Brendan O'Carroll

I wonder if consideration of the jobs being run might throw further light on the (cash flow) economics of data center operation? Eg. at a basic level, training vs inference. If you're renting to a consumer running inference on a Stable Diffusion session, it is likely that for a fair amount of time, the gpu will be idling. (Session set up, prompt creation, image analysis/critique, prompt revision etc.... Over the course of a day, this can add up to a lot less power use than you may expect. On the other hand, an intelligently cued and controlled 8gpu instance running inference on (e.g., any publicly facing LLM), may end up using more power, but making a good margin by spreading that out over a large number of customers.
Properly managed power is the classical variable cost, and when you don't add in the cost to the Commons, heat, pollution etc., it should closely match actual billable machine use. (All in theory, and excluding the current loss-leader environment of buying customer volume (and maybe loyalty as they gain proficiency and seek in the future to avoid retraining on another suppliers system).