I saw a paper about this in early 2020 (pre-COVID shutdowns) at the ScaledML conference. I looked into it and had the same conclusions. At some point, running your own models in the clear is just more practical.