Grid carbon intensity
gCO 2
e per kWh Dirtier
— 800
— 600
— 400
— 200
“ YOU DON’ T NEED PERFECT DATA TO CALCULATE AND REDUCE YOUR EMISSIONS”
Dr. John Bistline Head of Science Watershed
0 Cleaner
Sustainability measurement only works if the methods are shared and transparent. A customer should be able to understand why a number appears in its footprint and what would change it. Openness makes the analysis more useful, and criticism makes it better.
Q. WHAT’ S THE NEXT OPEN QUESTION IN AI EMISSIONS MEASUREMENT YOU WANT TO TACKLE?
» Attribution at the interaction level. Aggregate estimates are useful for inventories, but dynamic routing requires a different question: What emissions would one additional query cause at a particular place and time? That requires provider data on hardware and utilisation, tied to hourly grid conditions at the data-centre location.
The other one I’ d flag is embodied and training emissions. As the grid gets cleaner, the manufacturing footprint of chips, data centre construction and training a model becomes a larger share of the total, which are less well characterised than operational electricity.
Q. AFTER TWO DECADES IN CLIMATE AND ENERGY MODELLING, WHAT KEEPS YOU DOING THIS WORK?
» I like problems where the answer keeps changing because the system changes. Twenty years ago, the question was whether clean energy could compete on cost. Now that it can, the questions are about speed, transmission, unexpectedly rapid load growth. AI makes those questions more urgent.
Watershed also shortens the distance between analysis and action. Research we publish in the spring can change how a company chooses models or procures compute by the fall. Seeing the work change a real decision is what keeps it interesting for me.
sustainabilitymag. com 81