Scope 3 Emissions

Scope 3 is divided into upstream emissions (purchased goods and services, capital goods, fuel-and-energy-related activities, upstream transportation, waste, business travel, employee commuting, and upstream leased assets) and downstream emissions (downstream transportation, processing of sold products, use of sold products, end-of-life treatment, downstream leased assets, franchises, and investments).

• Upstream Example: Embodied carbon in GPUs and server racks purchased from hardware vendors.
• Downstream Example: Electricity consumed by end-users running inference on a deployed AI model.

This is typically the most significant category for AI companies. It captures the cradle-to-gate emissions of all products and services purchased during the reporting year.

• Hardware: Emissions from mining rare earth minerals, semiconductor fabrication, and assembly of servers.
• Cloud Services: Emissions from third-party data center construction and upstream energy infrastructure.
• Software & Consulting: Emissions from the operations of SaaS providers and professional services firms.

Calculation methods include supplier-specific data, hybrid models combining supplier and secondary data, and spend-based methods using environmentally-extended input-output (EEIO) databases.

For AI companies offering model-as-a-service or on-device inference, this category captures the energy consumed by customers when using the product. It requires estimating the total lifetime energy consumption of the deployed system.

• Direct Use-Phase Emissions: Electricity consumed by end-user devices during inference.
• Indirect Use-Phase: Energy consumed by networking infrastructure transmitting data to and from the model.

Calculation requires defining a functional unit—such as emissions per million inferences—and multiplying by projected usage volumes over the product's lifespan.

This category covers the upstream emissions from the production of long-lived assets like data center buildings, custom ASIC fabrication facilities, and server fleets. Unlike purchased goods, capital goods are depreciated over their useful life.

• Data Center Construction: Emissions from concrete, steel, and cooling infrastructure.
• Custom Silicon Development: Emissions from the research, design, and fabrication of proprietary AI accelerators.

Emissions are accounted for in the year of acquisition, not amortized, per the GHG Protocol.

This category captures emissions from the extraction, production, and transportation of fuels and energy consumed by the reporting company that are not already included in Scope 1 or Scope 2.

• Transmission & Distribution Losses: The electricity lost as heat when transmitting power from the generation source to the data center.
• Upstream Fuel Emissions: Emissions from extracting and refining natural gas used for on-site backup generators.

This category ensures that the full lifecycle of energy procurement is captured, not just the combustion or generation phase.

Scope 3 accounting is inherently uncertain due to reliance on secondary data and supplier estimates. The GHG Protocol defines a data quality hierarchy:

• Tier 1: Supplier-specific emissions data (highest quality, rarely available).
• Tier 2: Hybrid models combining supplier activity data with industry-average emission factors.
• Tier 3: Spend-based methods using EEIO databases (lowest quality, highest uncertainty).

For AI hardware, embodied carbon estimates can vary by 2-3x depending on the semiconductor foundry's energy mix, making supplier engagement critical for accurate reporting.