byteleap per gpu-hour undercuts the big clouds by a lot if you can tolerate distributed nodes. did a 3-day training run, one node dropped once, checkpoint resume saved it. worth it.

BitFan
Public Service Atlas for Bittensor
byteleap per gpu-hour undercuts the big clouds by a lot if you can tolerate distributed nodes. did a 3-day training run, one node dropped once, checkpoint resume saved it. worth it.