The context length scales quadratically in terms of compute, so barring algorithmic improvements, there's definitely a limit.
OK, so we'll have to deal with this for a good while