I think you're conflating privacy and security. Serving the model on the client side means that the server is never aware of who put the inputs in.

Also he isn't contradicting himself, serving static data off a CDN or server is a lot easier and a lot less expensive than having a cluster setup to do inference.