We point out that the observed time delay between the detection of the signal at the Hanford and Livingston LIGO sites from the gravitational wave event GW150914 places an upper bound on the speed of propagation of gravitational waves, $c_{gw}lesssim 1.7$ in the units of speed of light. Combined with the lower bound from the absence of gravitational Cherenkov losses by cosmic rays that rules out most of subluminal velocities, this gives a model-independent double-sided constraint $1lesssim c_{gw}lesssim 1.7$. We compare this result to model-specific constraints from pulsar timing and cosmology.