This issue seems to be related to floating point precision issues when camera is far from the world origin. In deferred rendering mode the screen pos + depth is converted to world space and further transformed for shadow mapping which suffers from floating point precision issues when camera is far from the world origin (e.g. camera is at [495, 1.5, -495] in the repro sample). In HDRP this is addressed with "camera-relative rendering" but URP doesn't have this option. For example when expanding the plane ("Cube" object) scale from [1000, 1, 1000] to [1100, 1, 1100] the artifact goes away but there's quite evident shadow jittering from the shadow casting cube on the plane, which is indication of the float precision issue. Also, further debugging this issue the depth differences are quite minor in the artifact area between shadow map and pixel coordinates projected to the shadow map depth. We'll need to implement camera-relative rendering in URP to fix this issue.