Running a large model on rented GPU is still meaningfully more private than handing your chat logs over to FAGA