Hmm...I don't think you would need that much of a performance drop.
If the data was received encrypted, you could have a relatively small decryption key sent when necessary, thus solving the "prefetch from server" problem.
If the model files were broken up into chunks and xored with each other in various ways, and saved in a redundant manner, I could imagine that there is a setup in which there are a combinatorially large number of possible ways to fetch a set of blocks which reconstruct a given model, and if you prefetch a few red-herrings as well, it would be infeasible to determine which model you are loading based just on the commands the sever sent you on which blocks to load. Then, when the time comes, they send a very short command which tells you which blocks to use and how to reconstruct the model. Hmm... but now I'm thinking about it, I think you'd need a large overhead in both directions to make it difficult to just match the patterns of which blocks you used.