Im guessing its over by 108 VMs for CPU demand? and not allocation.
Under the effective policy on the cluster Under setting 3a capacity and time remaining, do you have use stress to account for spikes and peaks checked? It usually is by default. I have seen this cause so many issues in environments that may having something like av scans or backups that cause a really big peak for all VMs at a particular tie of day and because of that even though it doesn't matter for day to day it effects the capacity management really badly.
Also under 3b you can remove HA out of the equation and buffers if they are enabled
While some things will change instantly vCOPS Does its main capacity calculations once a day. so you may not see the effects of the above changes until tomorrow.
Cheers