I would say this is precisely the reason, with normal drivers on a server, all LFS needs to do essentially is to draw the other cars and interprate the driver inputs and update the cars position on track, as the damage/physics for your car are the only cars physics that have to be created by your machine, but with AI's its different, not only would your machine have to draw everything and calculate your 'real' cars physics, it would also have to do all that work to every AI car you add as well, literally create every single thing that AI's do, in real time, as there could be x amount of things which could cause the environment to change, hence why LFS would need to literally do every calculation for the AI's near enough frame by frame, it couldnt be pre-defined completely, the rough driver paths could be, and are, but anything else can and does change, and all that calculating takes a heck of a lot more resources.
Now, i dunno what kinda spec PC you have to test this, put try running in a 20 car grid on a server online, then create a 20 car grid of AI's offline and compare the results, i bet there will be a significant performance hit due to all the extra work that LFS has to do.
You have probably seen what happens when 2 cars hit and there is lag, and thats when there are 2 seperate PC's running 2 cars worth of phyics, just imagine the outcome of that when 2 PC's are having to essentially 'run' lets say their own 'real' car and 10 AI's each, doesnt bare thinking about does it??