It seems likely you're measuring something other than what you intended to measure, like the start-up time of the Python interpreter, time to byte-compile Python source, etc. fib(40) itself runs in 1.1us/loop on this Ryzen 3700X running Debian Buster and Python 3.7.3, excluding start-up costs using timeit:

    python3 -mtimeit -s"
    def fib(x):
        a = 0
        b = 1
        for i in range(x):
            a, b = a+b, a
        return a

    assert fib(40) == 102334155
    " "fib(40)"

Even this takes only 3.5ms on the "task clock", counting all interpreter start-up time, bytecode compilation, etc:

    perf stat python -S -c "
    def fib(x):
        a = 0
        b = 1
        for i in range(x):
            a, b = a+b, a
        return a

    assert fib(40) == 102334155
    "

(oops, that example changed to python2.7; python3.5 is a bit slower at 9ms)