what is the source type of your data? Featureclass tables? dbase? excel?
And I assume you want to automate the whole process rather than doing the process table by table and field by field.
You could use numpy, and convert the tables to numpy arrays and do the columns all at once.
b
Out[6]:
array([[ nan, 5., nan, nan, 5.],
[ nan, nan, 8., 6., nan],
[ 5., nan, 7., 4., 7.],
[ nan, 9., nan, 9., nan],
[ nan, nan, 4., 9., 6.],
[ nan, 4., nan, 6., 5.]])
np.nanmean(b, axis=1)
Out[7]: array([ 5. , 7. , 5.75 , 9. , 6.33333333, 5. ])
Then you could cycle through the tables and get the mean
PS 'nan' is the numpy equivalent of None in python and nodata in tables.