Greenhouse Gas (GHG) Protocol

Ensure the GHG inventory appropriately reflects the emissions of the company and serves the decision-making needs of users—both internal and external. For AI governance, this means selecting a boundary that captures the training compute, inference serving, and data center overhead that materially contribute to the organization's carbon footprint. The inventory must include all emission sources that are significant enough to influence stakeholder assessments and decisions.

Account for and report on all GHG emission sources and activities within the chosen inventory boundary. Disclose and justify any specific exclusions. For sustainable AI reporting, this requires tracking:

• Scope 1: On-premise GPU cluster natural gas backup generators
• Scope 2: Purchased electricity for cloud instances and colocation racks
• Scope 3: Embodied carbon in purchased servers, networking gear, and downstream customer usage of deployed models

Use consistent methodologies to allow for meaningful comparisons of emissions over time. Transparently document any changes to the data, inventory boundary, methods, or any other relevant factors in the time series. For AI workloads, this means applying the same emission factors, PUE values, and allocation methods year-over-year. If you switch from average grid carbon intensity to marginal emissions rates, you must recalculate the base year to maintain trend integrity.

Address all relevant issues in a factual and coherent manner, based on a clear audit trail. Disclose any relevant assumptions and make appropriate references to the accounting and calculation methodologies and data sources used. For model lifecycle assessments, this means publishing:

• The hardware specifications used for training
• The grid region and emission factor source
• The measurement tool (e.g., CodeCarbon, Cloud Carbon Footprint)
• Any exclusions (e.g., hyperparameter tuning experiments)

Ensure that the quantification of GHG emissions is systematically neither over nor under actual emissions, as far as can be judged, and that uncertainties are reduced as far as practicable. Achieve sufficient accuracy to enable users to make decisions with reasonable assurance as to the integrity of the reported information. For AI compute, prioritize direct measurement (e.g., server power draw via IPMI) over estimation (e.g., TDP-based modeling). When estimation is necessary, calibrate models against real-world joules per inference benchmarks.

Define which operations are included using either the equity share or control approach (financial or operational control). For cloud AI workloads, this determines whether emissions from a Platform-as-a-Service instance are Scope 2 (purchased electricity) or Scope 3 (upstream leased assets). The choice of consolidation approach must be applied consistently across the full inventory and clearly disclosed in the report methodology section.

Direct greenhouse gas emissions from sources owned or controlled by the reporting organization. In AI contexts, this primarily includes on-premise data center backup generators burning diesel or natural gas, and company-owned vehicle fleets. For most software-centric enterprises, Scope 1 is negligible compared to Scope 2 and 3. Reporting requires direct fuel metering and emission factor application.

Indirect emissions from the generation of purchased electricity, steam, heating, or cooling consumed by the organization. This is the dominant category for cloud-based AI workloads. The GHG Protocol mandates dual reporting using both location-based (average grid emission factor) and market-based (contractual instruments like PPAs) methods. Granular hourly matching is the emerging best practice.

All indirect emissions occurring in the reporting company's value chain, divided into 15 upstream and downstream categories. For AI, critical categories include:

• Category 1: Purchased goods and services (embodied carbon in GPUs)
• Category 2: Capital goods (server manufacturing)
• Category 11: Use of sold products (customer inference emissions) Scope 3 typically represents over 90% of total emissions for technology companies.