SPI is fine with multiple devices per bus, but does suffer from the need to dedicate an I/O pin per device for "device select".

I2C of course has device "address" as part of the protocol, so there it's just a matter of how many different device addresses any particular type of chip can do.

But either way, a pair of 16-bit expanders give a LOT of extra I/O. But if one really needs tons of I/O, then perhaps the solution is to start with a larger microcontroller. There are Arduino boards with bigger chips ("Mega 2560" board), and then there's the Wiring-S board (very, very similar to Arduino), both of which offer 54 I/O pins.

Cheers