That's not really how quantum works. It will bend the curve if we can have enough qbits. But it doesn't make P=NP or anything like that. Parallelization is still going to be more useful for neural nets, because LLMs are most computationally intensive at training time, not execution time. What quantum might be able to do is explore more alternatives (e.g., quantum is really good at speeding up classical search) in a fixed amount of time with a world model. That world model doesn't come for free, however, it can only be acquired through interaction with the real world, and quantum doesn't speed that up. Also, this assumes planning in a fully-observable world where all variables necessary to generate a sound plan are available in advance. For most real problems, that information can only be acquired at inference time, and quantum cannot speed that up earlier.