Numerical representation and climate projections of sea surface winds over Southeast Asia (SEA) are assessed here using an ensemble of the Coupled Model Intercomparison Project Phase 5 (CMIP5) downscaled simulations performed over the 20th and 21st centuries under Representative Concentration Pathways (RCPs) 4.5 and 8.5 scenarios within the Coordinated Regional Climate Downscaling EXperiment (CORDEX)‐SEA project. The ensemble is based on two regional climate models (RCMs: RegCM4 and RCA4), and CMIP5 simulations are performed with five global climate models (GCMs: CNRM_CM5, HadGEM2, GFDL, MPI‐ESM‐MR, EC‐Earth). Comparison with QuikSCAT satellite data shows that dynamical downscaling improves sea surface wind speed representation, mainly by reducing its underestimation. The level of improvement depends on the RCM choice, GCM performance, and wind strength. Our results reveal significant differences in modelled projections of sea surface wind, depending on the model, RCP, region, and season. GCMs simulate weak and contrasted changes, stronger for RCP8.5, with no clear common trend. RCA4 simulates weak changes, with high similarities between pairs, but contrasted results between RCPs. RegCM4 simulate stronger changes, with a weakening of average and intense winds for all seasons, stronger in June–August, and in RCP8.5 than in RCP4.5. RCA4 and RegCM4 simulate different changes, with no clear common trend except a weakening of seasonal and intense winds and an increase of seasonal wind interannual variability for June–August in RCP4.5, stronger for RegCM4. This corresponds to a weakening of the boreal summer monsoon and a slight increase of its interannual variability and presumably to a decrease of the tropical cyclone frequency. Differences in seasonal sea surface wind changes between models are related to differences in sea level pressure gradient changes. For a given RCM, those differences are partly related to the differences between parent GCMs. Finally, results suggest that uncertainties related to the RCM choice are larger than those related to the GCM choice.