Why not adopt the SSH model of accepting the server's cert at the first connection and then complaining loudly if the fingerprint changes? Or even go a step further and have browsers manage private keys (client certs) which identify the user? Some Linuxes already do this with SSH: if you have a private SSH key, an SSH agent will remember its decrypted form on first use, and use it for SSH auth.
The client certs can be exchanged in a standard format throughout browsers and devices, and mobile phone-based two-factor-authentication can be used to increase security.
So - bam, good enough MITM protection (sure, let the banks retain the PKI system), no passwords, and a completely ready implementation requiring no new standards. There is no requirement for the client certs to be "blessed" with strong identification with the person, so anonymity is achievable by having an arbitrary number of different client certs.
However, the bigger thing a self-signed cert allows you is the ability to verify later if your communication has been snooped by an active attacker--which is very significant. It's problematic that the communications aren't authenticated before they occur, but that doesn't mean they can't be authenticated later (via CAs or other means).
A slightly more sophisticated attack, indeed — but one that doesn't scale well. It ups the cost of mass surveillance tremendously.
As far as I can tell, both insertion of a passive listener and a MITM are algorithmically O(n) on the number of connections being surveilled. All you're increasing for the MITM is the constant factor.
When you're on the order of billions of connections being surveilled, even linear growth is hard, but we know that the NSA already has done that. Increasing the difficulty by a constant factor is not much harder, and there's no question that the NSA has the budget to do so. And in fact, the constant factor isn't even large: it's whatever the resource costs of two connection handshakes per connection is, plus decryption/encryption on the data flowing across, all of which are highly optimized algorithms at this point.
That's point 1.
Point 2: Detection.
Actively MITM-ing an SSL connection requires you to (if the CA Chain-O'-Trust works as supposed) give away the fact that you own a (valuable!) compromised CA authoritative cert. You would want to use such a cert for a targeted attack, not waste it on blanket surveillance and get found out (and called out) within a couple of days.
If you don't own a root cert but rely on vulnerable implementations, or implementations such as this one which do not rely on the CA infra, same story. You do not want to waste that on blanket surveillance and get caught. You'd save it for the /special occasions/.
SSL-MITMing everyone, all the time, as in blanket surveillance, is unfeasable even without CA chains. You'll get called out on it by people that do check cert fingerprints once in a while.
This is a different kind of scalability than the computational order-of-complexity one that you seem to be thinking about.