Model Lifecycle Assessment (LCA)

This stage accounts for the embodied carbon of the physical infrastructure required to develop and run the model. It includes the extraction of rare earth minerals, semiconductor fabrication, and assembly of GPUs, TPUs, and networking equipment.

• Key Metric: Embodied Carbon (kgCO2e)
• Scope: Scope 3, Category 2 (Capital Goods)
• Example: Manufacturing a single high-end GPU server can emit over 1,500 kgCO2e before it is ever powered on.
• Data Source: Hardware vendor Product Carbon Footprints (PCFs) and Life Cycle Inventory databases.

This phase quantifies the energy consumed by storing, cleaning, and transforming massive training datasets. It involves the continuous power draw of data lake infrastructure and the compute cycles used for ETL (Extract, Transform, Load) jobs.

• Key Metric: Joules per GB stored/processed
• Scope: Scope 2 (Energy) & Scope 3 (Cloud Services)
• Example: Storing 1 PB of data in a cloud region with a high grid carbon intensity can generate significant ongoing emissions before training begins.
• Optimization: Data deduplication and efficient columnar storage formats reduce this footprint.

Often the most visible phase, this captures the dynamic power consumption of compute clusters during pre-training, fine-tuning, and hyperparameter sweeps. It must account for both productive runs and failed experiments.

• Key Metric: Total kWh consumed, multiplied by the grid's Marginal Emissions Rate.
• Scope: Scope 2 (Power) & Scope 3 (Cloud/Colocation)
• Example: Training a large language model can consume over 1,000 MWh, equivalent to the annual energy use of 100 average US households.
• Reporting Tool: CodeCarbon or Cloud Carbon Footprint integrated into the MLOps pipeline.

This stage measures the operational energy cost of serving predictions. For high-traffic services, the cumulative energy of billions of inference calls often dwarfs the training footprint over the model's lifetime.

• Key Metric: Joules per Inference or gCO2eq per 1,000 API calls.
• Scope: Scope 2 (Operational Energy)
• Example: A generative AI search query can consume 10x the energy of a standard keyword search.
• Optimization: Quantization, model distillation, and carbon-aware scheduling to low-carbon regions.

The final stage accounts for the environmental impact of retiring hardware and deleting data. This includes e-waste processing, hazardous material management, and the potential for recycling precious metals from compute components.

• Key Metric: kg of e-waste generated and recycling rate (%)
• Scope: Scope 3, Category 5 (Waste generated in operations)
• Example: Secure data sanitization via cryptographic erasure has a negligible carbon cost, while physical destruction of drives increases the footprint.
• Circularity: Refurbishing and redeploying decommissioned GPUs to less latency-sensitive workloads extends hardware life.

A transversal impact category that spans all lifecycle stages. It quantifies the energy consumed by networking equipment (routers, switches, optical interconnects) when moving training data, model checkpoints, and inference payloads across data centers or to end-users.

• Key Metric: kWh per GB transferred.
• Scope: Scope 3, Category 4 (Upstream Transportation & Distribution)
• Example: Distributing a multi-terabyte open-source model to thousands of download mirrors globally has a non-trivial network backbone footprint.
• Optimization: Edge caching and efficient serialization formats reduce transfer volumes.