The GHG Protocol is the accounting standard behind almost every emissions report, developed by the World Resources Institute and the World Business Council for Sustainable Development. It defines the three scopes and the fifteen Scope 3 categories. Under it, AI inference a company buys is Scope 3, Category 1, and it can be measured per token with a stated uncertainty.
Almost every emissions figure a company publishes traces back to the GHG Protocol. It is not a law, but the laws, the investor frameworks and the auditors all use it as the common reference, so a figure that does not follow it cannot be compared with one that does. This guide covers what the protocol is, how it sorts emissions, and how a company’s AI usage is reported under it.
What it is
The GHG Protocol is a partnership between the World Resources Institute and the World Business Council for Sustainable Development that has published emissions accounting standards since 2001. The Corporate Standard sets out how a company draws its boundary and sorts emissions into three scopes. The Scope 2 Guidance covers purchased electricity, with a location-based method that uses the grid average and a market-based method that uses the contracts a company holds. The Corporate Value Chain Standard, usually called the Scope 3 Standard, covers the fifteen categories of indirect emissions in the value chain, and its Calculation Guidance walks through each one. Other standards cover products, projects, policies, mitigation goals and cities.
The rules that require emissions reporting adopt it rather than replace it. The EU’s European Sustainability Reporting Standards are built on it, California’s SB 253 requires reporting in conformance with it, and the investor frameworks reference it. So whether the protocol itself is mandatory matters less than it sounds: the reporting that is mandatory is written in its terms.
The three scopes, and where AI sits
| Scope | What it covers | Where AI usage goes |
|---|---|---|
| Scope 1 | Direct emissions from sources the company owns or controls | Nothing, unless the company runs its own generators |
| Scope 2 | Purchased electricity, heat and steam | The provider’s data centre electricity, in the provider’s accounts; a company’s own servers, in its own |
| Scope 3 | Everything else in the value chain, fifteen categories | Purchased inference, Category 1, in the buyer’s accounts |
The split matters because the same kilowatt-hour appears in two places with two owners. For OpenAI, Anthropic or Google, the electricity behind a query is Scope 2 and the hardware and training behind it are Scope 3. For the company that sent the query, all of it is a purchased service in Category 1. Both are correct, neither is double counting, and a defensible AI line says which side it is reporting from.
How the methodology maps onto it
CrbonFree’s methodology v1.2 adopts the corporate framework published by Bistline and colleagues and written up by Watershed, and reports three layers that line up with the protocol’s boundaries.
| Layer | What it counts | Protocol boundary |
|---|---|---|
| Active energy | Joules per token on the accelerator, through PUE and grid intensity | Part of the provider’s Scope 2 |
| Facility operation | Host servers, idle capacity and cooling, at 1.18 times host power and 30% utilisation | The provider’s full Scope 2 |
| Full lifecycle | Embodied hardware at 0.020 and training at 0.008 kg per million tokens | Adds the provider’s Scope 3 |
The buyer reports the full lifecycle figure in Category 1, and keeps the two inner layers available because an auditor may ask for the operational boundary on its own. Every figure carries a plus or minus 28.3% band, from a 20% activity uncertainty and a 20% factor uncertainty combined by root sum of squares, which is the IPCC’s Tier 1 method and the level of rigour the protocol’s own guidance on uncertainty asks for.
The calculation methods, ranked
For Category 1 the protocol’s guidance lists four methods: supplier-specific, hybrid, average-data and spend-based, in descending order of data quality. Most purchased services are reported spend-based because spend is the only number available. AI usage can be reported at the average-data level on measured activity, tokens per model from the provider’s own usage API multiplied by published per-token factors, and at the supplier-specific level wherever a provider publishes its fleet’s power usage effectiveness and grid intensity, as Google has. The methodology uses the provider’s figures for Google’s tiers and published averages for the rest, and says which.
Two things the protocol insists on that AI reporting often skips
The first is a stated boundary. The protocol asks a company to say what is in and what is out, and AI figures in the wild rarely do, which is why two published numbers for the same query can differ by ten times. The three layers exist so that the boundary is explicit.
The second is the separation of measurement from action. The protocol treats credits and their retirement as something reported alongside an inventory, never subtracted from it. CrbonFree keeps the same rule: the measured footprint is reported unchanged, and credits retired against it appear beside it with their serial numbers and a signed receipt. An inventory that nets the two together does not meet the protocol.
Reporting your AI line, in practice
- Meter tokens per model and per provider from the provider’s usage API, the CLI for coding agents and the browser extension for chat.
- Apply the per-tier factors and the three layers, and record the factor version on the period.
- Report the full lifecycle figure in Scope 3, Category 1, with the boundary, the version and the uncertainty band stated.
- Report any credit retirements alongside it, not against it.
The methodology page has every factor and formula, the case study shows a year of real gateway traffic worked through this way, and the sample account opens without a login.
Sources
- 01GHG Protocol, Standards
The family of standards and guidance documents.
- 02GHG Protocol, Corporate Value Chain (Scope 3) Accounting and Reporting Standard
The three scopes and the fifteen Scope 3 categories.
- 03GHG Protocol, Scope 2 Guidance
Location-based and market-based accounting for purchased electricity.
- 04Bistline et al., Estimating GHG Emissions from AI Use: Framework for Corporate-Level Measurement (2026)
The corporate framework for AI usage that CrbonFree’s methodology v1.2 adopts, with an operational Scope 2 boundary and a lifecycle Scope 3 boundary.
- 05Watershed, An Open Framework for AI Emissions Measurement
Watershed’s write-up of the same framework.
- 06IPCC 2006 Guidelines, Volume 1, Chapter 3: Uncertainties
Tier 1 error propagation, the method behind the plus or minus 28.3% band.
- 07California Air Resources Board, Corporate greenhouse gas reporting (SB 253)
Reporting in conformance with the GHG Protocol, Scope 3 from 2027.
Figures attributed to CrbonFree come from methodology v1.2, published in full with every factor and formula. Read the methodology.
About the author
Cory Bergh
Cory leads Crbon Labs, which originates its own climate projects and builds CrbonFree, the platform that measures the footprint of AI usage per token. He was previously VP of Technology and Innovation in the energy industry and holds a BComm, an MBA and the CFA.