The problem mostly stems from the vast risk you take on from making a large investment in an unstable/unproven platform vendor.

Servers are relatively fungible, given ops automation; it's painful but not the end of the world if you have to migrate away.

But the technology is still relatively immature in that building your own ML service in house - and having it scale, etc - is still a big pain.

I would immensely prefer it if we first brought ML libraries up to a higher level of maturity - as simple as apt-get install and adding `includes ActiveLearning::Bayes` to your models.

But if a client came to me tomorrow and said "there's this great Amazon API that we're thinking of using" I wouldn't consider that insane on first principles.