Corporate greenhouse gas inventories rarely account for the emissions from purchased AI inference, even though it falls under Scope 3 Category 1. When companies do try to include it, they often rely on a spend-based factor—multiplying the dollar amount spent on AI by an emissions per dollar figure. That approach, according to a new arXiv paper, can be wildly inaccurate.
The paper proposes a token-based method instead. Rather than using money as a proxy, it estimates emissions from the physical quantity of AI inference—measured in tokens, the basic units of text or data that a model processes. This ties the carbon footprint to actual compute usage, which the authors argue is a more faithful representation of the energy consumed.
The discrepancy is significant: the spend-based factor yields estimates 1.7 to 12.3 times higher than the physical estimate. That means companies using spend-based accounting could be over-reporting their Scope 3 emissions by a wide margin, or, if they omit AI entirely, under-reporting them. The paper's method offers a way to bring purchased AI inference into inventories with greater precision.