Open-Source Tool Shows the Same AI Query Emits 8x More Carbon in Texas Than Sweden
A new open-source method shows the same AI request can emit 8.5x more carbon in Texas than in Sweden — the kind of open, auditable benchmark AIforE wants to see become standard.
Swedish AI infrastructure provider Berget AI, working with the Stockholm Environment Institute, WattTime, and Klimatkollen, this week released a free, MIT-licensed, open-source method for calculating the carbon emissions of individual AI inference requests — testable publicly at co2.berget.ai. Running the identical query on identical hardware, the method found emissions ranging from about 0.4 grams of CO2 in Sweden to 3.2 grams in Texas, an 8.5x gap driven almost entirely by each region's electricity grid mix. The method follows the Green Software Foundation's SCI for AI specification and accounts for hardware manufacturing emissions alongside operational energy, rather than relying on rough averages.
This is close to a model case for the kind of tooling AIforE wants to see become routine: a named, open specification that any provider can adopt, run, and publish results against, rather than an unverifiable marketing claim about "green AI." Request-level carbon accounting that factors in grid mix and hardware lifecycle, released under a permissive license so competitors and auditors alike can inspect it, is exactly the building block a future sustainable-AI certification scheme could require providers to report against. AIforE will be watching whether other inference providers follow Berget's lead and publish comparable, independently checkable numbers of their own.
Source: Stockholm Environment Institute
