| I follow it a bit. Right now we know that it's theoretically possible but it's very slow, too slow for uses other than maybe very niche things. In the next five years I fully expect to see an acceptably fast implementation that will broaden use to the application areas you mention: high security stuff and crypto junkies. At that point is where things will get interesting. To make homomorphic crypto really fast will take custom silicon. As soon as I see acceptably fast software homomorphic, I expect to see either a startup or one or more projects inside big companies very quickly pop up with the goal of producing either homomorphic silicon to pair with an existing CPU or -- more interestingly -- a CPU that computes homomorphically as its native mode of operation. That could get interesting. I predict that by 2030 you will be able to spin up a 100% "blind" sealed VM in the cloud for <$20/month that is fast enough to run what can run on an average cloud instance today. All compute doesn't have to be homomorphic to get the benefits. You only really need it for encryption, decryption, key generation, authentication, and indexing of security critical data. Boilerplate stuff can be done in the clear. By 2030 standard cleartext compute will be so cheap I wonder if cloud providers will even bother charging for it below very large levels. Edit: can't make this any deeper but to reply to your reply: I'm not sure. I think that's up in the air. Right now you can do a whole lot with a small cloud instance with <1gb RAM and 20gb local storage. Software bloat seems to be increasing less rapidly than it was 10 years ago, and there are even movements toward more efficient languages today. So I just have to shrug on that one. We'll see. |