That would buy you some time.

My thought was that if a simulation that centered around a single individual had a simulation running within it, the simulation would only need to be convincing enough to appear real to that one person. Even if the nested simulation runs a third level simulation within it, or if the one individual runs two simulations, aren't you still basically exploring the idea space of that one individual? That is, me running a simulation and experiencing it through virtual reality is limited in cognitive/sensory scope and fidelity to the qualia that I can experience and the mental processes that I can cope with... which may still be very impressive from my point of view, but the computational power required to present the simulation can't be much more complex than the computational power required to render my brain states in the base simulation. I may simulate a universe with very different rules, but these rules are by definition consistent with a full rendering of my concept space; I may experience new sensory inputs (if I use VR), but I won't be experiencing new senses.... and what I experience through VR replaces, rather than adds to, what I would have experienced in the base simulation.

Even in the worst case scenario that I build 1000+ simulations, they only have to run for the time that I check on them. The more time I spend programming them and checking that they are rendering what they should, the less time I have to do additional simulations. This seems at worst an arithmetic progression.

Of course, if I were specifically trying to crash the simulation that I was in, I might come up with some physical laws that would eat up a lot of processing power to calculate for even one person's local space, but between the limitations of computing as they exist in the base simulation, the difficulty in confirming that these laws have been properly executed in all of their fully-complex glory, and the fact that if it worked, I would never know, I'm not sure that that is a significant risk.