if those users were exposed to the full financial cost of their toy they would find other toys

And what is that cost, if you have it handy? Just as an example, my Radeon VII can perfectly well run smaller models, and it doesn't appear to use more power than about two incandescent lightbulbs (120 W or so) while the query is running. I don't personally feel that the power consumed by approximately two light bulbs is excessive, even using the admittedly outdated incandescent standard, but perhaps the commercial models are worse?

Like I know a datacenter draws a lot more power, but it also serves many many more users concurrently, so economies of scale ought to factor in. I'd love to see some hard numbers on this.

IIRC you can actually get the same kind of hollow praise from much dumber, locally-runnable (~8B parameters) models.