Definitely not true for these workloads. TPUs are interconnected, one dying makes the whole cluster significantly less useful.