Also worth considering how many years it is going to take to offset the sizeable cost of such a migration.

Forget the price of the hardware, how long is it going to take to offset the cost of the programming and administration labour to pull off this sort of move?

What about maintenance? We've got years of experience with Nvidia cards in datacentres, we understand the failure modes pretty well, we understand the tooling needed to monitor and triage these systems at scale.

What guarantees do I have that if I fill my racks with this hardware they won't be dying or catching on fire within a year?

What guarantees do I have that Intel won't unilaterally decide this is a dead cat for them and they want to scrap the project? Like they have for almost every GPU adjacent project they've had.