+1 to GPGPU specifically. It feels like there is enormous opportunity here, as existing languages have serious limitations (CUDA is effectively Nvidia-only, choices for compute shaders are extremely low level), and at the same time I think there's a lot of very cool stuff you can do, largely held back right now by poor tooling.

I suspect that progress on this front will overlap a bit with the previous category, as more often than not you run machine learning workloads on GPU anyway.