Product Carbon Footprint (PCF)

A rigorous PCF employs a cradle-to-grave assessment scope, capturing emissions from raw material extraction through manufacturing, distribution, use, and end-of-life disposal. For an AI server, this includes the embodied carbon of silicon fabrication and the operational emissions of data center electricity. This contrasts with cradle-to-gate assessments, which stop at the factory door and ignore downstream usage—a critical omission for energy-intensive software products.

The PCF is calculated relative to a defined functional unit, which quantifies the service provided by the product. This enables fair comparisons between solutions. In AI, functional units must be carefully scoped to avoid misleading efficiency claims:

• Per inference request: Useful for comparing API-based AI services.
• Per training run: Used for foundation model development.
• Per token generated: Increasingly standard for large language model (LLM) comparisons. Without a precise functional unit, carbon efficiency metrics become meaningless.

A significant challenge in software PCF is emission allocation in shared, multi-tenant cloud environments. A single physical server hosts multiple virtual machines or containers. The PCF must define a logical partitioning rule to attribute a fair share of the hardware's embodied and operational carbon to a specific software product. Common allocation methods include economic allocation (based on compute cost) and physical allocation (based on CPU-hours or vCPU share).

A credible PCF includes a data quality assessment and an unciquity analysis. Primary data (e.g., direct power meter readings from a data center) yields high accuracy, while secondary data (e.g., industry-average emission factors for GPU manufacturing) introduces uncertainty. For AI products, the most significant data gaps often lie in Scope 3 upstream emissions, particularly the fabrication of advanced logic chips, where supplier-specific data is proprietary and difficult to obtain.