The hashing that is used within a single memcached server is different than the hashing used to determine which server to communicate with. The memcached server does the former, which only needs to ensure low hash collisions for the keys itself is storing, the client software does the latter. You couldn't have a single memcached instance that is terabytes anyway -- can you even get a machine that would take a terabyte of RAM? -- which is why you scale memcached by adding more memcached instances.
