Whether there would be a benefit is an open scientific/engineering problem. We certainly haven't reached the limits if Jnon-neural interfaces (I still can't casually mumble to my computer and have it reliably understand), but there are definitely limits. For instance, regardless of how much visual data you can present to a user, she can only focus on one place/element at a time. A "neural uplink" might give us ways to bypass or finesse that bottleneck, which would by itself be huge.
Or perhaps not! But Musk has no interest in sure-fire bets. :)