Evaluating the Inclusion of LLM Training Emissions in SCI for API Consumers

[navveenb]

Large Language Models (LLMs) like GPT-4, Gemini, and Claude are widely consumed via APIs. Users interact with these systems by sending prompts and receiving outputs—without any visibility or control over how the model was trained, what infrastructure was used, or the carbon impact of the development lifecycle.
This raises a critical question within the context of the Software Carbon Intensity (SCI) standard:
Should the emissions generated during the training phase of an LLM be partially allocated to end-users who only access the model via API-based inference?
This document explores both sides of the argument: the case for excluding training emissions in standard SCI measurement for API consumers, and the case for including them via amortization. It presents conceptual approaches, identifies challenges, and ends with a recommendation framework for practical implementation
Evaluating the Inclusion of LLM Training Emissions in SCI for API Consumers.pdf