I wonder how much of Elon's conviction on this is because he's forced himself into a corner, years ago, by stating that Tesla would make FSD work just using cameras, plus the forward looking radar and close range ultrasonic sensors? Back when Tesla first announced the self-driving goal, LIDAR was mega expensive, around £100k for a reasonably good 360° unit. Prices have really tumbled, though. Google have kept their sensor development in house, so we don't know how much the Waymo sensor suite costs, but it's estimated to be under £5k. At that price it's not a lot more than the multiple cameras and radar fitted to a Tesla, I suspect, and there's almost certainly room for that price to fall further.
Making LIDAR work over a wide range of conditions is pretty easy, too. Around 35 years ago I made a homebrew directional LIDAR ranging unit for cave surveying, using a salvaged IR fibre optic drive LED and some lenses salvaged from old cameras. That worked reliably out to about 50m, using a retroreflector, even when the lenses were a bit muddy, and inevitably wet (nothing stays dry for long down a cave). That homebrew unit was mainly analogue, in that the ranging system used sine wave modulation of the laser diode, together with a phase measurement system that looked very like two radio IF strips, feeding a comparator. The digital bit was one of the early CMOS versions of the 6800, that measured the pulse width from the comparator, did some simple arithmetic and displayed the distance on a display, with an accuracy of around +/-20mm and a resolution of 100mm (good enough for cave surveying).
The really big advantages of shifting more of the heavy lifting into more capable sensors has to be the reduction in processing that gives, together with the likely reduction in erroneous target ID. The output from a LIDAR sensor, for example, is already pre-processed, in that it gives the dimensions and position in 3D space of everything it sees, with no need to do this using software. The software can then spend more time doing other stuff. This approach probably reduces the power demand, too.
The problem of keeping a LIDAR sensor clean has already been solved for units being used in farm and industrial machines, and farming has to be a more challenging environment for sensors overall. If I had to pick one industry that will see the most rapid growth in autonomous vehicle use then I think it will be farming. There are so many farming activities that are well-suited to autonomous vehicles, and where both costs could be driven down and productivity improved, just by removing the limitations that human operators have, that it's a no-brainer to use autonomous machines for ploughing, soil conditioning, sowing seeds, weed control, fertiliser application, harvesting etc.