean, J. (2021). Carbon emissions and large neural network training. arXiv 2104.10350 (GPT-3). arxiv.org/abs/2104.10350
  31. Patterson, D., Gonzalez, J., Hölzle, U., Le, Q., Liang, C., Munguia, L.-M., Rothchild, D., So, D., Texier, M. & Dean, J. (2022). The carbon footprint of machine learning training will plateau, then shrink. Computer 55(7), 18–28 (about 3/5 of Google's ML energy is inference, 2019–2021). doi.org/10.1109/MC.2022.3148714
  32. Luccioni, A. S., Viguier, S. & Ligozat, A.-L. (2023). Estimating the carbon footprint of BLOOM, a 176B parameter language model. Journal of Machine Learning Research 24(253), 1–15 (BLOOM; Gopher in Table 4). jmlr.org/papers/v24/23-0069.html
  33. Li, P., Yang, J., Islam, M. A. & Ren, S. (2025). Making AI less "thirsty". Communications of the ACM 68(7), 54–61. doi.org/10.1145/3724499
  34. Microsoft (2024). 2024 Environmental Sustainability Report, covering fiscal year 2023. Emissions 15.4 MtCO₂e, 29.1% above 2020; water consumption. www.microsoft.com/en-us/corporate-responsibility/sustainability/report
  35. TSMC (2024). ESG Data Hub, SASB index, 2023. Total water withdrawn 113,610 thousand m³. esg.tsmc.com/en-US/ESG-data-hub/framework-disclosure/sasb
  36. Company announcements: Talen Energy and Amazon, expanded Susquehanna PPA, up to 1,920 MW through 2042 (11 June 2025); Constellation and Meta, Clinton, 1,121 MW, 20 years (3 June 2025); Constellation and Microsoft, Crane (Three Mile Island Unit 1), 835 MW, 20 years (20 September 2024); Google and Kairos Power, 500 MW of small modular reactors by 2035 (14 October 2024). www.powermag.com/talen-amazon-launch-18b-nuclear-ppa-a-grid-connected-ipp-model-for-the-data-center-era/
  37. NERC (2026). 2025 Long-Term Reliability Assessment, released 29 January 2026. Summer peak demand forecast to grow 224 GW over ten years, most of it from new data centres. www.nerc.com/pa/RAPA/ra/Pages/default.aspx
  38. US Geological Survey (2024). Mineral Commodity Summaries 2024: gallium (China 98% of primary low-purity production), cobalt (DR Congo 74% of mine production), rare earths (China 240,000 of 350,000 t mined), 2023 estimates. doi.org/10.3133/mcs2024
  39. Baldé, C. P. et al. (2024). The Global E-waste Monitor 2024. ITU and UNITAR. 22.3% of 2022 e-waste formally collected and recycled; e-waste recycling meets about 1% of rare earth demand. ewastemonitor.info/the-global-e-waste-monitor-2024/
Back to top