Just to be clear, the effective bandwidth is 5 MHz, but the spectrum width before the edge transition is around 4.5 MHz = (301 non-edge subcarriers) * (15 kHz subcarrier width).
I think your confusion is very common.
The sampling theorem tells you that you need a sampling rate at least 2x the signal bandwidth to prevent signal information loss (perfect reconstruction possible). Remember, however, that isn't necessarily important for communication.
With matched filtering, it is pretty straightforward to show that the maximum SNR can be obtained for communication while sampling at only the symbol rate (7.68 MHz), which is less than 2x the signal bandwidth (~10 MHz).
Typical receivers still oversample anyway so that functionality can be pushed into the digital domain (AGC, bandlimiting front-end filters) and/or so that practical digital operations such as time and frequency synchronization have maximum performance.
Long story short, for simplicity I recommend sampling at an integer (>=2) multiple of 7.68 MHz if you have the computational resources. What that multiple is will depend on how the other functions of your receiver behave (especially synchronization). After synchronization you can easily decimate down to the symbol rate for equalization. You can probably get away with sampling at 7.68 MHz, but your performance will suffer.