It does need some horse power, yes. But you wouldn't do the calculations on the stage. The same way you wouldn't make a profile on stage.
I looked a bit at it, and it does seem very interesting. No desire to switch out my kemper (which is collecting dust for no fault of its own), but I find this impressive. Even moreso as it is not a company making this (an open source, no less).
I haven‘t yet, time is a bit limited right now. But of course it is impressive to develop sth like this on an open source base.
Nervertheless, also applying neural nets requires quite a bit of power, not just the training. On the other side, it is remarkable how good the kpa sounds with its limited capabilities (compared to e.g. a GPU).