Around one third (30%) was the educated guess that Josef Ka used in Coach Parrot, but nobody really knows. There are no data points for this because mostly because you don't know exactly when the player becomes capped. Potential has sublevels so a Superstar player can cap at 8.00 or at 8.99, but you won't know until you see that pops are not coming as they were supposed to and even then you cannot track exactly when when you went above the cap, since we don't know the exact potential sublevel.
Also as hrudey pointed out, because of crosstraining it is very hard to gauge how much is the effect from slowed down training and how much was due to previously accumulated sublevels due to crosstraining (you can track the sublevels due to primary and secondary training, but not crosstraining). Since the slowed down training is actually slow, it is actually comparable to the full speed crosstraining effect (that should be around 10% of total training and hits random skills). You would need a player who is definitely over the cap and then train 2 straight levels in a single skill to have an good estimate of the actual speed after capping.