One more comment / suggestion: I watched the video and saw there are lots of pre-configured environments. I see this a lot, but to be honest this is not to helpful for me because I'm not doing development in your environment, I'm writing and checking the code somewhere else and then training on the gpus.

For the same reason, I'm only getting value out of the gpus when I am right ready to train. So I would be much happier if I could push a docker, or maybe even a conda environment spec, along with my code, attach data storage, and run e.g. train.py (or more likely a shell script that calls it) to completion and then immediately release the GPUs. Everything else is just the overhead of as quickly as possible trying to get the environment right, run my script, and then shut down the instance as soon as it's done. It would be awesome to have this kind of functionality, or is there a way to do that I missed?