More and more companies are implementing artificial intelligence into both their internal workflows and their external products, and that AI use comes with an environmental impact. Already, AI data centers are driving a surge in electricity demand that is outpacing supply.
But that impact likely isn’t showing up on all corporate sustainability reports yet, because accounting for corporate AI emissions is a challenge—particularly when companies are using closed AI models that don’t disclose their energy use.
Watershed, a startup that helps companies track their emissions, is working on this challenge.
The startup recently published a framework for how companies can estimate their emissions from AI. It takes into account the data center infrastructure, a functional unit of kilograms of CO2 per million AI tokens, and calculations based on the number of AI tokens a company uses.
“Companies are already tracking AI usage at the token level for cost management,” John Bistline, Watershed’s head of science, tells Fast Company via email. “The emissions math plugs into that same data. So this isn’t asking companies to build something entirely new. Cost and sustainability go hand in hand here.”
When companies quantify their carbon footprints, they take into account not only direct emissions from their own energy use or products, but indirect emissions, like from the flights their employees take for business travel—or all the power needed to answer their workers’ AI queries. Those are called Scope 3 emissions.
In some cases, Scope 3 emissions disclosures are already required by law, like in California. The Greenhouse Gas Protocol, which sets corporate standards, is considering requirements around cloud and AI services.
And aside from those requirements, companies are already being asked about these numbers.
“Investors, auditors, and regulators are asking about AI emissions, and most companies don’t have a defensible way to answer,” Bistline says.
AI may be a small part of most companies’ footprints currently. “But nobody expects that to stay the case for long,” Bistline adds. “The companies that build their measurement infrastructure now will be better prepared than those who wait.”
By accounting for AI emissions, corporations will also be able to take steps to reduce both the emissions and their operating costs.
“The [Watershed] framework reports electricity alongside emissions specifically, so that measurement connects to concrete reduction levers: which model you use, which region serves your query, how you structure your prompts,” Bistline says. “Even with data gaps, these are all things companies can control in how they deploy and use AI.”
AI models can vary widely when it comes to energy use—a reasoning AI model may use about 30 times more energy than a smaller model for the same task, according to Watershed. “Region” also matters because different parts of the power grid are powered by different energy sources, which changes their carbon intensity.
Watershed’s framework only estimates the emissions from AI use. That’s because there’s no real way to precisely measure these emissions yet.
“Many of the most widely used AI models are closed, meaning you can’t independently test their energy consumption the way researchers can with open models,” Bistline says.
“The only empirical, published energy figure for a closed frontier model is Google’s Gemini data from mid-2025, and even that is a single data point for one model at one moment in time,” he adds.
Figuring out AI use emissions is also complex because of the research and development that goes into training these models. AI companies may not want to disclose the figures needed to do such calculations, either.
Those figures—concerning total training emissions and total lifetime tokens served—are “commercially sensitive,” Bistline says.
But even if AI providers won’t share those details, Watershed hopes they’ll share the ratio of emissions per token. (Tokens themselves are often a vague unit of measurement, adding to the challenge.)
That leaves an estimate of emissions as the best answer. As AI providers share more information, Bistline says, those estimates will get more precise.
And as AI providers share that info, it may show that their AI infrastructure is actually more efficient than the estimates assumed. That sort of disclosure, then, helps AI companies demonstrate their own efficiency gains as well.