First of all: There are maybe 2^160 possible adresses, but not every base number leads to an actual adress. There are about 2^96 possible private keys so far, which can lead to diffenrent adresses if you hash them compressed or uncompressed.
So I dont need to hash 2^160, just 2^96 keys to uncompressed.

With 10 Bk/s and 10 M adresses I am looking for I have the probability of finding an adresses by luck once every 300 years.
However I will increase the rig number and gpus involved to have a speed of 10 Tk/s = 10 000 Bk/s so than I will have have an adress by luck in less than a year. Maybe vanitysearch will also increase speed in the future, so that I don´t need that many gpu hardware.

However we could also make a deeplearning process out of it where the systems learns which binary input change will lead to a prefixed output?
Maybe my maths is not correct, but who cares... I do it for fun and the chance is bigger than 0 and can be increased by speed and number of prefixes. And the softwares in this field increased the speed dramatically over the last years. For example the LBC was happy to have 150 Mk/s with 50 people in a pool and vanitygen was fast if you had 400 kK/s. So yes, it becomes more likely that someone will have luck.
I have 2 rigs at work and 1 comes next week. So as more rigs I am putting to work as more likely it is that I will have luck.