Most expensive if you discount reliability, longevity and persistence in energy generation.
A predictable energy generation source for a predictable work load makes sense.
A predictable energy generation source for a predictable work load makes sense.
I'm curious, do you actually run servers for a living? I'm wondering what a typical use case is (aside from crypto mining) where powering off and on physical servers due to demand, cost, or availability of power is a thing.
In practice…yeah. The machine’s your running your models on are expensive, the people working on the model waiting extra time to check the new model are expensive, and not keeping up with your competitors because you only work half the time is not generally a wise choice for someone like Microsoft.
I have worked for a smaller company that trained a model on AWS when spot instances were cheap…until they got large enough to purchase some hardware best suited to their use case that could run almost constantly, then doing everything except training their model in the cloud. So yeah I also doubt this makes sense in practice.
Another consideration is that servers can be located in the best places for renewables, not in places with high seasonal variability and relatively low winds.
-- A++ customer experience
Dispatchable load usually means oversizing the dispatchable consumer to get the same overall output. This is already uneconomical for even particularly energy intensive industries (e.g. aluminum smelting). I would assume server hardware is a lot more capital intensive than that.
Pausing the plant doesn’t just mean pushing a button. Safely shutting it down is a long process of draining every single part of molten metal first, and might not even be doable; almost any other factory would have an easier time.
A related problem is that the walls of the pot cannot be allowed to get too hot. If this were not the case, they could simply insulate the pots to reduce heat losses to some arbitrarily low level.
The benefit of your services having 25/7 availability aside, if the capital costs are high compared to operational costs, reducing operational time to save on operation costs is a bad trade.
Debating total cost depends almost entirely on the particular grid in question. For all we know, those "nuclear fans'" estimates could be spot-on but completely irrelevant to you.