In https://en.wikipedia.org/wiki/Geometric_complexity_theory it is mentioned that ".. Ketan Mulmuley believes the program, if viable, is likely to take about 100 years before it can settle the P vs. NP problem".
It seems to indicate that the only currently viable program could face serious obstacles.
What are some of the obstacles where the program could fail?
It depends what you count as "the GCT program."
Note, however, that if instead of multiplicities you merely want a separating module, then the strong perm v det conjecture is true if and only if there exists a separating module.
If your goal is the original permanent versus determinant conjecture, there is an earlier step in GCT, namely (as pointed out by chazisop) moving to the strong perm v det conjecture by considering the orbit closures. It is conceivable that the original permanent versus determinant conjecture is true but the strong version is false. However, this seems highly unlikely to me. Also, if this is the situation, then none of our current methods can even come close to resolving the perm v det conjecture, since they all currently work for the "strong"/"approximative"/"border-"/Zariski-closed version of whatever algebraic complexity statement they are proving.
If your goal is not perm v det but Boolean $\mathsf{NP} \not\subseteq \mathsf{P/poly}$, there are additional claimed steps in GCT that have yet to be published. It is possible that one of these unpublished steps could fail as well, but obviously it is hard to comment on the details of mathematics one hasn't seen...
[Potential failures of lower bounds in general, not specific to GCT.] GCT is currently aimed at nonuniform lower bounds; that is, even in the GCT approach to Boolean lower bounds, it is aimed at showing $\mathsf{NP} \not\subseteq \mathsf{P/poly}$. But of course, it is consistent with current theorems that $\mathsf{P} \neq \mathsf{NP}$ yet $\mathsf{NP} \subseteq \mathsf{P/poly}$. Of course, it's also technically possible that $\mathsf{P} = \mathsf{NP}$ and the perm v det conjecture is false!
However, let me point out that the GCT program as it currently exists still seems to me like the first thing to try. If it turns out that any one of (1)-(3) above actually doesn't work, it will mean that the perm v det conjecture (and hence, $\mathsf{P}$ versus $\mathsf{NP}$) is almost unimaginably harder than we currently think it is. (It may be worth noting that this statement is coming from someone who already thinks that the following analogy may be roughly correct, if not inadequate: $\mathsf{P} \neq \mathsf{NP}$ is to our current state of knowledge as the Classification of Finite Simple Groups is to Fermat's Little Theorem). And even if that is the case, understanding the exact way in which the failure occurs will likely be important to making further progress.
I believe the "100 years" statement refers that the theory is general, but requires deep understanding and new results in representation theory and algebraic geometry to progress, something that might be slow to progress (I want to make a comparison to number theory, but I am not sure how apt it is).
Also, there is a loss of precision when translating to the algebro-geometric world: Instead of proving a lower bound against the properties of a complexity class (i.e. polynomials that vanish when objects in that class are given as input), you are proving it against its Zariski closure (of aforementioned polynomials). It is conceivable that in order to separate the two, one has to examine the boundary of that closure (those polynomials that occur only in the closure but on the original set). It is believed that in the determinant vs permanent variant of the GCT program, this is likely the case.
Finally from personal experience, the skillset required to understand GCT in depth is quite different than what is usually the focus of undergraduate or even master programs in CS, essentially picking up the prerequisites is a natural follow-up of choosing to study GCT.
