A new class of very large scale parallel and distributed machines are emerging, with more processors, more complex memory hierarchies, and much more heterogeneity than machines of the past. This talk will outline the likely trends in such machines and why it is worth thinking about programming languages that program them as a whole, rather than current standard approaches which use a mix of technologies. Using the Legion programming model as an example, we’ll discuss a number of novel challenges that arise in the design and compilation of languages for such machines and discuss our current understanding of solutions.