Every public figure for what a model query costs the grid is a number I cannot verify, so I have stopped repeating them and started metering my own agents instead.
People ask me, because I run a fleet of coding agents most of the day, roughly how much power the whole thing draws. I used to answer with whatever number I had most recently seen quoted somewhere, rounded to sound authoritative. I do not do that any more, and the reason is not that the topic does not matter, because it obviously does. It is that I went looking for where those numbers actually came from and could not find a source I would trust with anything else I care about being right.
The public figures for what a single query costs in energy vary by large multiples depending on which one you land on, and every version I found relies on assumptions its author does not fully show: which hardware, which model size, how much of the surrounding infrastructure gets counted, whether the number is a single pass or an amortised share of training. Nobody publishing these figures shows the arithmetic in a way I could check line by line, and stacking three of them side by side gives three different answers that do not reconcile. I am not in a position to referee that argument, and I would rather say so plainly than pick whichever number best supports the point I wanted to make.
So this article is not going to give you a number, and I mean that as the actual argument, not as a disclaimer before I sneak one in anyway. Repeating an unverifiable figure because it sounds concrete is worse than saying nothing, because it launders someone else’s guess into something that reads as an established fact once enough people have quoted it. I would rather tell you honestly that I do not know the true cost of a query than hand you a number I picked because it was the one I had seen most recently.
What I can measure, and what I have started actually measuring, is my own consumption. I put a meter on the machines I run my agents from, at the socket, and I log the draw against the sessions running at the time. That gives me a real number for a real thing: what my own estate costs, in the electricity I am actually billed for, running the actual workload I actually run. It is a smaller and less impressive-sounding number than a headline estimate, and it has the one property that matters to me, which is that I can stand behind it because I measured it myself rather than inherited it from somewhere I cannot audit.
It also turns out to be a more useful number for the decisions I actually make. A global average tells me nothing about whether switching one workflow from a larger model to a smaller one changes anything, because it was never measuring my workflow in the first place. My own meter tells me exactly that, before and after, on the same hardware, under the same conditions. That is the shape of number worth acting on, one you can attribute to a specific cause, rather than the shape of number that is mainly good for a headline.
I do not think the energy question is unimportant, and I do not think operators should shrug it off because the public numbers are contested. I think the contested state of the public numbers is itself the reason to stop relying on them and start metering what is actually mine to measure. If enough operators did that, the honest aggregate would eventually be better evidenced than any of the current estimates, because it would be built from real meters rather than from assumptions nobody outside the room can check. Until then, the correct answer to what something costs is either a number I measured myself, or an honest admission that I do not know, never a number I read somewhere and liked the sound of.