"Suppose we use the subset construction to convert a $7$-state NFA $M = (Q,\Sigma, \delta, q_0, F)$ into a DFA $M' = (Q', \Sigma, \delta', q_0, F')$ for the same language. Then $M'$ will have $|Q'| = 128$ states. If $|F| = 2$, then $M'$ will have $|F'| = 96$ final states."
From the subset construction, I know there are $\mathcal{P}(Q)$ new states in the machine, which means there are $2^7 = 128$ new states in $M'$. I am not sure mathematically why there are $96$ final states.
Observations:
I see that if I have an initial NFA with $3$-states $\{1,2,3\}$ and a $|F| = \{1\}$, then $P(\{1,2,3\}) = \left\{ \emptyset, \{1\}, \{2\}, \{3\}, \{1,2\}, \{1,3\}, \{2,3\}, \{1,2,3\} \right\}$. I can also note that there are now $4$ final states and can conjecture $|F| = 2^{n} - 2^{n-k} = 2^{3} - 2^{3-1} = 4$, where $k$ is the number of final states. I also notice that each particular state occurs $2^{n} - 2^{n-k}$ times in the construction.
However, if I let there be $2$ final states (say $\{1,2\}$) in the initial machine, the new machine will have $5$ final states, not $2^3 - 2^{3-2} = 6$. I can see this is because $\{1,2\}$ is a state that contains both elements, and is being double counted.
I would prefer simple well-written intuitive answer rather than an overly mathematical one.