This rests on a lot of assumptions that the published figures for "planned" datacentres, "committed" AI spend, etc. are irreversible. I suspect that at least some of it is possible to back out of.