But if you're thinking about scientific programming, it's not so much effort to use AWS GPUs. I do that, and don't think I would buy an eGPU even if it was CUDA capable.
AWS is expensive if you are an extremely heavy user.
Checking out whether the pricing works for you is an easy exercise. A p3 instance costs $3.06 per hour at the time of writing. It runs an Nvidia V100, which costs around $10k. So you've got yourself around 3000 hours of use before it makes sense to buy your own.
If you want something cheaper, you can go with the p2 instances at $0.90 an hour. These cost around $2k, so you're looking at around 2200 hours of use before it might become economical to buy your own.
I don't want to sound like an AWS fanboy, but I do believe it's a good democratising catalyst for deep learning and scientific computing.
EDIT: p2 instances run on K80 GPUs, which I neglected to mention above.
It depends. Small instances are a few bucks an hour, so if you just need a GPU for like a hundred hours then it’s great. If you are running the instance 24/7 for months on end, the equation changes. The big instances are cool though because they give you access to hardware that would be pretty unobtainable otherwise
