Assessing the carbon footprint of generative AI systems

This article introduces TokenFlop, a modeling methodology for assessing the environmental footprint of Generative Artificial Intelligence Systems (GAIS). It is based on best practices and existing standards, including the ISO 14040 LCA standard.


The goal is not to produce a certified measurement of emissions, but to provide reliable and comparable estimates that are useful for decision-making. Beyond simply reporting in a GHG Protocol report, this approach enables an organization to understand the materiality of its AI usage, place it within an environmental and societal context, and steer its choices toward more energy-efficient systems.

Adopting such a methodology offers tangible benefits

• A unified approach covering all use cases for generative AI—training, fine-tuning, and multimodal inference—based on the ISO 14040 LCA standard.
• A methodological framework for estimating GHG emissions and incorporating them into structured reporting, with full transparency regarding the assumptions used.
• The ability to identify growth trends in order to forecast future consumption and costs.
• A framework for establishing a carbon budget consistent with a pathway toward alignment with the Paris Agreement.

The theoretical modeling approach described below also makes it possible to:

Simulate the impact of scaling up and anticipate the sizing of production equipment.

Compare different solutions on a common basis, in order to guide decisions toward the most efficient options.

Assess in advance the impact of digital energy-saving policies before they are implemented.