Given the strong prior pattern, I forecast this is due to (a) changes in methodology/non-random sampling, with 60 % probability, (b) random sampling effects t

Given the strong prior pattern, I forecast this is due to (a) changes in methodology/non-random sampling, with 60 % probability, (b) random sampling effects that go away on longer timelines, with 20 % probability, (c) any actual change in CPUs, with 20 % probability. In my experience, when a metric shows weird behavious these are usually the reasons with roughly those proportions.