It's not the ML itself, it's the ability to quickly store large amounts of data and troll through them in a few hours and then shut everything down and never us

It's not the ML itself, it's the ability to quickly store large amounts of data and troll through them in a few hours and then shut everything down and never use it again. One example is the case files I told you about. It's basic image recognition really, but it's millions of poorly scanned pages. 10 years ago, we would've gone through it by hand, and it would've taken 8-12 full time employees 3 months to do it. Now we can have one PHD student do a 2-3 month training project that he/she uses for their research, and we have an algorithm that can go through the case files in 5 hours, provided we rent enough iron to do so. With AWS or Azure that's not a big problem, depending on size you can do it on-demand, but we had to call Microsoft and prepare 5000 server instances when we did it. That was some years ago by the way, so maybe it's on-demand now. Anyway, they went through the case files in 5 hours, and had a lower error margin than our employees would have had. After those 5 hours, you shut everything down. There is really nothing comparable to that. You can certainly do that without moving other things to AWS or Azure, like I said, but it's often the financially solid decision. And what's the problem with it exactly? Our own internal network has been breeched more than once, our Azure instances have never been hacked.