Yup- if you have control over the entire deployment, more reasonable options open up.

For my purposes, CUDA isn't really an option due to needing to run simulations on both servers and unknown clients (ideally with a single codebase), but for other services that never leave the datacenter I'd definitely consider it.