Develop on a tiny slice of you dataset locally of you must. Then train properly elsewhere (in the cloud of your choice)