The above code can also be viewed as: Input | A(x) | B(x) | C(x)
Evaluating this lazily will get you: C . B . A . Input
What the program wants to return is the final output of C. When the program tries to return C, it will go down the chain of expressions it must evaluate until it reaches one that it can, in this case A(Input). Once A(Input) is calculated, we no longer need Input it memory, so it can be garbage collected. At this point, the program is: C . B . Input' Because Input' is the same size in memory as Input, we are, at this point, using no more memory than the strict version of the program. We can continue this argument until we eventually arrive at the final state, in which we have our answer.
With this, we can see that any strictly evaluated program can be done lazily with no more than twice the memory requirements (there is a point where we need both Input, and Input'), which from an asymptotic standpoint is equvilent.