We analyze the effective depth of LLMs and find that they are unlikely to compose subresults, and deeper models spread out the same type of computation as the shallow ones.