Startup Offers Framework for Estimating Corporate AI Emissions

Watershed, a startup focused on emissions tracking, has introduced a framework designed to help companies estimate the environmental impact of their artificial intelligence (AI) usage. This development addresses a significant challenge in corporate sustainability reporting, as the energy demands of AI data centers are rapidly increasing. Many companies currently struggle to accurately account for AI-related emissions, particularly when utilizing closed AI models that do not disclose their energy consumption data.

The proposed framework by Watershed incorporates several key components for calculating AI emissions. It considers the underlying data center infrastructure and establishes a functional unit of kilograms of CO2 per million AI tokens. The calculation then scales based on the total number of AI tokens a company processes. John Bistline, Watershed's head of science, explained to Fast Company that this approach leverages existing data tracking, as companies already monitor AI usage at the token level for cost management. He noted that the emissions calculation integrates seamlessly with this existing data, making it a practical addition rather than a completely new system, and highlighted that cost and sustainability efforts are interconnected.

The increasing scrutiny on corporate AI emissions stems from investors, auditors, and regulators. Companies are expected to report not only direct emissions but also indirect emissions, known as Scope 3 emissions. These indirect emissions include the energy consumed to power AI queries for employees. Some jurisdictions, such as California, already mandate Scope 3 emissions disclosures. Furthermore, the Greenhouse Gas Protocol, a global standard-setting body for corporate emissions, is considering introducing requirements for cloud and AI services. This growing demand for transparency means that companies are increasingly being asked to provide data on their AI-related carbon footprint, a request many are currently unable to fulfill comprehensively.