Those are completely different levels of heterogeneity. A GPU/TPU might draw a lot of power, but does it have to be separate power? Do you need to put that stuf

Those are completely different levels of heterogeneity. A GPU/TPU might draw a lot of power, but does it have to be separate power? Do you need to put that stuff on a separate slab to isolate it from vibration? Does it need its own air filtration system? Do you need a new runbook, because your usual redundancy approaches can't be used for million-dollar systems? Do your DC techs need new training, because the usual "yank and replace" isn't applicable either? I've watched colleagues in a team adjacent to my own work through these issues. Have you? Slapping a commodity card into a box and loading some commodity software on it seems like a cakewalk by comparison. Obviously people thought it was worth it anyway, but I think you're seriously misunderstanding where the difficulties lie and what kind of resources are needed to overcome them. 200+ engineers would be overkill (unless FB engineers are ~4x more productive than Google engineers) but you'll need other kinds of specialists as well and the installation/operational costs will still be high. It's unlikely to be worth it for an organization much smaller or different.