I’d never even heard of A100s. Those are very new and using 7nm transistors, no company is going to be unloading them cheap for a long time. K80s are 28 nm architecture but the value is there. If you’re looking at a100s though value might not be a concern for you lol.
In my opinion, you don’t need a100 power starting out. You’re not going to get ‘smarter’ networks with a more powerful gpu, your training will just be faster. For deployment, you generally need less compute power than for training. Possible vastly less, like even a cpu might be enough for deployment depending on your use case.