[chrismartin]

Hey Sam, does 0.34 watt-hours per query include an amortization of the energy consumed to train the model, or only the marginal energy consumed by inference?

I used to believe this wasn't worth considering (because while training is energy-intensive, once the model is trained we can use it forever.) But in practice it seems like we switch to newer/better models every few months, so a model has a limited period of usefulness.

> (is watt-hours the right unit of measurement here?)

Yes, because we want to measure energy consumed, while watts is only a measure of instantaneous power.