“We’ve heard from companies that they’re already being asked about AI emissions from investors, from auditors, from regulators, and right now most of them are guessing,” John Bistline, Watershed’s head of science, told me. “We wanted to give them something that was more defensible.”
For most frontier AI models, including those developed by OpenAI and Anthropic, there’s very little information to work with. Google is the only proprietary AI developer that has published a transparent estimate of its model’s operational energy use and related emissions. In a paper last August, researchers at the company found that “the median Gemini apps text prompt consumes 0.24 watts,” which is “less energy than watching nine seconds of television,” and released 0.03 grams of CO2-equivalent. These numbers may be out of date by now, however. In the paper, the authors note that this already represented a 33-fold reduction in energy consumption compared to the previous year.
That’s one challenge with estimating AI-related emissions — tech companies are both growing and innovating rapidly, expanding their energy footprints while also finding greater efficiencies, which may be one reason they don’t disclose this information yet.
Another obstacle is that the exercise involves making a number of carbon accounting decisions, and there’s no consensus yet on best practices. For instance, where do you draw the line on which emissions to include? You could just look at the energy required to operate the model, or you could include the energy used to train the model, or even the emissions related to fabricating and manufacturing the hardware it’s running on. Training a model tends to be more energy-intensive than running it to respond to queries, but it only happens once. If you’re going to include training emissions, the next question is, how should responsibility for those be allotted across the lifetime of the model and its use by hundreds of thousands of customers?
Another decision is how to account for differences in user behavior. A model’s energy intensity can vary widely depending on whether the user is asking a simple question, requesting complex research, generating images, or dispatching agents to conduct multiple tasks simultaneously. Models capable of “reasoning” use an estimated 30 times more electricity than those without that ability, according to research by HuggingFace, a company that creates tools for AI developers. A per-prompt emissions average would not capture these differences, and therefore would not give companies actionable information to help them reduce their emissions.
A “per token” average might be more useful in that sense. When AI models process queries, they break the sentence or code down into smaller components called tokens. One token might be just the first few letters of a word. When the model generates a response, it also processes it in terms of tokens. Estimating emissions per token is not a perfect system either, however, since a token’s value can vary across AI providers. Input tokens, i.e. user questions, also tend to be less energy-intensive than output tokens, or user responses, and a single per-token average will conceal that difference.
Then there’s the question of how to get from a model’s energy intensity to an emissions estimate. Should you use the real-world average carbon intensity of the electric grid? What about any clean energy agreements the AI company may have signed? And how should you factor in companies that decide to bypass the grid entirely and build their own on-site generation, which tends to use natural gas?
The Watershed paper proposes some answers to these questions, and also offers guidance for how companies can develop emissions estimates based on the data available to them.
While most of the published research on AI emissions to date has calculated energy intensity on a per-query basis, Watershed advocates for a per-token approach. — i.e. “kilowatt-hours per thousand tokens.” The authors reason that electricity use scales more directly with the number of tokens used than the number of queries submitted. AI application customers are also often billed based on their token usage, so there’s a business case for tracking tokens and trying to use them more efficiently.
For those companies working with essentially zero data — not even the number of tokens they’re using per year — Watershed recommends they approximate their AI emissions using a “spend-based” method. This means simply multiplying the amount they spend per year on AI services by an emissions factor of 0.134 kilograms of carbon dioxide equivalent per U.S. dollar, which is based on U.S. Bureau of Economic Analysis numbers for the data processing sector of the economy.
The Watershed paper concedes that whatever number this method spits out will be wrong, noting that it “can misestimate true AI emissions by several times in either direction,” and advising companies to treat this as a “provisional placeholder.” But publishing these numbers, even though they are wrong, could help push AI companies toward more transparency if they want to correct the record.
For companies that do track their token volumes, Watershed has a more rigorous solution. The paper proposes a formula companies can use to calculate their AI emissions, accounting not just for inference energy use, but also training emissions, embodied emissions of the data processing equipment, and a figure known as “power usage effectiveness.” This captures the energy consumed by cooling systems, power conversion, and other data center infrastructure that’s not directly serving AI processing. Since model-specific values for the various inputs to the formula are mostly not available today, Watershed has provided default values gathered from previous studies, including papers by Microsoft and Google. Companies can substitute the actual numbers disclosed by AI providers into the formula as that information becomes available.
I reached out to Google, Microsoft, Anthropic, and OpenAI to ask why they didn’t share token carbon intensity, and whether they planned to in the future. Only Microsoft responded to my inquiry, pointing me to its blog post and peer-reviewed paper estimating general AI energy use across frontier models.
To get the most accurate estimate, companies would also need to know where, geographically, their AI queries are being serviced, since emissions from the electric grid varies by region. In some cases, companies may be able to actually choose where their queries are being processed, offering another lever by which they could potentially reduce their emissions.
The right data, disclosed in sufficient detail, will unlock companies’ ability to reduce their AI-related emissions, Watershed argues. Employees would have more reason to choose the most appropriate model for a given task, for example, like avoiding using energy-intensive reasoning models for basic questions.
“I think about a John von Neumann test here,” Bistline said, referring to the mathematician and proto-computer scientist. “You wouldn’t ask an advanced model like Fable anything that you would be embarrassed to ask John von Neumann, or Marie Curie, right? You wouldn’t want to ask ‘how many R’s are there in Strawberry?’ or ‘which restaurants would you recommend I go to in Miami?’”
Of course, companies can already implement this recommendation today, but there will be no way to account for and prove that they are reducing their emissions as a result until AI providers disclose distinct model-based energy estimates.
As Bistline mentioned, this information isn’t just nice-to know — companies are already being asked for it. Upcoming regulations in California and the European Union will require large companies to disclose their total direct emissions, and will eventually require them to disclose indirect emissions like AI energy use. The EU’s AI Act will also require AI companies to disclose a breakdown of the energy consumption of its general purpose AI models.
“There are customer-side disclosure rules and provider-side ones developing in parallel,” Bistline said, “and right now there’s no agreed methodology connecting the two, which is the gap we’re trying to address with our AI emissions framework.”
Courtesy Polydrops
Courtesy Pebble
Courtesy Lightship
Courtesy Winnebago
EPA Administrator Lee Zeldin. Kevin Lamarque-Pool/Getty Images
