Tailpipe

Celebrating the Release of the Software Carbon Intensity for AI Specification!

15/01/2025

Did you know that a short conversation with GPT-4 generates 3 grams of CO₂e¹? With 2.5 billion messages sent to ChatGPT every day², it has never been more important to accurately account for the carbon impact of AI. Fortunately, in a milestone for AI sustainability, the Green Software Foundation has just released the Software Carbon Intensity for AI (SCI for AI) Specification!

The SCI for AI Specification provides a ratified framework for organizations to measure the carbon footprint of AI usage. Built out of the Software Carbon Intensity ISO specification, it addresses the specific lifecycle stages of AI and the distinct measurement metrics required for different model types, covering machine learning, LLMs, GenAI, Agentic AI, document analysis, translation, and speech recognition.

Importantly, it provides distinct methodologies for consumers and providers of AI services, apportioning responsibility to the following:

Lifecycle Stage		Responsibility
Inception	→	AI Service Provider
Design & Development	→	AI Service Provider
Deployment	→	AI Service Provider
Operation & Monitoring	→	Consumer
Retirement	→	Consumer & AI Service Provider

In short, the specification lays out how emissions should be calculated as:

(Operational Carbon + Embodied Carbon) / Functional Units

Where:

Operational Carbon is emitted during the usage of the hardware within the consumer or provider boundary.

Embodied Carbon is emitted during the manufacture, transport, and disposal of the hardware within the consumer or provider boundary.

Functional Units are specific to the model type, for example, tokens, images, inferences, FLOPs, or parameters.

For the full specification, visit the SCI for AI GitHub.

Many businesses are facing regulatory pressures to report their emissions across their entire value chain – including AI usage. For the first time, they now have a framework to do so. Tailpipe is already developing its methodology for measuring cloud-hosted AI usage in line with the SCI for AI specification. To find out more about the impact of AI on cloud emissions, read our Insights!

¹ Ecologits, via HuggingFace, ‘impact of a small conversation with gpt-4′.

² OpenAI, 2025.