Have model training make checkpoints every five minutes. When the training gets interrupted, just restart it on a new instance from the last checkpoint. Repeat as necessary until hundreds of hours of training is completed.