I'm willing to credit this type of feature creep for some of the problem, but only some.
Decades ago, we had hardware that was so much less capable than anything anyone has put in a car in ten years that the comparison is absurd. Nonetheless, those old systems were often very capable of doing useful things, and the UIs were often very, very fast by modern standards. And a lot of this was that the software stack was tuned for performance and latency and did not do things that hurt latency too much, and this came at a substantial modularity cost.
Back when a TI calculator with a whole OS running on a Z80 was an order of magnitude or so faster than your average new car, it didn't have a compositor (which still eats a good frame or so of latency, maybe more, and, as an industry we still don't know how to solve this). It ran an OS that was surely far harder to tweak than a modern car stack, but it had fewer layers and could be kept performant. An Apple ][ and its clones were, of course, much faster than a TI calculator, and we still can't match that level of performance.
A group of engineers that were willing to keep the managers away and make a simple, limited-purpose system, could surely get excellent performance out of any currently available car computer hardware, but it would come at a cost. Buzzwords would not happen :) Also, the software developers had access to full data sheets for the hardware, and when you told the hardware to put something in its framebuffer, it was there, and no blobs or firmware were in the middle.
I still remember when Windows 2000 launched with "layered windows": an application could opt in to being composited, and it would gain the ability to be translucent, but it came at a very obvious performance cost. Sadly, 23 years later, that performance cost is still there, but it's not opt-in any more.