@simon I've found the self-reported Llama2 training (!) emissions interesting. See section 2.2.1 here:

https://arxiv.org/abs/2307.09288

That doesn't say much about inference, tho, but it's a start.

You'd also have to include not just GPU energy but e.g. emissions of the data center and supporting software around it, cluster utilization, the emissions of people working on this, traveling, and so on.

There's Scope1/2/3 etc. categories to make this a bit more structured but in the end it's tough.