Modulation with Maximum Likelihood decoder and capacity

My question comes from the high spectrum efficiency claim of OFDM. This seems not correct IMHO, at least theoretically.

For me, we lose energy in using cyclic prefix, hence OFDM should be a waste of theoretically achievable rate. That said, OFDM has advantage in implementation and it is easier to get a high throughput.

Long story short, my question is: if complexity is not my concern, can (any) modulation (without channel coding) combined with a Maximum Likelihood (or Maximum a Posteriori) decoder achieve asymptotically the Shannon capacity ?

One thing to keep in mind is that OFDM is designed for the wideband (frequency selective) fading channel. The capacity of such a channel is not the usual $$C=B\log(1+SNR)$$, which only applies in the AWGN channel. Since the fading channel is time-varying, other measures of capacity (such as outage probability) become more important.
To answer you specific question: to my knowledge, it is impossible to reach capacity without coding, except when $$\text{SNR} \rightarrow \infty$$.