Aims. We present a measurement of the dependence of galaxy clustering on galaxy stellar mass at redshift z ∼ 0.9, based on the first-epoch data from the VVDS-Deep survey. Methods. Concentrating on the redshift interval 0.5 < z < 1.2, we measured the projected correlation function, w p (r p ), within mass-selected subsamples covering the range ∼10 9 and ∼10 11 M . We explored and quantify in detail the observational selection biases due to the flux-limited nature of the survey, both from the data themselves and with a suite of realistic mock samples constructed by coupling the Millennium Simulation to semi-analytic models. We identify the range of masses within which our main conclusions are robust against these effects. Serious incompleteness in mass is present below log (M/M ) = 9.5, with about two thirds of the galaxies in the range 9 < log (M/M ) < 9.5 that are lost due to their low luminosity and high mass-to-light ratio. However, the sample is expected to be 100% complete in mass above log (M/M ) = 10. Results. We present the first direct evidence for a dependence of clustering on the galaxy stellar mass at a redshift as high as z ∼ 0.85. We quantify this by fitting the projected function w p (r p ) with a power-law model. The clustering length increases from r 0 = 2.76 Comparison to the SDSS measurements at z ∼ 0.15 shows that the evolution of w p (r p ) is significant for samples of galaxies with M < 10 10.5 M , while it is negligible for more massive objects. Considering the growth of structure, this implies that the linear bias b L of the most massive galaxies evolves more rapidly between these two cosmic epochs. We quantify this effect by computing the value of b L from the SDSS and VVDS clustering amplitudes and find that b L decreases from 1.5 ± 0.2 at z ∼ 0.85 to 1.33 ± 0.03 at z ∼ 0.15, for the most massive galaxies, while it remains virtually constant (b L ∼ 1.3) for the remaining population. Qualitatively, this is the kind of scenario expected for the clustering of dark-matter halos as a function of their total mass and redshift. Our result therefore seems to indicate that galaxies with the highest stellar mass today were originally central objects of the most massive dark-matter halos at earlier times, whose distribution was strongly biased with respect to the overall mass density field.