It should also be noted that C++ itself isn't fundamentally a fast language from a design perspective (and if you make the mistake of having a lot of "new"s and "delete"s going off at inopportune times you'll see its true potential sluggishness). Its primary benefit is that it lets you talk to hardware more-or-less directly which lets you the intelligent programmer (with the help of the compiler) align the interests of the hardware with the interests of the problem you're trying to solve, whereas Python is mostly about just solving the problem even if it takes a long time to compute. Which makes it an excellent platform to learn anything abstractly (like game design) off of while not simultaneously fighting with the language on account of hardware being a particular way. You can focus on learning algorithms and data structures that improve efficiency in a language-agnostic way and with Python's flexibility you can also explore algorithms that don't usually get mentioned in the C++ world that may fit your particular case (and learn to love native dictionaries/hash tables and lists).
Sufficiently smart compilers have been able to compile Haskell and (slightly annotated) Common Lisp to code that beats C++ for a long while now, and FORTRAN is still used in some circles because it's still very fast for number crunching. Projects others mentioned like PyPy continue to narrow the margins though they've got a ways to go. Unfortunately talking to hardware directly with Python usually means doing it with C/C++ then talking to those libs from Python. (As is the case for PyGame's core (SDL in C), PyOpenGL, etc.) I'd like to say this is the standard approach for most Pythonistas: write everything in Python, then rewrite the slow bits in C/C++. Fortunately a lot of the reasons we might need to talk to the hardware (graphics drivers, screens, fast number crunching) are general enough that other people have done most of the work for you to take advantage of. Even in the FPGA world where you get to design your own hardware with code, there's the MyHDL project which provides Python with nice sugar for HDL code and compiles your source to Verilog or VHDL which you'd normally write instead at your own peril and frustration.
