I will present the simulation-based cosmological inference from Dark Energy Survey (DES) Year 3 data, using machine learning data compression of gravitational lensing "dark matter maps". DES is a ground-based galaxy survey that has observed hundreds of millions of objects in order to understand the cosmological model that describes the Universe. I will present a new framework for DES: using simulation-based inference to estimate posterior probability distributions of unknown parameters using simulated mock data. This approach improves reliability through more realistic modelling (including measurement systematics, cosmological modelling and neutrinos). It also allows us to use more of our data by inferring cosmology directly from the gravitational lensing maps, which leads to more than a two-fold improvement in our inference of dark energy parameters. The improved reliability and discovery power of this method will have an even greater impact using data from the recently-launched Euclid mission.