WATERSHED
THE SAME AI PROMPT, VERY DIFFERENT CARBON
Grams of CO 2
e from one Gemini-class text prompt( 0.24 Wh), by US state grid. The energy is identical everywhere; only the grid changes. gCO 2 e per prompt Dirtier
– 0.200
– 0.175 – 0.150 – 0.125 – 0.100 – 0.075 – 0.050 – 0.025
Cleaner
Cleanest grid: Vermont 0.006g Dirtiest grid: West Virginia 0.212g about 34x difference
The good news is that the two choices that dominate are available to companies today. Sending a query to a large reasoning model instead of a smaller one can mean roughly 30 times the electricity for the same task. The second is where the model runs, which can change its carbon intensity by more than five times. AI emissions vary with deployment choices. They are not a fixed consequence of using AI. Procurement teams and developers can act now through model and region choices, even before providers disclose data to refine these estimates.
Q. WHO DID YOU BUILD THIS WITH AND HOW DID YOU STRESS-TEST IT?
» The framework came out of work with Dr. Steve Davis at Stanford and Dr. Sangwon Suh at Tsinghua, along with the Watershed team, including Shaena Ulissi, James Joyce, Mo Li and others. We consulted with companies managing these emissions day to day, including Block and Okta, and with the Business Council on Climate Change. We had formal review and feedback from industry, academia, and the standards community, which helped to make the framework scientifically sound and useable by sustainability teams.
78 October 2026