There are several well-documented, major production deployments of MongoDB already, that seem to contradict your assertion:
* Etsy
* CERN
* BoxedIce (600MM+ documents)
* BuzzFeed (400MM+ datapoints/month)
I don't know about riak, but we are forced to use couchdb at work, for the wrong reasons I think, and I was not impressed. First, its performances are pretty bad: document insertion is slow unless you batch them, but even though it is slower than in mysql, even though the guarantees about durability are certainly not the same (couchdb was not configured to fsync at each write). Building views is excruciatingly slow, and the view engine does not use all the CPU available (even when IO wait is low, so not a simple IO issue). There are bug reports about this issue (cannot find it at the moment). Replication is not very reliable either for large data (where large against means a few GB, so actually not that large).
I am sure we are using couchdb the wrong way at work (out of my hands), and for the wrong application, but its design choices as well as my limited experience does seem to imply couchdb is not adapted for large amount of data, especially one which are written often (there was an interview last summer with D. Katz who said that there were not yet much optimization for large data).
here is the interview of Damian Katz concerning couchdb: http://howsoftwareisbuilt.com/2010/06/18/interview-with-dami..., which mentions that large data is not a focus.
There are some companies which seem to use couchdb for large usage, for example bbc (http://enda.squarespace.com/tech/2010/3/4/couchdb-at-scale-4...). I don't know their infrastructure, but they claim to server 4 billions requests as of 4th march 2010 since summer 2009 on a 32 nodes (16 master, 16 backups). Assuming that summer starts in september to get an upper bound of the traffic, this means 250 rq/sec on average, which is nothing impressive for an infrastructure with 32 machines without more information about what they do. Generally, I would not say much about this kind of usecases, but since it is often advertised by couchdb proponents, the burden of the proof is theirs.
As far as performance, the key with Couch is to keep the view generation speed faster than the insert rate. So if your users are generating less than about 1000 changes a second, you should be fine on a single server.
More than that and you may need to shard / partition, which can be done in a few different way. The leading option is BigCouch: https://github.com/cloudant/bigcouch
I am actually working on simple benchmarks representative of what we do to see if there is something worth submitting as bug issues.
With MongoDB you simply add a cheap second box and use it as at least a replica of your database. When you get tired of "we're updating the site with shiny new code" interrupting your users, you simply make it into a web-node as well.
Here are some reasons:
* excellent documentation
* runs right out of the box
* excellent libraries like Mongoid
* user can easily perform deep queries (e.g person.address.zip = '90901')
* services are available like MongoHQ
If your data was worth anything you'd also be using replication in which case the single server durability point becomes moot.
In each of the cases where the customer had a true standby system, implemented via replication, log shipping, or mirroring, they were able to failover with little (log shipping) or no data loss.
In the cases where they had a single, standalone server, the option was to restore the last known good backup, or sent the database files had to Microsoft for analysis and repair.
ANY system (RDBMS, NoSQL, or otherwise), should have a standby replica to prevent data loss. If you data is stored on a single machine, you are doing it wrong.