In this presentation we will show how pathwise XVA metrics and the embedded conditional risk measures (value at risks and expected shortfalls used for assessing dynamic initial margins and economic capital) can be learnt by training neural net parameterizations to simulated labels. The challenge posed by the hybrid nature of the features, which includes both diffusive (market risk) and discrete (default) components, is addressed by a hierarchical simulation technique, whereby an optimized number of default paths is simulated given each market path. The regression error of the schemes can be assessed by an a posteriori Monte Carlo procedure. Based on joint work with Bouazza Saadeddine and Lokman Abbas-Turki (mainly).