I think the key paragraph from the paper is this: > We would like to emphasize that the output of our logic gate can be fed directly as an input to a second si

I think the key paragraph from the paper is this: > We would like to emphasize that the output of our logic gate can be fed directly as an input to a second similar device (as shown below) and so on, to perform all the calculations desired. Clearly, our logic gate is a combinational device; thus, the removal of the inputs makes the system revert to the initial state S0 (regardless of whether the final state was S0 or S1). If we want to remember the final state, we need to couple this device to a sequential device where a Landauer reset might be required and a minimum dissipation of kBT log 2 is needed. In other words, the Laundauer limit applies to state machines, but not to combinatorial logic blocks. I guess in retrospect this should not be surprising: the derivation of the principle considers a computing machine which is in thermal equilibrium with its environment, so implicitly we model the machine as having some memory unit which has time to thermalize after each operation.