In addition there is a need to look at the total cost including supporting functions. Dev test servers, build pipelines, redundancy, failover, deployment environments. Sometimes these may be 5% of emissions, but at other times much higher. And some languages will have significantly less emissions when idle.
A good example here may be, a machine learning model may have 90%+ of it's carbon emissions in the training, vs the operation.
A good example here may be, a machine learning model may have 90%+ of it's carbon emissions in the training, vs the operation.