We do not really know how to do RSA encryption of messages which span multiple blocks. At least not securely. The problem is already hard for block ciphers (see all the defined modes of operation).
Also, the worst performance issue with arithmetic cryptography like RSA is not about CPU, but about size. When you encrypt a piece of data with RSA, you get some overhead. E.g. with a 1024-bit RSA key, you can encrypt at most 117 bytes at a time, but this yields 128 bytes. When transmitting a long message, a fixed overhead of a few hundred bytes is much more tolerable than a +10% size increase.
CPU efficiency is often quoted as the reason we do hybrid encryption, but it is rarely as compelling as the two problems explained above.
(Additional point: since proper RSA encryption needs randomness, you must have a cryptographically strong PRNG -- also known as "a stream cipher". So even non-hybrid encryption does not avoid using symmetric primitives.)