IMO long format is much more preferable as you will have much less metadata overhead (information about column names, dtypes, etc.).
In term of memory usage they are going to be more or less the same:
In [22]: long = pd.DataFrame(np.random.randint(0, 10**6, (10**4, 4)))
In [23]: wide = pd.DataFrame(np.random.randint(0, 10**6, (4, 10**4)))
In [24]: long.shape
Out[24]: (10000, 4)
In [25]: wide.shape
Out[25]: (4, 10000)
In [26]: sys.getsizeof(long)
Out[26]: 160104
In [27]: sys.getsizeof(wide)
Out[27]: 160104
In [28]: wide.info()
<class 'pandas.core.frame.DataFrame'>
RangeIndex: 4 entries, 0 to 3
Columns: 10000 entries, 0 to 9999
dtypes: int32(10000)
memory usage: 156.3 KB
In [29]: long.info()
<class 'pandas.core.frame.DataFrame'>
RangeIndex: 10000 entries, 0 to 9999
Data columns (total 4 columns):
0 10000 non-null int32
1 10000 non-null int32
2 10000 non-null int32
3 10000 non-null int32
dtypes: int32(4)
memory usage: 156.3 KB