Investigation I · Data Centres

Methodology &
Data Sources

How we built the data, what sources we used, what decisions we made, and where the limitations lie.

All data in this investigation is drawn from publicly available, independently verified sources. No proprietary databases were used. Where multiple sources existed for the same data point, we prioritised primary institutional sources (IEA, Lawrence Berkeley National Laboratory, EPA) over aggregators, and flagged disagreements in the text.

This page documents every major data decision, source, limitation, and update. We believe methodological transparency is as important as the findings themselves. If you identify an error, a better source, or a more recent data point, please contact us — corrections will be listed in the changelog below.

01 · Primary Sources

Where the Data Comes From

Energy — Primary
Energy and AI Report
International Energy Agency (IEA) · April 2025

The primary source for all global energy consumption figures, regional breakdowns, and 2030 projections. The IEA built a new global model and comprehensive dataset of data centre electricity demand, enriched by direct consultation with policymakers, the tech sector and energy industry. The 415 TWh (2024) and 945 TWh (2030 base case) headline figures come directly from this report.

iea.org/reports/energy-and-ai
Energy — US Detail
2024 US Data Center Energy Usage Report
Lawrence Berkeley National Laboratory · December 2024

The most comprehensive national-level audit of US data centre energy use, produced by the US Department of Energy's research arm. Used for US-specific figures, historical actuals, and the 2023 direct water consumption estimate of 17.5 billion gallons. This report forms the empirical basis for many IEA US figures.

eta-publications.lbl.gov
Water
Data Centers and Water Consumption
Environmental and Energy Study Institute (EESI) · 2024

Used for the 449 million gallons/day (163.7 billion gallons/year) US consumption estimate — based on 2021 EPA data covering 5,426 US data centres. Also sourced the Northern Virginia 2 billion gallon figure (2023, up 63% from 2019) and the statistic that only 51% of operators track water usage.

eesi.org/articles/view/data-centers-and-water-consumption
Carbon Accounting Gap
Data Centre Emissions Location-Based Analysis
The Guardian / Uptime Institute · 2024

Source for the finding that actual emissions from Google, Microsoft, Meta and Apple data centres were approximately 7.62 times higher than officially reported (2020–2022), based on location-based vs. market-based scope 2 accounting. The Uptime Institute's Jay Dietrich provided the methodological basis: location-based accounting reflects the actual carbon intensity of grids at point of consumption.

theguardian.com / uptimeinstitute.com
Carbon — Context & Charts
Five Charts on Data Centre Energy and Emissions
Carbon Brief · September 2025

Used for contextualising IEA projections, the Dublin 79% city electricity share (citing Oeko-Institute analysis), Ireland's national 21% share, and the US data centre share of national electricity (4% in 2023, rising to 7–12% by 2028). Carbon Brief draws directly on IEA primary data with clear methodology notes.

carbonbrief.org
Corporate Disclosures
Big Tech 2024–25 Sustainability Reports
Google · Microsoft · Amazon · Meta

Google Environmental Report 2024 (6.1bn gallons water; 32.2m MWh electricity; 13% emissions increase YoY 2023). Microsoft Sustainability Report 2024 (29.8m MWh; 23.4% emissions increase vs 2020 baseline; location-based scope 2 nearly doubled 2020–2024). Amazon Sustainability Report 2024 (68.25 Mt total corp. emissions, +6% YoY). All reports filed with CDP.

sustainability.google · microsoft.com/en/sustainability · sustainability.aboutamazon.com
Water — Spain / Europe
The Cloud is Drying our Rivers
EthicalGEO · July 2025

Source for the Amazon Aragon data: 500 million litres consumed annually at Europe's largest data centre; December 2024 application to increase water permit by 48%; Aragon's March 2025 EU drought aid application; Tu Nube Seca Mi Río campaign. Cross-referenced with France 24 reporting (2025).

ethicalgeo.org
Regulation
Sierra Club AI Data Centre Investigation
Sierra Club · 2024

Source for regulatory landscape: 17 Virginia bills defeated in 2024 session; Senator Markey's AI Environmental Impacts Act stuck in committee; Maryland diesel generator exemption. Also used for context on industry lobbying and the absence of federal oversight mechanisms.

sierraclub.org
Reporting Analysis
Not Greenwashing, But… Big Tech 2025 Sustainability Reports
Policy Review Institute · 2025

Source for Microsoft's location-based scope 2 emissions rising from 4.3m to ~10m metric tonnes (2020–2024); Google's 27% increase in data centre electricity use 2023–24; analysis of enabled emissions not reported; REC-based vs. location-based divergence in detail.

policyreview.info
02 · Data Decisions

Key Choices and Why We Made Them

Decision Choice Made Rationale Alternative Considered
Headline energy figure 415 TWh (IEA, 2024) The IEA figure is the most comprehensive global estimate, built from a new model and direct industry consultation. It is the most cited in peer analysis and policy. Energy Intelligence Group estimate of 460 TWh (2022); Goldman Sachs 160% growth projection. Both were referenced as context but not used as primary.
AI server energy breakdown AI workloads approximated using IEA's accelerated server category + Goldman Sachs / SemiAnalysis research No single source disaggregates AI from non-AI demand with full confidence. The IEA estimates AI accounted for 15% of total data centre energy in 2024; we used this as a floor. Some researchers (Alex de Vries, VU Amsterdam) argue the IEA AI figure is an underestimate. We noted this dispute in the text.
Emissions accounting method Dual presentation: market-based (as reported) + location-based (estimated) Market-based figures are what companies publish. Location-based figures better reflect physical reality. Presenting both makes the gap visible and allows readers to assess it. Using market-based only would replicate corporate PR. Using location-based only would require significant estimation. Dual presentation is most transparent.
Water consumption EESI/EPA 2021 US figure (449m gal/day); LBL 2024 direct figure (~17.5bn gal/yr); company disclosures The EESI/EPA figure covers the full US data centre fleet. The LBL figure covers direct consumption only (not indirect from power generation). Company figures are direct on-site only. We used all three with clear labelling of scope. IEEE Spectrum / LBL note that indirect water use (from electricity generation) typically exceeds direct on-site use by 4–5×. We referenced this in text but did not present a combined figure due to high uncertainty.
Map cluster positions Approximate centroid positions based on known cluster locations Precise operator-by-operator mapping would require proprietary data. Cluster centroids reflect documented concentrations (N. Virginia, Dublin, Singapore etc.) from IEA, Carbon Brief, and Oeko-Institute analysis. MSCI GeoSpatial data covers 13,558 individual assets but is not publicly available. We used cluster-level representation as the transparent alternative.
Carbon trajectory projections IEA base case (180 Mt 2024 → 300 Mt 2035) and lift-off scenario (→ 500 Mt) The IEA base case is the most credible central scenario. The lift-off case represents higher AI adoption. Morgan Stanley's 2.5 billion Mt cumulative figure to 2030 was referenced in text. We did not use third-party growth models from banks or consultancies as primary chart data, treating them as contextual references only.
03 · Limitations

What This Investigation Cannot Tell You

Limitation

Disclosure opacity. The majority of data centre operators do not publicly disclose energy or water consumption. The IEA's global model is built from a sample covering ~50% of global data centres (Malmodin et al. 2024) with geographic extrapolation. Uncertainty ranges are substantial — the IEA's own review acknowledges this explicitly.

Limitation

AI workload isolation. There is no standardised methodology for separating AI compute energy from general cloud workloads. The IEA's 15% figure for AI's share of data centre energy in 2024 is an estimate with wide uncertainty bands. Some researchers believe it significantly understates AI's true footprint.

Limitation

Location-based multipliers. The 7.62× emissions multiplier applied in our carbon accounting comparison is based on 2020–22 data from Guardian / Uptime Institute analysis. It is an average across four companies and will vary significantly by geography, year, and the specific grid mix at point of consumption.

Limitation

Water scope differences. Direct and indirect water consumption figures are not directly comparable across sources. Company disclosures report on-site direct use only. The LBL figure covers direct US consumption only. The 449m gallon/day EESI figure includes indirect power plant water. We have labelled each figure's scope clearly.

Context

Projections are uncertain. The IEA base case is a central scenario, not a forecast. The IEA explicitly notes "substantial uncertainty" in both 2024 actuals and forward projections, particularly regarding the pace of AI adoption and efficiency improvements. We present projections as projections, not predictions.

Context

Enabled emissions not counted. Neither we nor any major company disclosure includes "enabled emissions" — the downstream carbon produced by systems running AI (e.g., AI used in oil exploration, logistics optimisation). Policy Review Institute (2025) argues these should be included in scope 3 but they are currently absent from all reporting frameworks.

04 · Chart Notes

Specific Decisions Behind Each Visualisation

Energy Growth Chart (01): The stacked bar chart disaggregates total data centre electricity into conventional compute (including cooling and IT infrastructure overhead) and AI/accelerated server workloads. The AI bar from 2015–2024 is derived from IEA's estimate of AI's share of total consumption (15% in 2024, growing from near-zero in 2015) combined with LBL historical actuals. The dotted line shows the IEA base case total, which should match the stacked bars — minor discrepancies reflect rounding. All values 2025–2030 are projections.

Regional Comparison Charts (02): 2024 actuals are directly from IEA Energy & AI (April 2025): US 180 TWh, China 102 TWh, Europe ~62 TWh. "Europe" covers EU-27 + UK + Norway, as used by IEA. "Rest of World" is the residual from 415 TWh total. 2030 projections apply the IEA regional growth percentages (US +130%, China +170%, Europe +70%, Japan +80%) to 2024 actuals.

Water Charts (04): The US total water use trend line is constructed from EESI/EPA's 2021 figure (163.7bn gallons/year, the most comprehensive available) and LBL 2024's 17.5bn gallons direct figure for 2023. These measure different things: the EESI figure appears to include indirect water use from power generation; LBL is direct on-site only. The trend line is therefore indicative rather than a single consistent series — we have labelled this in the chart caption. Company water figures are direct on-site from sustainability reports.

Carbon Pledges Chart (05): The IEA base case and lift-off lines come directly from IEA Energy & AI (2025): 180 Mt in 2024, ~300 Mt by 2035 base case, ~500 Mt lift-off. The net-zero pledge trajectory is a composite constructed from company target dates: most commit to net-zero or carbon neutrality by 2030 or 2040. The location-based reality line applies the Guardian/Uptime Institute 7.62× multiplier to the IEA base case and projects forward — this is an estimation, not a primary data series, and is displayed on a separate right-hand axis with a different line style to distinguish it clearly.

Cluster Map (03): Cluster positions are approximate centroids of known data centre concentrations based on: IEA regional narrative; Carbon Brief/Oeko-Institute analysis; Uptime Institute and CBRE market reports. Bubble sizes are proportional to approximate annual energy consumption as estimated from regional allocations. They are illustrative of relative scale, not precise measurements. The animated pulse effect is aesthetic only and does not encode data.

05 · Changelog

Corrections and Updates

March 2025
Initial publication. All data sourced as described above.
Pending
Will update water consumption figures when LBL publishes 2024 actuals (expected Q4 2025). Will update company emissions table when FY2024 sustainability reports are finalised (expected Q2–Q3 2025).
Ongoing
IEA data portal (iea.org/data-and-statistics) is monitored for updated regional consumption figures. Carbon Brief's live tracker and Corporate Knights annual disclosure analysis will be used for emissions updates.

Read the Full Investigation

Return to the data visualisations, charts and analysis in the main piece — The Invisible Furnaces.

Read Investigation →

Reset your
password.

Enter the email address on your account. We'll send you a reset link valid for 24 hours.