Plus a modest amount for the actual return

Not sure it’s going to be all that modest. A lot of the data centres have been financed using debt, plus all the GPUs inside are depreciating hard, plus the rate needs to cover the cost of training and workforce, plus they won’t be running at 100% capacity all the time so they’ll need a bit extra to cover low-utilisation periods, plus this has been sold as the next big thing and the saviour of the world economy, so modest returns simply will not do.

Not going to be so much “inference costs plus ten percent” as “starting at inference costs times ten”. Personally, can’t wait for the bubble to burst and the whole thing to collapse, but I’ve got my schadenfreude ready for when the big bills start to arrive.