Methodology
How we approach environmental intelligence for the AI ecosystem.
Our approach
Environmental data on the AI industry is fragmented, inconsistent, and frequently missing. Our methodology is designed to work with this reality, not to paper over it with false precision.
For each entity we profile, we gather evidence from every available source, assess the quality of that evidence, and produce estimates with explicit confidence tiers. Where multiple sources disagree, we document the disagreement. Where no evidence exists, we say so.
All estimates are reviewed on a scheduled basis as new disclosures, regulatory filings, and research are published. The date of last review is recorded for every data point.
Data sources
Confidence tiers
Every figure we publish carries one of the following confidence designations.
Measured & Verified
Direct measurement by an accredited third party, or government-reported metered data. The highest confidence level.
Disclosed (Unverified)
Figure reported by the entity itself in a sustainability report, regulatory filing, or press release, but not independently verified.
Estimated (Documented Method)
Our estimate, derived using a documented and reproducible methodology based on proxy data. The method is published alongside the figure.
Estimated (Uncertain)
Our estimate, but with significant uncertainty in the inputs or the method. Wide confidence intervals apply.
Unknown / Modeled
No direct or proxy data available. Figure derived from sector-level models or analogous cases. Treat with substantial caution.
Not Disclosed
The entity has not published this information. Non-disclosure is recorded and treated as a signal, not a blank. It affects our overall disclosure quality assessment.
Estimation methods
When an entity does not disclose a figure, we estimate it using a documented cascade of public-data methods, tried in order of reliability. The first method with sufficient inputs is used, and the resulting figure carries the confidence tier shown below.
Known data center capacity
Energy modeled from disclosed data center capacity (GW), applying hyperscale PUE (1.10) and an 85% utilization assumption (IEA, 2025).
Query-volume model
Energy modeled from reported query or request volume and per-query energy intensity.
Data center count × average MW
Energy modeled from the number of facilities and an average per-facility capacity.
Capex proxy
Capacity inferred from capital expenditure at roughly $5M per MW (McKinsey, 2025), then converted to energy.
Revenue ratio
Energy inferred from revenue against sector energy-per-dollar ratios. The least reliable method, used only as a last resort.
Energy modeling applies a power usage effectiveness (PUE) matched to the entity's category: 1.10 for hyperscale operators, 1.44 for colocation, 1.30 for semiconductor fabrication, and a 1.58 global average otherwise (Uptime Institute, 2025), with an 85% utilization assumption (IEA Energy and AI, 2025).
From disclosed or modeled energy we derive the remaining metrics. Carbon uses a grid carbon intensity of 0.366 kg CO₂e/kWh (US EIA, 2023), with per-entity overrides where the operating grid is known (for example 0.495 for Taiwan, 0.42 for South Korea, 0.40 for TVA-supplied sites); Scope 1 and Scope 3 are estimated with per-category ratios calibrated against disclosed FY2023 figures. Water applies a usage effectiveness of 1.5 L/kWh for hyperscale operators and a 1.9 L/kWh industry average (The Green Grid / EESI, 2025). Continuous property-line noise is modeled from cooling load, anchored to roughly 55 dBA for a typical 30 MW campus and bounded between 45 and 75 dBA; it is always an estimate, never a company disclosure. A 100% renewable-matching claim does not mean zero emissions: we apply a 5% residual for timing mismatch. Every constant is cited at the point of use in our open estimation engine, which ships with a calibration self-test asserting model outputs against disclosed real-world figures.
Extracted figures are sanity-checked before publication. A reported energy value that deviates more than 3× from the capacity-based expectation, or a renewable share that disagrees with the curated company-wide figure by more than 40 points, is discarded and replaced with a documented estimate — a wrongly extracted value is never published at disclosed-tier confidence.
Disclosure quality framework (T1–T6)
Separately from the per-figure confidence tiers above, we audit each entity against a fixed checklist of disclosure requirements grouped into six categories. Every check is scored full, partial, none, or contradicted, and checks are referenced by code (e.g. T2.1 for Scope 1 emissions).
Energy
Total electricity consumption and AI-specific energy use disclosed and broken out from general operations.
Carbon
Scope 1, Scope 2 (location- and market-based), and Scope 3 greenhouse gas emissions disclosed.
Water
Total water withdrawal and water usage effectiveness (WUE) at data center facilities disclosed.
Land
Physical footprint of campuses and land-use change disclosed. Coverage of this category is still being built out.
Supply chain
Semiconductor manufacturing impacts and the methodology used to attribute environmental cost to customers disclosed.
Methodology
GHG accounting methodology, boundaries, and assumptions documented so reported figures can be interpreted and reproduced.
Transparency grade
Each entity's checklist results roll up into a single disclosure-quality score out of 100, which maps to an overall transparency grade from T-1 (most transparent) to T-5. The grade is shown on the Data page and each entity profile.
Score 80–100: comprehensive disclosure across categories.
Score 60–79: substantial disclosure with some gaps.
Score 40–59: partial disclosure; major categories missing.
Score 20–39: minimal disclosure.
Score 0–19: little or no meaningful disclosure.
Why “Not Disclosed” matters
When an entity does not publish a figure, that fact is not neutral. It tells us something about how the entity relates to environmental accountability. Consistent non-disclosure across multiple metrics is itself a meaningful data point, one that we record and include in our overall assessment of disclosure quality.
This approach is grounded in the literature on voluntary disclosure: companies that disclose unfavorable information tend to do so selectively, while companies that decline to disclose at all often have more to hide than those that disclose imperfect data. We do not treat silence as evidence of good performance.
What we do not claim
We do not claim to produce definitive figures. Environmental accounting in the AI sector is genuinely hard, and anyone who presents precise numbers without wide uncertainty bounds is likely overstating their confidence.
We do not have inside access to any company's infrastructure. Our estimates for undisclosed figures rely on public information, proxy data, and documented modeling assumptions. They can be wrong.
We do not update in real time. Our profiles reflect the state of evidence at a specific review date. The AI industry changes quickly; a figure that was accurate six months ago may no longer reflect current operations.