>>> x = complex(1e200, 1e200)
>>> x/x
0j
>>> x = complex(1e-200, 1e-200)
>>> x/x
(1.#INF-1.#INDj)
>>> 

This is nothing new; it's always been this way; the 
algorithm we use is numerically naive, ignoring the 
possibility for overflow and underflow in internal 
intermediate results.  The results will vary across 
platforms in such cases, in unpredictable ways (the 
above was run on a WinTel box).